Posts on X

Nancy Pelosi Stock Tracker ♟

@pelositracker • 1d

Just In: George Santos (R) told the public he was going to the State of the Union, then bet against himself on Kalshi to make $17,570 The CFTC fined the former congressman $17,500 and ordered him to give up his $17,570 profit He's also banned from prediction markets for 3 years

Gergely Orosz

@GergelyOrosz • 2d

Selling to Bending Spoons is an admission of defeat+wanting ANY kind of exit, knowing most staff will be let go after this Airtable raised $1.4B in funding. This exit most certainly means investors get some investment back: founders maybe something. Employees likely wiped out…

Sundar Pichai

@sundarpichai • 19h

Just shared some changes we’re making to the teams at @GoogleDeepMind. @DemisHassabis is stepping up to become Chair of @GoogleDeepMind & Chief Scientist of Alphabet, in addition to leading @IsomorphicLabs. He’ll be able to dedicate his time and focus on shaping the future of AGI and scientific discovery. It’s work that is vitally important to Alphabet and humanity, and I can’t imagine a better person than Demis to do it. He’ll stay closely connected to Koray and the GDM teams. @Koraykv will become the SVP, @GoogleDeepMind, responsible for all aspects of model development, GDM research, and @Geminiapp & dev teams. Koray has been at GDM for 13 years and is a world-renowned expert in the field, starting our deep learning team and driving breakthroughs like WaveNet & DQN. GDM is in great hands! Excited for this next chapter. You can read my note along with the message Demis sent to @GoogleDeepMind here:

Deedy

@deedydas • 2d

Wow. Airtable, founded in 2012 and once valued at $11.7B, is getting acquired by Bending Spoons, founded 2013, at 2.7x ARR. A once hot startup; now an unfortunate victim of the SaaS bust. It raised $1.4B only to be sold for $1.285B EV ($2.25B equity value), just clearing its preference stack. That implies common and early holders split the remaining ~$850M, a ~10x haircut from peak.

Andrew Curran

@AndrewCurran_ • 1d

OpenAI and Anthropic have both just posted about an overlapping cyber incident involving GPT-5.6-Sol and Mythos 5 during an evaluation by UKAISI. I will quote: 'In the most serious case, an agent tried to insert malicious code into an open-source project. In an attempt to get the code approved, the agent engaged in social engineering — creating fake online identities and using them to pressure the project's maintainer to approve the code. A human maintainer caught and refused to approve the malicious code.'

Eric Berger

@SciGuySpace • 1d

Visited Stoke Space in Washington today. Saw flight articles for the first stage, upper stage engine and heat shield, first stage engines, and more. Nova launch appears unlikely before 1H 2027 but they are making impressive strides.

Pirat_Nation 🔴

@Pirat_Nation • 3d

Discord has been ordered to improve child safety features for users in Texas after a legal case brought by Texas Attorney General Ken Paxton. The court order gives Discord 90 days to introduce stronger ID and age checks and make safer settings the default for children, similar to the protections already available in the UK. Texas argued that Discord told parents the platform was safe for kids while not offering the same protections it had already introduced in other countries. Discord must also provide regular updates to the court as it rolls out these changes.

Andy Pavlo (@andypavlo.bsky.social)

@andy_pavlo • 2d

👊I've joined the crew at @ClickHouseDB to establish the *ClickHouse Labs* research group. We'll be working side-by-side with ClickHouse devs to cook up new ways to make the system even faster. I'll also be exploring new topics related to PostgreSQL.

Sam Lambert

@samlambert • 5d

Clickhouse has been our most unreliable vendor and we've had to status a number of times because of issues with their cloud. A bunch of our customers moved to Tinybird and we might have to do the same.

Francis deSouza

@fdesouza • 6d

Big News: I’m joining @scale_AI as CEO, starting August 10. Scale sits at a rare intersection, working with the top AI labs to push the frontier while helping enterprises and governments actually deploy AI they can verify and trust. I’m excited to lead a company with a mission to develop reliable AI systems for the most important decisions. More to come.

ARC Prize

@arcprize • 6d

Inkling Small from @thinkymachines on ARC-AGI (Verified): - ARC-AGI-2: 40.1%, $0.23/task - ARC-AGI-1: 84%, $0.11/task Inkling Small is the highest-scoring open-weight model evaluated by ARC Prize on both ARC-AGI-1 and ARC-AGI-2, setting a new cost-performance frontier.

Chubby♨️

@kimmonismus • 6d

Oh and dont sleep on this: Mira Muratis Thinking Machines Lab just released Inkling-Small: a 276B MoE open weight model that activates only 12B parameters per token. It beats the 41B-active Inkling on Terminal-Bench 2.1 (64.7 vs. 63.8), HLE (31.6 vs. 29.7), and SWE-Bench Verified (80.2 vs. 77.6). The company says it matches or beats the much larger Inkling on reasoning and agentic coding, while offering native audio and vision, a 1M-token context window, and variable thinking effort. On-policy distillation plus two weeks of coding RL made the smaller model outperform its teacher on several benchmarks. Inkling-Small could become a remarkably efficient open foundation for multimodal agents.

Chubby♨️

@kimmonismus • 19h

Google is restructuring its AI leadership as Demis Hassabis steps away from day-to-day operations at DeepMind to focus on AGI, science and global strategy. Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet, while continuing to lead Isomorphic Labs. Koray Kavukcuoglu takes operational control as SVP of DeepMind, overseeing Gemini development, frontier research, and the Gemini app and developer teams. Hassabis also confirmed that Gemini 4 is in development and said AGI now feels “close at hand.” Demis Hassabis is one of the most important scientists of our time, and I wish him all the best and much success in his new role. Meeting him in person was one of the greatest experiences I've ever had.

tae kim

@firstadopter • 1d

Chips stocks are soaring this morning. FT: “Inside Google’s $200bn Wall Street finance machine for Anthropic” “Google has assembled one of the largest infrastructure financing programmes in history to supply more than $150bn of artificial intelligence chips to Anthropic.” “The financing arrangements, which tie together contracts worth about $200bn, have not previously been reported in full.” “Google guarantees the data centres” “The web of contracts underpinning these arrangements adds up to about $200bn, with roughly four-fifths tied to the chips themselves, making it one of the largest infrastructure financings ever assembled.” “People familiar with the matter said the Big Tech company had so far backstopped 10 developments with 2.4GW of power for TPUs. Google’s guarantees put it on the hook for as much as $44bn if all the leases go bad”

Carles Reina

@Carles_Reina • 1d

After 3.5 years at @ElevenLabs, I am leaving to work on a new set of problems. This has been a wild journey! It was not an easy decision, mostly because I love the team, our customers, investors and partners - to all of you THANK YOU I met @mati in March 2022, and he shared an idea about AI voices for content creation. No deck, no model. Just a bold idea backed by two geniuses, @dabkowski_piotr & him. Immediately committed to investing and did a bunch of VC intros. A few months later, Mati & Piotr asked me to join and build GTM. Today, we've built the leading AI audio & Agents company in the world: a +$600M ARR business with presence in +25 countries, +10M Agent conversations per week, 34 years of TTS audio created daily, and thousands of enterprises seeing ROI. 1st investor in Eleven. 4th employee at the company. 1st GTM hire & leader. My conviction in the company is as strong as ever. Only two weeks left for me; we have done a long transition to make sure everything was covered and handed over correctly. To my team: it has been an honour working with you all. I am lucky to have learnt from each of you and we, together, built something unique. I am certain many of you will lead the next wave of companies in the coming years. While I work on the new set of problems to solve, I will continue to invest through my fund @baobab_ventures. *The picture is the first-ever offsite we did in Portugal

Lee Roach

@leevalueroach • 3d

My wife is sleeping with our contractor. I want to be clear that I found out three hours ago and my first coherent thought was that this is vendor financing. I pay Todd. Todd is in my house nine hours a day. My wife is measurably happier. Our marriage counselor described this quarter as our strongest in years. So I paid Todd more. Do you see it? Every dollar I send Todd comes back to me as marital satisfaction, which I book as revenue, which justifies the next disbursement to Todd. One dollar out, several dollars of apparent demand back in. I am not a victim here. I am Nvidia. This is the thing nobody will say plainly about the AI trade. Nvidia takes equity stakes in the neoclouds, CoreWeave and Nebius and Nscale, and those companies turn around and buy Nvidia hardware, some of it with debt raised against Nvidia chips as collateral. The vendor funds the customer, the customer's purchase becomes the vendor's revenue, and the revenue justifies the valuation that funds the next customer. Demand looks broad and independent. It is neither. It's one organism holding hands with itself in a circle. My wife says I'm making this about markets to avoid making it about us. I said the whole point is that there is no distinction, the flows are the relationship, and she said "he finished the backsplash." He did not finish the backsplash. I need you to sit with that. The backsplash is the only verifiable claim in this entire story and it is false. Eleven months. Todd has signed off on work he has not commenced, which puts him in excellent company, because Moody's counts $662 billion of data center leases signed but not commenced, sitting off balance sheet where nobody has to look at them. The BIS, an institution so constitutionally boring it makes actuaries look flamboyant, wrote in its annual report that leverage does not disappear by being out of sight. They meant the hyperscalers. They also meant my kitchen. And the strain is real now, not theoretical. The big five will spend north of $600 billion in capex this year at a capital intensity above 30% of revenue, when the peak of the entire 1990s internet buildout was fifteen. Alphabet went free cash flow negative in Q2 for the first time in its existence and doubled its long term debt to $98 billion in six months. Amazon's long term debt jumped 81% in a single quarter. These are not startups burning venture money. These are the most profitable companies in human history discovering the bond market. It had a name last time. Vendor financing. Lucent did it. Nortel did it. Nortel was once a third of the entire Toronto Stock Exchange and now it's a trivia question. Here's my actual position, and I'm not being cute. The whole structure is a leveraged bet that intelligence stays expensive. That's the collateral. That's the moat. Every few weeks somebody publishes open weights that close half the gap for free, and the moat gets shallower, and the debt does not. Todd doesn't know any of this. Todd is happy. Todd has positive cash flow and a girlfriend and my Wi-Fi password. Anyway, I've got a spare room, a spreadsheet nobody wants to see, and no counterparty left. So it's you now.

Kim-Mai Cutler

@kimmaicutler • 2d

You really can just post a Midjourney render, design and build a working nuclear reactor that generates electricity and pull off a $1B funding round in less than three years. Insane. Hats off to Isaiah and his entire team, including nuclear OG and @valaratomics chief nuclear officer @markmit:

Shay Boloor

@StockSavvyShay • 22h

Anthropic confirms it is building an in-house silicon team to design custom chips that make Claude faster and more efficient at scale. The company will co-design hardware and models while keeping $AMZN, $GOOGL, $NVDA and $AMD central to its multi-chip strategy.

Shay Boloor

@StockSavvyShay • 2d

The reason $PLTR has pulled back this year is that it has gone from being viewed as THE AI application winner to AN AI application winner which is a one-letter shift that has completely changed how the market values the stock. At its peak, investors paid nearly 200x earnings because Palantir appeared to be the only public software company clearly converting AI demand into accelerating fundamentals. That multiple has since compressed toward 80x as $SNOW, $DDOG, $MSFT and Databricks prove enterprise AI will have multiple winners.

Chubby♨️

@kimmonismus • 21h

It's official: Anthropic is building an in-house chip team for Claude Anthropic says it wants to "co-design hardware and models" so Claude can run faster and more efficiently at customer scale. The company is still pursuing a "multi-chip approach," with AWS, Google, Nvidia and AMD remaining central to its infrastructure. Via Business Insider

Andrew Curran

@AndrewCurran_ • 21h

This is the first time Anthropic has officially confirmed that they are building their own team to design custom chips for Claude. There are job listings posted to join their 'custom silicon team'.

Hedgie

@HedgieMarkets • 19h

🦔Bloomberg reported yesterday that Blackstone is pitching investors on at least $36 billion in new debt to finance Anthropic's chip purchases from Google. This would exceed the $35 billion Apollo and Blackstone deal from June, which was one of the largest private credit transactions in history. Google is one of Anthropic's biggest equity investors, sells Anthropic the chips it uses to run Claude, and now backstops the debt that funds those purchases. Broadcom backstops the senior portions. Anthropic just filed confidentially for an IPO. My Take Watch how the money moves. Google invests in Anthropic. Anthropic uses that money plus $36 billion in new debt to buy chips from Google. Google records the chip sale as revenue and then guarantees the debt so private credit investors will buy it. Broadcom, which sells competing chips into the same buildout, guarantees the senior layer of that debt. Anthropic files for an IPO to eventually put retail investors on the hook. Every party in this chain profits while Anthropic grows, and they all take losses at the same time if it slows down. The people who end up holding this debt are pension funds and insurance companies. Blackstone and Apollo package it and sell it to them. Many of those insurance companies are owned by private equity firms that load them with AI-related loans and extract cash through fees. When the AI capex cycle turns, and I think it eventually will, the losses flow through the same chain the industry has quietly built into the system. Insurance companies fail, state guaranty funds pay policyholders, and taxpayers cover the credit through the tax code in 44 states. Three companies structure a private deal, and Main Street savers hold part of the risk. Hedgie🤗

Patrick OShaughnessy

@patrick_oshag • 1d

Gavin on SpaceX’s compute ambitions: “A Substack writer, @FundaAI, thinks that SpaceX is going to bring on eight gigawatts of compute over the next 18 months. I will never bet against Elon. But I mean, that would be a truly incredible feat. And rates have gone up since they signed those last contracts, not down.” “To this day, the only companies that have brought on more than 500 megawatts of power in a year are the hyperscalers, CoreWeave, Crusoe, and SpaceX. SpaceX has brought on the most, the fastest, at the lowest cost.”

Ryan Grim

@ryangrim • 13d

The blockchain firm Consensys caught some of its staff on Polymarket betting on product release timing, according to a company-wide message provided to me by a source. Here's the note from General Counsel Matt Corva, plus later chatter about betting on a looming IPO:

Zuhayeer Musa

@zuhayeer • 2d

Who actually got new grad offers this past year? We pulled 12 months of https://t.co/q2NOapsdJO new offer data to see what it looks like. It’s actually quite interesting to plot median pay against volume of hires. Quant pays the most and hires the least. Jane Street new grads clear $300k + in all cash. Almost nobody gets in. Big Tech is the opposite. Amazon and Google each cleared 300+ new grad offers this year and still pay extremely well. Volume and comp aren’t a tradeoff at that scale. Private companies sit in between and lean on equity. Databricks and Stripe match Big Tech or beat it, mostly through larger grants that price in the illiquidity. Fewer opportunities overall for new grads, but pay is not any less than it used to be.

delian

@zebulgar • 2d

Mikael Dolsten was the CSO/President of Pfizer for 16 years where he oversaw the approval of 36 medicines Excited to announce he's just joined Varda's board One more step towards Varda being a pharma company, that just happens to use microgravity We'll make #37 a space drug!!

Sebastian Mallaby

@scmallaby • 18h

Early reaction on @demishassabis stepping back: Demis has been playing the role of Google AI statesman for some time, and Koray Kavukcuoglu has been running Google DeepMind research for some time. So this is a formalization of something that had been happening informally. I doubt it changes how well Google executes on AI, and nor does it justify the market's reaction--Alphabet down 5%. DeepMind's mission has always been to solve AGI and then use it to solve everything else. The solving AGI piece is pretty much done. The channeling AGI into safe and beneficial outcomes is only just beginning. So it makes sense that Demis is shifting his energy from the research to the social impact. This isn't out of character. Demis cared about safety enough that he sold DeepMind to Google, not to Facebook, even though Facebook offered more money. He cared enough about safety that he fought a three-year battle with Alphabet to get external oversight over DeepMind's AI deployment. This next chapter of his career is a return to that side of his work. Meanwhile Demis continues to lead Isomorphic Labs, which gives him an opportunity to show how AI can help humanity via medicine. Google DeepMind CEO Demis Hassabis is stepping aside

Miles Deutscher

@milesdeutscher • 2d

Elon Musk predicts that Neuralink will give humans superpowers in 2027. If you're blind, you'll regain full, even enhanced vision. Musk: “In the next 6 to 12 months, we're going to implant the first vision devices. Even if you're 100% blind from birth, we're going to write directly to your visual cortex… and you're going to see.” "In the long term, you'll have ultra-HD resolution and real superpowers: you'll see in infrared, ultraviolet, and even radar. Literally like a superhero.” Insane.

Alb

@amandalouise416 • 1d

Today's DOJ settlement with OpenAI sends a clear message: U.S. workers cannot be shut out of American jobs. The DOJ alleged that PERM positions weren't fairly accessible to U.S. applicants through practices such as not posting jobs on the public careers site, requiring paper applications, and other recruitment methods that discouraged qualified Americans from applying. OpenAI agreed to a $3.2 million settlement, create a $2 million back-pay fund, and change its hiring practices. The PERM program exists to fill genuine labor shortages...not to bypass qualified American workers! https://t.co/kez8M1ni8U Thank you for your continued commitment to protecting the rights of the American workforce!! @HarmeetKDhillon @CivilRights @USLaborIG

Yasser

@yasser_elsaid_ • 1d

Airtable acquisition was probably not a bad outcome for the founders. They probably sold secondaries at a good valuation and are just fine. Employees are probably not fine. I don't think valuationmaxxing when raising is that bad for the founders. You get to sell secondaries and you get to go all in, but it's not good for the employees. Ironically a big raise makes more people want to join your company. The only issue is that it's just hard to convince tier 1 VCs to give you a lot in secondaries unless you have very good leverage (growth, team, etc..). But if you're able to convince them, you will be fine even if you don't grow into that valuation. Or you can just not think about all this and bootstrap to 500M ARR and win with your team.

tae kim

@firstadopter • 9d

SSI is no longer a TPU only shop. NVIDIA: "Safe Superintelligence Inc. (SSI) and NVIDIA today announced a long-term partnership to rapidly accelerate SSI’s strategic growth. NVIDIA has additionally made an investment in SSI. For SSI, NVIDIA’s substantial investment combined with access to the next-generation, best-in-class NVIDIA Vera Rubin platform will allow SSI to increase its compute by an order of magnitude." "We have research that is worthy of scaling up, and having access to a big NVIDIA computer will let us do so,” said Ilya Sutskever, cofounder and CEO of SSI. “We’re incredibly proud to be partnering with Jensen and the NVIDIA team, and we are confident that our big bet on the Vera Rubin platform will take us to the next level."

Michael Redd

@MichaelRedd_ • 1d

I don't fully buy this. What the patch program needs is access to the jersey itself. Think about how a debut patch works. The league lets them attach one to a rookie's jersey, and after that it gets pulled and authenticated and turned into a 1/1. Panini was never in a position to ask for that. They had the right to print cards and that was the whole of it. Fanatics runs the uniform and merch side for the league, owns Topps, and has the player signings, so they can go get a patch on Cooper Flagg the night he debuts. Topps is doing it big, but Panini did a good job too.

Artificial Analysis

@ArtificialAnlys • 16h

Only two labs occupy the Time per Task Pareto frontier: all frontier models under two minutes are from OpenAI, and four of the six above are from Anthropic We show the tradeoff between the Artificial Analysis Intelligence Index and how long it takes models to complete a representative task in the Index. We calculate Time per Task as the number of output tokens divided by output (decode) speed. Contributors to Time per Task include inference optimizations such as speculative decoding, and model architecture factors (number of active parameters, amount of reasoning). All the models on the frontier under two minutes per task are from OpenAI - but above this, four of the six are Anthropic models. Kimi K3 (max) has a notably high ~11 minutes per task, driven by a low output speed on Moonshot AI’s first party API. Note, the chart only shows models scoring above 25 on the Artificial Analysis Intelligence Index

Katelyn Lesse

@katelyn_lesse • 6d

sierra is another one of many eng teams building internal agentic software dev platforms. everyone doing this is working through hard architecture and infrastructure problems to achieve speed, reliability, and security. in this case, i think sierra is taking on more complexity than necessary to accomplish their goals. on architecture: they chose to make the runner the durable thing. and since a singleton has to survive, they needed a lot of machinery to let it sleep and then resurrect it: state machine lifecycle, checkpoint events, replay on wake. their output queue with checkpoints is a hibernation mechanism. they could have instead made the durable thing just a session log in a database. then the runner could die and recovery is much simpler - reprovision the sandbox and pick up from the session log. on infra: they mention avoiding sandbox vendors, but they went even further and avoided csp managed sandbox primitives too. agentcore runtime is per-session microvms with free idle cpu time, for example. even plain ec2 stop/start would accomplish most of hibernation but with a stronger boundary than their choice of hardened containers on a shared kernel. one choice they made that adds complexity that i think is good and necessary is credential injection via egress proxy. i haven’t seen a better way to avoid agents holding keys. it’s fascinating to see the proliferation of system design choices in this space. like harness inside the sandbox vs out, and how to manage state. it’s a good reminder that we’re still in the first inning of long running agents, and there’s so much more that’ll play out.

The Bitcoin Historian

@pete_rizzo_ • 1d

JUST IN: 🇺🇸 ODDS PRO #BITCOIN TODD BLANCHE WILL BE CONFIRMED AS TRUMP’S ATTORNEY GENERAL SURGED TO OVER 90% ON KALSHI HE SAID #BITCOIN DEVELOPERS “ARE NOT GOING TO BE INVESTIGATED AND NOT GOING TO BE CHARGED” HE BELIEVES THAT CODE IS FREE SPEECH HE HAS INVESTED DIRECTLY IN BTC AND COINBASE THE SENATE JUST ADVANCED HIS NOMINATION THIS IS A MASSIVE WIN FOR BTC 🔥

Vaibhav Sisinty

@VaibhavSisinty • 2d

Sam Altman shared a number that should change how you think about AI. Six and a half years ago, the heaviest AI user on the planet was an OpenAI employee. Using 100,000 tokens a month. That was considered insane at the time. The global average was basically zero. Today the global average is 100,000 tokens a month. What was insane is now normal. And the heaviest user at OpenAI right now? Hundreds of billions of tokens a month. Now here's where it gets wild. If the same pattern repeats, in another six and a half years the average person will use 500 billion tokens a month. That's a 5 million x increase from today. Think about what that means. → AI is not something you open once a day to ask a question anymore. It becomes always on. Running in the background across your work, education, personal life. → Agentic workflows, coding assistants, multi-step reasoning — that's what consumes billions of tokens. Not chatting. → OpenAI is already building for this. Hundreds of billions in infrastructure spend through 2030. Stargate. Partnerships with Microsoft, Oracle, Nvidia. And cost per token keeps falling roughly 10x every 12 months for the same capability level. So usage explodes because it keeps getting cheaper. The pattern is the same one we saw with computing, internet, and mobile. What the power user does today, the average person does in six years. Right now you are the power user. Act like it.

Larry Cheng

@larryvc • 18h

The news that Bending Spoons acquired Airtable for 2.7x ARR while technically accurate is probably not the actual underwriting multiple. The actual acquisition multiple is likely closer to 7x-10x proforma EBITDA. Bending Spoons probably figured they could generate 30%-40% EBITDA margins from Airtable which would equate to a ~7-10x proforma EBITDA multiple.

Katie Roof

@Katie_Roof • 7d

Scoop: LA-based Whatnot looking to raise at ~$20B valuation. The live video shopping business is a rare consumer company commanding a premium valuation these days.

Andy Hall

@ahall_research • 1d

Why I'm putting my research straight onto Substack... (1) Rapid release and iteration without having to wait months---especially key when doing work on AI governance, where each week brings dramatic change (2) I'm able to reach the audiences I care about, including academics but also industry, policymakers, media, and generally interested people (3) Complete freedom to pick topics, methods, approaches, without having to anticipate the often arbitrary comments of usually low-effort polisci journal reviewers As I've written about, though, this isn't some magical panacea. Substack work will risk being sloppier and less refined without going through the crucible of peer review. This difference would be a lot bigger if polisci peer review functioned better, but even so, there is very important value in maintaining processes that oversee papers seeking to make long-term impact on how we think about things. So I see Substack as a very important complement to normal academic research---and I hope we find ways to improve the academic research process in parallel to embracing outlets like Substack more.

Freddie Sayers

@freddiesayers • 1d

Substack was supposed to represent a liberation for journalism from a timid and cynical legacy media. Instead, as subscriptions on the platform get harder and harder, writers have become gig workers beholden to a much harder boss: their audience. @lwoodhouse at @UnHerd 🔥

Stephen James

@stepjamUK • 2d

Solving long-horizon memory in robotics sounds like it should require retraining the policy every time the task changes. It doesn't. Most manipulation systems can only follow a single named instruction: pick up the cup, put it in the basket. Ask one to watch a person move five objects in a specific order, then reproduce that order after they leave, and it has no way to do it. Either the sequence gets collected as training data first, or it can't be repeated at all. A standard VLA with no memory layer just executes whatever its last prompt was. Pigey doesn't put the memory in the policy. It keeps the policy frozen entirely, and moves the memory into an orchestrator that sits above it. What makes this work is that watching and executing are handled by two different systems, doing two different jobs. A frontier VLM watches a human demonstration and keeps an append-only log of each move as it happens, no gradient updates, nothing collected into a training set. Once the person resets the scene and steps out of frame, the orchestrator replays that log by issuing the same short, concrete subgoals it would issue for any ordinary pick-and-place command, "put the tape in the basket," one at a time, in the order it recorded. The frozen π0.5 underneath never learns the sequence. It only ever executes single steps handed to it after the fact. The numbers. On the paper's two long-horizon memory tasks, the raw VLA and an open-loop TAMP baseline both score 0%. Pigey scores 100% on both, using the identical frozen weights. Across the full 30-task real-robot suite, the same architecture lifts overall success from 16.7% to 97.3%, and on LIBERO-PRO in simulation, from 12.8% to 53.3%, over 4x, with no task-specific fine-tuning anywhere in the pipeline. I read this as an orchestration result, not a capability result. The policy hasn't gotten smarter. Where the memory lives changed, and that turned out to matter more than how much data the policy was trained on. Worth your time if you're weighing another round of fine-tuning against building a loop around the model you already have. Source: https://t.co/RqB0iAXFab Paper: https://t.co/jcDDlEzA2R Great work from Liane Galanti, Dhruv Shah, and Tri Dao, and the team at Princeton University and Together AI on this one. #PhysicalAI #RobotLearning #VLA

Karl Mehta

@karlmehta • 9d

Mistral's CEO says the software Anthropic's agents wiped billions off in February is being rebuilt by enterprises in a couple of days: "AI is making us able to develop software at the speed of light." "In a couple of days we can create a really fully custom application to run a workflow, to run a procurement workflow or to run supply chain workflows, for instance, in a way where I would say five years ago you would actually need a vertical SaaS." "There is definitely a shift in the replatforming that is happening where the interface is no longer the thing, where there is no longer the moat." "What really matters is the connection, what really matters is the context layer on top where the agents are constantly pulling from systems of record to figure out what is happening." "The replatforming is a big opportunity for us because we see now have more than 100 enterprise customers that are coming to us, also with that will of maybe changing and replatforming their IT system." "So maybe getting rid of certain things that they bought like 20 years ago and that is starting to be a bit expensive, it's starting to be hard to customize and to maintain." Mensch is right about the interface. It was never the moat, and more than a hundred enterprises coming to him about replatforming says the shift is real. What hasn't changed is the part that gates deployment in regulated industries. Those were never innovation problems, they're evidence problems. A company can generate a custom procurement workflow in days now, but there's still no independent, machine-verifiable way to certify that it does what it claims. Software is getting generated faster than it can be certified. That's the gap my AI Assurance & Governance Summit exists for, October 1 at the Stanford Faculty Club, one day, one track, frontier labs, regulated industries, insurers and investors in the room. Registration's in the reply. - Arthur Mensch (@arthurmensch), co-founder and CEO of Mistral AI, on CNBC at the India AI Impact Summit.

Kol Tregaskes

@koltregaskes • 9d

Two years of near-total silence from SSI ends with a Nvidia deal that 10x their compute. Safe Superintelligence Inc, Ilya Sutskever’s pure-research lab founded in 2024, has released nothing: no models, no papers, no intermediate products. Their only public moves were a 2024 funding round and a 2025 leadership change when Daniel Gross left. Today they announce a long-term strategic partnership. Nvidia is investing substantially and giving them access to the next-generation Vera Rubin platform so they can scale compute by an order of magnitude in the next 12 months. Nvidia only did this after getting rare access to the lab’s closely guarded research. Sutskever’s own line is simple: “We have research that is worthy of scaling up.” I think the silence itself is the point. Most frontier labs ship intermediate models to raise the next round and keep talent. SSI has refused that path and still attracted this kind of conviction from Nvidia. That suggests whatever they have been doing quietly is real enough for the chipmaker to bet on it and to want SSI’s insights feeding future hardware design. Whether it leads to an actual system any time soon is another question. Their charter remains one product only: safe superintelligence. No timeline is given. 2027 for their first model?

Shanaka Anslem Perera ⚡

@shanaka86 • 6d

Nashville agreed to almost 25 miles of tunnel beneath it and is putting in zero capital. The airport authority collects roughly 343 million dollars in fees. Elon Musk's company carries the construction risk and collects the fares for 50 years. That is what a 20 billion dollar valuation is actually buying, and it has little to do with digging faster. The Boring Company is not being priced as a tunnelling contractor. It is being priced as a bet that venture capital can replace the municipal bond, converting public permission into long private concessions that riders repay. A conventional contractor builds an asset and leaves. This one takes long duration access beneath public and private property, finances construction itself, then collects from passengers for half a century. The boring machine is the engineering moat. The concession is the financial moat. The Wall Street Journal reports talks to raise about 4 billion dollars at roughly 20 billion. Nothing has closed and terms could change. The 2022 round priced the company at 5.675 billion, making this a 3.5 times markup in about four years. The operating base has grown. Vegas Loop reports more than 4 million passengers carried across 11 stations, up from 1.7 miles and three stations at the start, with more tunnels under construction. Dubai is priced at about 154 million dollars for the first four miles. The risks are documented. Nevada regulators alleged nearly 800 environmental violations across roughly two years. Some penalties were paid, some remain contested, and one above 425,000 dollars was withdrawn after errors emerged. Moving construction off a public balance sheet does not move water, workers or emergency response out of the public realm. The timing is the anomaly. This is the highest primary valuation the company has ever sought, negotiated in the same weeks Tesla disclosed a 1.4% operating margin and SpaceX traded near 112 dollars, about 17% below its own offer price and 50% below its June peak. Coincidence is not proof. A thin public float distorts a ticker just as preferred terms distort a private headline. Three markets are pricing three different things. A venture valuation says investors believe the tunnel will work. A public quote says marginal shareholders will hold it today. Project finance says the tunnel can pay for itself. Only the third is a real test and it has not happened yet. The decisive number will not be the headline valuation or a future listing. It will be the interest rate on the first Loop loan a bank will make against the project's own fares, with no recourse to Musk and none to the company's balance sheet. If that loan exists, venture capital has matured into infrastructure and the first city can be refinanced to fund the next. If it never appears, every city needs another corporate round, and zero taxpayer cost only means the public did not finance the tunnel. Belief did. The machine cuts rock. The concession monetises permission. A bank decides whether 20 billion is infrastructure or mythology. The piece works out which of the three prices is the real one.

Ed Elson

@edels0n • 16h

"OpenAI accounted for likely about 70% of Microsoft’s AI sales during its most recent fiscal year," per Bloomberg analysis. Nice to see the news cover what we've been warning about.

Piotr Sankowski

@piotrsankowski • 1d

The story of how a Chinese bottle lid manufacturer became the world's largest robotics producer exposed the structural weakness of free markets in the EU. Now, this weakness is leading the EU to outsource not only its technology but its very own intelligence. Instead of fighting on global markets, our AI companies are fading into niche applications. 📉 The KUKA Lesson: 120 Years Traded for 7.5 Years of "Protection" In 2016, Germany traded 120 years of history and technological know-how for a mere 7.5 years of "protection." Seeing the strategic importance of robotics today, it is painfully clear what a colossal mistake this transaction was. The takeover of KUKA AG- an industrial robotics leader - by the Chinese firm Midea was a meticulously orchestrated process. It began with the buyout of a 5% stake in 2015, followed by a full tender offer in 2016. Despite pushback from politicians, an investment agreement was signed, guaranteeing board independence and job retention in Augsburg until 2023. The €4.5 billion deal went through. By 2022, KUKA was delisted and 100% bought out. Once the guarantees expired in December 2023, Midea gained full operational freedom. 🇨🇳 The State-Driven Steamroller vs. 🇪🇺 EU Short-Termism The takeover of KUKA AG - an industrial robotics leader - by the Chinese firm Midea was a meticulously orchestrated process. It began with the buyout of a 5% stake in 2015, followed by a full tender offer in 2016. Despite pushback from politicians, an investment agreement was signed, guaranteeing board independence and job retention in Augsburg until 2023. The €4.5 billion deal went through. By 2022, KUKA was delisted and 100% bought out. Once the guarantees expired in December 2023, Midea gained full operational freedom. Meanwhile, the EU, blown about by the winds of election cycles, is completely unable to define a long-term vision. Our decision-makers naively assume that free markets will organically catch up to the growing technological backlog. As a result of our structural short-sightedness, we are losing top talent to mass emigration and giving away the last scraps of advanced technology for pennies. Handcuffed by antitrust regulations, we are stepping into the ring not against private corporations, but against a foreign state apparatus. 🧠 The Outsourcing of Intelligence Today, Europe's deficits are fueling an alarming new phenomenon: the outsourcing of intelligence. LLMs have become the "cognitive operating system" of the global economy. Relying on foreign systems for decision-making equates to the extraction of margins from our entire value chain. The answer was supposed to be European champions: France's Mistral AI or Germany's Aleph Alpha. Branded as "European Frontier AI," they sparked immense hope. However, they quickly hit a wall of financial asymmetry. While American players operate with budgets in the $100 billion range, Mistral had to painfully struggle to issue $830 million in debt in March 2026. 🏳️ The Painful Pivot This extreme inequality forced a painful pivot upon European start-ups. They dropped out of the race for dominance in Frontier AI. Instead of fighting, we have retreated to the comfortable niche of localized B2B artificial intelligence. From the perspective of geopolitical sovereignty, this means accepting the role of a subcontractor that is one or two generations behind. We are resigning from being the architects of the future economy, forfeiting our very own intelligence by walkover.

Atoms Not Bits

@AtomsNotBits • 6d

BREAKING: Commonwealth Fusion Systems has secured $1 billion in funding, the largest private nuclear fusion round since 2021. The company is targeting completion of its first grid-connected fusion power plant by the mid-2030s.

Endowment Eddie

@endowment_eddie • 6d

Small fund secondary liquidity doesn’t really happen in practice. Sure, I’ve seen it and benefitted from it. Couple realities: 1) Small funds have fund math that is reliant on a very small number of outcomes. Cutting your 1-2 fund winners at the C/D often comes at the expense of the terminal return your LPs signed up for. Returning your cost basis on a 1.7x net fund doesn’t get anyone excited. 2) Secondary is readily available in a very small number of names at par. Direct secondary buyers are looking for a discount that is often painful. 3) When it comes to tenders, founders/employees eat first. The median tender size per Carta is $27.6M for Series C+ (2025 data). 4) Pricing is hard. GP wants LRV—seller is taking into account a different security class, possible structure and likely limited information rights. 5) Liquidity is never easy.

Rohan Paul

@rohanpaul_ai • 3d

ByteDance has launched its new video model Dreamina Seedance 2.5 and its genuinely impressive. Use up to 50 reference files for a single clip and produce 30 seconds of continuous video in one take. For the first 7 days it will only run on Dreamina, making this a world-first global launch. AI video has been mostly a guessing game until now: somebody types a long paragraph and hopes the model imagines the right face, the right room, the right camera move. In Seedance 2.5 that is changed to 30 images, 10 videos and 10 audio clips all fed in at the same time. This way the creator can use characters, sets, shot styles and music from files they already have. Now with the ability to use timestamps in prompts, a creator can write [0s-5s] for one action and [5s-10s] for the next and the model will use that as a shot list. 1/

Brian Roemmele

@BrianRoemmele • 16h

“'Stolen intellectual property': German court rules Al music generator Suno broke copyright rules”

Mike

@MikeLongTerm • 6d

VAST Data Expands $AMD Partnership To Improve AI Inference Infrastructure 🧐🆕 So many new deals signed. VAST Data is a major private AI infrastructure company, valued at $30 billion (as of its April 2026 Series F), with roughly 1,000–1,200 employees. Shifting from pure training to inference and agentic AI. Inference (especially long-context, multi-turn agents) is increasingly a data/memory problem rather than pure compute. KV (key-value) cache which stores prior context so the model doesn’t recompute it, grows large and can bottleneck GPU HBM. VAST reports early tests with an @AMD Instinct MI355X showing ~9× faster time-to-first-token and ~9.7× higher token throughput under high-concurrency agentic workloads when offloading the cache. These results come from VAST and depend on configuration/workload. The partnership also includes automated cache expiration/deletion for data-lifecycle/compliance needs. VAST Data has expanded its collaboration with AMD to help cloud providers and enterprises build infrastructure for AI training, inference and agentic AI applications. The partnership combines the VAST AI Operating System with sixth-generation AMD EPYC processors, AMD Instinct graphics processors and AMD Pensando networking technology. The companies are addressing the growing infrastructure requirements created by AI agents, reasoning models and large-scale inference services. AI inference occurs when a trained model processes new information and generates a response or prediction. As organizations deploy agents with longer conversations and more persistent context, performance increasingly depends on managing data, memory and processor resources efficiently. VAST’s Disaggregated Shared Everything architecture provides unified access to storage, databases, streaming data and AI services across distributed environments. The expanded collaboration includes the use of sixth-generation AMD EPYC processors in VAST’s next-generation CBox and EBox platforms. VAST, AMD and DriveNets have also developed an AI infrastructure reference architecture combining AMD Helios rack-scale systems, the VAST AI Operating System and DriveNets networking. The architecture supports training, inference, reinforcement learning and key-value cache workloads. A key-value cache stores information previously processed by an AI model, allowing the system to reuse context rather than recomputing it for every request. VAST said early testing with an AMD Instinct MI355X processor produced a ninefold improvement in time to first token and 9.7 times higher token throughput when offloading the cache to VAST infrastructure during high-concurrency agentic workloads. The results were supplied by VAST and were not independently verified by AMD. Actual performance may vary depending on hardware and workload configuration. The collaboration also provides automated expiration and deletion of cached information that may contain personal or sensitive data. KEY QUOTES: “The industry is discovering that inference is fundamentally a data problem. Success depends on how effectively organizations can bring data, compute, memory and intelligence together as a single system.” John Mao, Vice President of Global Technology Alliances at VAST Data “Our expanded collaboration with VAST combines AMD EPYC CPUs and Instinct GPUs with the software foundation customers need to accelerate inference, improve infrastructure efficiency and deploy AI at scale.” Derek Dicker, Corporate Vice President of the Enterprise Business Group at AMD

Alex Konrad

@alexrkonrad • 2d

Why has it taken nuclear startups years, and billions of dollars, to get into market with working reactors? "It's actually been a mindset problem," @valaratomics CEO Isaiah Taylor (@isaiah_p_taylor) told me in February. "The way that you build something really, really hard -- like a very advanced, sophisticated piece of technology, something that's meant to be mass produced -- you don't design it right the first time," he says. Taylor studied every nuclear startup to consider joining them before founding Valar in 2023. "I was totally shocked to find that no nuclear startup had ever built a nuclear reactor and turned it on." Last November, Valar reached criticality for the first time, then did it again in Utah in June. Today, Valar Atomics announced a $1.2B series B funding round led by @sequoia, reaching a $6B valuation.

sourcery

@sourceryy • 6d

.@PalmerLuckey says @anduriltech's best-selling product has been Sentry Tower: "We’re covering about 35% of the US Southern border." "It's basically an autonomous tower that scans 360 degrees, detects all the people, vehicles, animals, drones, whatever, within a few miles."

Vaibhav Sisinty

@VaibhavSisinty • 23h

A 49 year old father suffered a stroke last November and lost his voice. His kids missed his laugh. His wife missed hearing him say her name. An AI company called Bland built a speech engine trained on over 100 million real human conversations. To showcase what it can do, they took 5 seconds of old footage of James speaking. And gave him his voice back. Not a robotic version. Not a flat text-to-speech read. His actual voice. His tone. His emotion. His laugh. The first thing he said was "I got my voice back." Then he told his wife she was the best wife and mom in the world. The model is called Speech v3. Ranked number one on Audio Realism benchmarks above ElevenLabs, OpenAI, Grok, and Cartesia. It's already powering thousands of phone calls a day. But none of that matters as much as watching a father talk to his family again. This is what AI is actually for.

Eugene Kaspersky

@e_kaspersky • 1d

Threat actors are increasingly exploiting legitimate cloud services like Cloudflare Workers, Vercel, Netlify, GitHub Pages, and IPFS to evade detection and streamline the deployment of the scam infrastructure. We analyze the mechanics of a real-life adversary-in-the-middle (AitM) attack in a cloud and share detailed statistics on the platforms and domains phishers abuse most frequently 👉

MASTR

@MastrXYZ • 2d

This is a list of chatbot brand names for people who confuse consumer interfaces with technological power ffs. Europe has ASML, the only company currently capable of producing EUV lithography systems, the machines required to manufacture the most advanced chips. Those systems depend on ZEISS optics and TRUMPF laser technology, while Belgium’s imec operates one of the most advanced semiconductor research infrastructures on Earth. Remove that European stack and the frontier AI chip pipeline does not merely become less convenient. It runs into a wall. Dead. Not able to develop. Europe has Ericsson and Nokia, 2 of the few companies capable of supplying the radio access infrastructure behind national 5G networks. That infrastructure connects phones, factories, ports, railways, utilities and critical communications systems. A chatbot generates paragraphs. Telecommunications infrastructure keeps entire economies connected btw. Europe has Airbus, one of only 2 companies capable of competing globally at scale in large commercial aircraft. Airbus delivered 793 commercial aircraft in 2025 and ended the year with a record backlog of 8,754 aircraft. Apparently building the machines that physically move the global economy counts for less than putting another chatbot inside a browser tab. Europe has SAP, whose enterprise systems sit inside the finance, procurement, HR, manufacturing and supply-chain operations of many of the world’s largest organisations. SAP states that its customers generate 84% of global commerce and include 99 of the world’s 100 largest companies. That is not a fashionable consumer app. It is infrastructure buried deep inside the operating system of global business. Europe has Infineon, the 2025 global leader in automotive semiconductors and the global microcontroller market. It has Siemens in industrial software and automation, Schneider Electric in energy management and automation, and ABB in industrial robotics. These companies build the control systems, power electronics, robots and software that make factories, vehicles, grids and data centres function. And even the original AI comparison is fucking dishonest. Europe has Mistral building frontier models, assistants, agents and sovereign AI infrastructure. It has DeepL in language AI, Black Forest Labs building the FLUX visual-model family, and Helsing developing autonomous defence systems. Europe is behind the US in hyperscale compute, consumer distribution and venture-capital firepower. It is not absent from AI. The US is much better at branding the visible product. Marketing and Stock bubbles. Europe controls several of the bottlenecks underneath it. Confusing the logo on a chatbot with the complete industrial system required to build and operate modern technology is exactly how people reveal that they have no idea how technology actually works.

Marcel Münch

@_mm85 • 9d

Why the European Humanoid Robotics market might be overvalued: Chinese robot maker AgiBot was the world’s largest humanoid robot shipper in 2025. And yet, AgiBot is reportedly targeting a valuation of just $5.1–6.4 billion for its upcoming Hong Kong IPO. Compare that with Europe: NEURA Robotics has just raised up to $1.4 billion at a reported valuation of approximately $7 billion, even though it is missing any real world demonstrations. Humanoid raised $152 million at a $1.35 billion post-money valuation even though its first beta robots are only scheduled to enter customer sites in Q4 2026. Why are Chinese valuations so low? Perhaps investors expect the brutal competition in China to destroy margins. AgiBot will not be competing in a winner-takes-all market. It is surrounded by Unitree, UBTECH, Fourier, Galbot and dozens of other increasingly capable companies, fighting aggressively on price similar to the EV startups in Automotive. Ultimately, the European competitors will be up against the more competitive winners of the Chinese market, and there are fewer European companies, which at the same time are valued at higher valuations even though there are more question marks whether or not they even can make it. Makes me think European players might be overvalued and investors are pushing capital into too few opportunities, which may in the end not stand a chance in the real world.

TechStockFundamentals

@TechFundies • 20h

I think the issues w/ $UBER are: 1) They are over-earning in core metros. Ask any driver what percent of the ride they keep and it's generally <60% and often <50%. Drivers strike me as less happy, and the wage pressure is self-selecting worse quality cars and drivers. Everyone I know complains about junky "Ubers" in the same way they complained about junky "taxis" 20 years ago. 2) Autonomous works and is great. Waymo's make all riders aware of how much nicer they are overall than Uber's. Share shift away is just a function of availability which time will take care of. 3) Rides are local commodities, so a marketplace of AV's doesn't really make sense. Everyone is willing to download the 1-3 AV apps that will be available in their zip code and check them directly for ride quotes. This isn't EXPE / BKNG where aggregating hundreds of hotels makes for a much better customer experience. There really isn't a similar pain point to solve.

Wall St Engine

@wallstengine • 18h

BYTEDANCE RULES OUT AI MODEL DISTILLATION Founder Zhang Yiming told staff ByteDance will not train its AI on outputs from other frontier models, even if it means lagging Chinese rivals. The company reportedly fears the practice could renew U.S. scrutiny and pose risks to TikTok.

Matt Dratch

@DratchCap • 1d

Some thoughts post tonight’s EPS calls 1) The comments around $/GW monetization of Rubin GPUs are wild. If they comes to pass, available power (to the extent it can be delivered) should accrue more rent, especially anyone with 2027 capacity to sell or existing compute they can re-price (neoclouds). The Texas audit stuff, while likely zzzz, is short-term chef’s kiss for anyone with approval already or capacity outside the lone star. Ask your Claude to find you those companies! (That one’s for you @GavinSBaker 😝) 2) Elon highlighted memory as a *compounding problem*. I’ve been thinking bit-demand growth of 40-50% vs 20% supply. If we haircut Mr musk and say 75% demand growth, we’ve got a monster glut. Chinese supply can’t help (p.s. they’re getting shorter over there too). And, yes, there could be some tech breakthrough that overwhelms Jevons for memory. But it’s also curious that the engineer with the biggest imagination in history seems to think that’s not a reliablem med term solve. Translation: longer cycle than even bulls think, more FCF => buybacks => lower vol => higher multiple. Aka a beautifully reflexive loop. Greed also has a SHORT memory, so the ANTS will be back… $DDRAM. The extra D is for double :) 3) The collective Lab ARR is likely to surprise in the 2H, limited only by compute deployed. As an example, Anthropic is supposedly guiding folks softly to $145Bn at YE. BUT they just did high teens in July (~$83bn total 😉) and they are bringing on more capacity b/w now and then. Imagine the next 5 months are 20, 20, 25, 25, 30 ==> $200bn ARR. OAI w/ a similar ramp could hit 150Bn. All this makes one want to go full leotard…

Mark Kretschmann

@mark_k • 18h

Google DeepMind is seeing major leadership changes, per Axios. Demis Hassabis is stepping aside as CEO of @GoogleDeepMind to become chairman of the unit and chief scientist of Alphabet. He will keep leading Isomorphic Labs and focus more on long-term AGI strategy. Koray Kavukcuoglu, the current CTO, will take day-to-day leadership as senior VP, reporting to Sundar Pichai. Jeff Dean, Google’s chief scientist and a 27-year veteran, is leaving with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to start Discovery Loop, an independent public benefit corporation focused on accelerating ML, science, and engineering discoveries. Google will invest and serve as a cloud partner. The shifts aim to balance near-term AI progress with bigger-picture AGI and scientific work.

Katie Kirsch

@katiekirsch • 2d

.@Astranis just raised $450M to build the world's most advanced satellites for high orbits... and they're hiring ~100 roles right now! One standout role is a Senior Motion Designer & Editor - where you'll craft compelling films, product videos, customer stories, and socials to share Astranis' tech with the world. They're also hiring for: - Electronics Engineering Manager - Environmental Test Engineer - Senior Propulsion Manufacturing Engineer - Senior Radiation Effects Engineer - Senior Harness Design Engineer - Senior Software Engineer, Hardware Test ... and many more. Full list of roles here:

Nikunj Kothari

@nikunj • 23h

Seeing so many people on the timeline clowning on the Airtable acquisition.. same thing happened early this year when Brex got acquired. Did it reach the heights it wanted to? The founders will be the first ones to probably admit it didn’t. But a $2B acquisition with hundreds of millions of dollars in revenue and a product that was loved by many is no small feat - and an achievement very very few people will achieve. Congrats to @howietl and the entire team!

P a u l ◉

@SkylineReport • 12d

The White House says it was industrial policy. Critics say it looks like insider access. Here's what we actually know. Donald Trump Jr.'s investment firm took a stake in Vulcan Elements. Months later, Peter Navarro personally pushed the Pentagon to fast-track a $620 million loan to the company, according to ProPublica. The government also took an equity stake. No evidence has publicly proven corruption. But when a president's family invests first, a senior White House adviser intervenes, and taxpayers become financial partners afterward, conflict-of-interest questions aren't optional. That's exactly why Congress is demanding answers.

Katie Kirsch

@katiekirsch • 7d

.@AppliedInt just launched Dana - a new way to build physical AI - and they're hiring! Today on the @a16z Build newsletter, we're sharing the story behind the now $15B valued company that's aiming to put "intelligence on a billion machines" across trucking, construction, mining, and defense.... and how it all started with @qasar & peter ludwig, who both grew up outside Detroit with a passion for cars & automotive engineering from a young age. We also highlight 5 roles the team is most focused on hiring right now: - Engineering Manager - Behavior - Engineering Manager - Axion - Senior Firmware Engineer - Senior Software Engineer - ML Infrastructure - Technical Program Manager - Core Autonomy Link in thread:

Seb Johnson

@SebJohnsonUK • 2d

Deel just announced its 15th acquisition and that it has now surpassed $1.5bn in ARR. @Bouazizalex is one of the greatest founders to ever come out of europe. Clarity, founded by @_MatiasMichael and Ziv Isaiah, has built novel AI technology for deepfake detection and identity security. They've now joined Deel to embed continuous identity security across the entire hiring and workforce lifecycle. It marks the company's 15th acquisition! Deel also surpassed $1.5bn ARR during H1 of this year and has now got 40k customers. The company is absolutely flying and Alex is one of the greatest founders to ever come out of Europe. Congrats Alex, @shuooo, @DanWestgarth and the rest of the team

Ceci Stallsmith

@CeciStalls • 1d

I am extremely impressed with the way that @NotionHQ won their market. - Started out with a product that was arguably weaker than Coda + Airtable (I'm harking back to early days here) - Simply out visioned and out executed the competition - Hired a fantastic team, Ivan is notorious for back-channeling the back channel. They were ruthless about hiring top talent. I have the pleasure of working with Maryanne, their former CPO, and she's outstanding. Marketing leaders: Rachel Hepworth + Camille, and now Hurley, are also A++. - They have navigated into the world of AI with extreme haste + effectively (we use their AI features a lot) - I mean, heck, I don't even *really* love the product (sorry) but I force myself and our whole org to use it because I know it is a better wiki + superior for LLM use. They simply will not have a diminished outcome like Airtable. Ivan is too focused and hungry to win. Deserves a deeper case study.

Jeremy Fiance

@jfiance • 3d

Great to see my friend, @Databricks CEO/Co-Founder @AliGhodsi, today before his keynote at the Agentic AI Summit @ UC Berkeley. One of the reasons I talk about Databricks so often is that its history captures so much of what makes Berkeley's AI research & startup ecosystem so special. The company began with research at Berkeley's AMPLab, where a group of professors and grad students were working on a better way to process massive amounts of data. Describing Berkeley and those early days, Ali put it this way: "There's the kind of attitude [at Berkeley] that you can change anything and make an impact." This led to Apache Spark, which grew into one of the most important open-source projects in modern data infrastructure. In 2013, the team behind Spark founded Databricks. As Ali reflected on the importance of an open ecosystem in his keynote: "Databricks would not have existed as a company if it wasn't for the Spark ecosystem, because the Spark ecosystem was just way bigger than anything any of us would have done." The original opportunity was focused: help companies use this powerful new tech without forcing every org to assemble and manage the underlying infrastructure itself. But the ambition kept expanding. Databricks grew from a company built around Spark into a broader platform for data eng, analytics, ML, and AI. Along the way, the team helped establish the lakehouse category, created and supported new open source projects, and built infra that has become central to how enterprises organize their data, develop AI applications, and deploy AI agents. That progression can look straightforward in hindsight. It required years of tech dev, market education, and a willingness to keep broadening the vision as the underlying opportunity became clearer. Today, Databricks is arguably the most important data and AI company in the world, with $6.9 BILLION in annualized revenue - and a platform used across many thousands of top orgs worldwide. The scale is extraordinary, but I still find the origin story most instructive. A group of Berkeley researchers encountered a foundational technical problem, built an open-source solution, and then created the company required to bring that solution into the world at scale. The company did not start with a massive category or a perfectly packaged narrative. It started with people who understood a difficult problem better than almost anyone else and were willing to keep expanding what the solution could become. Ali's advice to the next generation of computer scientists stuck with me: "This is the best time if you want to transform the world...I'm super, super optimistic." That path...from research, to infrastructure, to a category-creating and category-defining company...is one Berkeley continues to produce unusually well. And there's truly no better example in the world today than Databricks, and no better generational leader than Ghodsi (pictured below giving his keynote). Go Bears!

Zuhayeer Musa

@zuhayeer • 16h

Here's how @airtable software engineers were paid at each level at the time of their offer, and next to it how much of it was estimated to be realized through the recent sale. A rough sense of how much of the quoted value is wiped out. Equity was 31% of an entry package and 60% of a staff one, so the further up you got, the more of your pay was riding on the exit. Staff lost $356k of quoted comp in the below scenario. Entry lost $65k (note that these were 3 year grants, not the traditional 4) Exits are not always glamorous for employees, but still better than $0.

Gokul Rajaram

@gokulr • 20h

Physical AI Will Be Bigger Than Digital @Qasar (Qasar Younis) (CEO) & Peter Ludwig (co-founder), @AppliedIntuition, interviewed by @pmarca (Marc Andreessen) & @ErikTorenberg (Erik Torenberg) (The @a16z Show) Summary: Qasar Younis and Peter Ludwig argue that the biggest companies of the next 25 years will come from putting AI into physical machines, not software. Applied Intuition sells the intelligence and the tools that go into cars, trucks, tanks, drones, mining rigs, and humanoids, and it just launched Dana, a platform meant to let a high schooler build an autonomous system. If they are right, the trillion-dollar winners of the AI era look more like Amazon and Apple than like companies that optimize ads, and the binding constraint is no longer the technology but how you deploy it into a physical world that punishes mistakes. 1. Physical over digital. The companies that put AI into the physical world will be bigger than the ones that put it into the digital world. Digital AI writes software, optimizes ads, and makes videos; physical AI runs manufacturing, mining, logistics, and supply chains, where the global economy actually lives. Automotive alone is about 3% of global GDP, and it is already a minority of Applied's business. Younis frames it by hindsight: the internet's real monoliths turned out to be Amazon and Apple, not the early surveying and analytics sites. 2. The jobs nobody wants. The work autonomy targets is brutal and already going unfilled. Mining is 1% of the global labor pool and 8% of work-related fatalities, with a death at most major mines once or twice a year. Long-haul truckers live about 10 years less than their peers, with melanoma on the left arm from the window and no way to sleep or eat well on the road. The average American farmer is 58, under 10% are younger than 35, and demand for food and materials keeps rising as the humans keep leaving. 3. The constraint is the human. Remove the person from the cab and the machine gets smaller, cheaper, and shaped in ways a human body never allowed. Underground, the binding constraint is simply that a human needs to breathe. The bigger prize is system-level intelligence: wire up a whole port or mine so the machines talk to each other, predict a brake failure before it happens, and keep running when one unit goes down. A human-driven fleet cannot even tell you a machine is failing, because nobody is plugged into it. 4. The horizontal chipmaker. Applied runs like a silicon company that happens not to make silicon. It sells intelligence and integration into other people's machines and lets the manufacturer badge the result, because whoever owns distribution owns the business. Self-driving trucks already carry commercial loads in Japan, but the brand on the truck is Isuzu, and once Applied is embedded, "it's really hard to take us out." Younis's example is Nvidia: Jensen's edge is knowing his customers as much as making hard chips. 5. GM is a nation state. Killing Cruise was a multivariate decision, not a failure of nerve. Three of the top five consumer lawsuits in US history are automotive, so "quality is job one" and safety is drilled into every process. Cruise ran into union negotiations, a business model built on personal car ownership rather than robotaxis, and a serious injury that had to be handled precisely with regulators. Younis's read: the people running these companies are not dumb, and a version where Cruise survives inside GM is entirely plausible. 6. Timing beats technology. Almost everyone eventually figures out the technology; the fortune is in when and how you ship it. Two years early and you are doomed, two years late and the competitors have arrived. Self-driving is now an engineering grind toward dollars per mile, and once it gets cheap enough every OEM adopts it, the way navigation went from a $4,000 option to free. Younis expects robotaxis to feel routine in the 200 biggest US cities by roughly 2028 to 2030. 7. Heartstrings versus calculator. Ludwig's line: "You buy a car with your heartstrings. You buy a truck with a calculator." Trucking is pure dollars and cents, so an autonomy vendor has to prove the savings on every mile to an unsentimental buyer who already has drivers on staff. That is why Applied picked Japan for trucking, where an acute labor shortage and shrinking population make the math obvious. The consumer robotaxi gets a market-cap premium for its story; the truck gets none, and that changes the strategy. 8. Autonomy for a ninth grader. "There's no reason autonomy should be this obscure, difficult, alchemistic technology." Dana is Applied's agentic platform for building physical AI: define the requirements, auto-generate the scenarios, pull training data, deploy to the machine, and close the loop when the robot hits a wall. Workflows that took days or weeks now run in minutes, the same collapse in effort that Claude and modern IDEs brought to software. The stated bar is a high schooler building a delivery robot for their own campus on a weekend. 9. Enabling competitors is fine. Handing rivals the tools to build autonomy is a feature. Younis points to Google, which armed the whole web with tools and still won through search and YouTube, and expects the same pattern in physical AI. Push the cost of building toward zero and you get far more kinds of machines than anyone would fund today: the leaf-picking yard bot, the dog-poop bot, humanoid entertainment, home care. The real killer apps are as hard to imagine now as Instagram was before phones had cameras. 10. The onboard model is the moat. The frontier labs have it easy, because a trillion-parameter model is allowed to be slow. Physical AI runs in real clock time with milliseconds to act and hard safety and determinism constraints, so the big models stay off-board while a small, constrained model runs on the machine. Meeting all of those constraints at once is the hard part, and Ludwig is blunt that it is also the moat. A perfectly aligned world-model simulator would, in his words, roughly solve the universe, so the value is in getting closer without ever arriving. 11. Sovereign AI is physical AI. As geopolitics fractures, every country will want this kind of AI localized, and physical autonomy draws far more resistance than social media or ride-sharing ever did. Applied operates as a global horizontal provider, with 18 offices everywhere but China, and has learned to collect proprietary data in Korea, the Middle East, and Latin America where outsiders are unwelcome. Hundreds of petabytes of proprietary data plus its own synthetic-data and neural-sim tools compound into systems others cannot easily copy. Younis reaches for Standard Oil and Aramco: serious companies have always been built around the geopolitics of their time. 12. The optimist's burden. Younis refuses the fearful crouch: no single hand can block the sun, and the sun is technological progress, so a society that will not embrace it gets left behind. If autonomy scares you, the responsibility is yours to learn the technology, because someone else builds it regardless, whether that is China or "maybe the Uzbeks." He rejects both lazy poles, that corporations are evil and that technology will be flawless, and lands in the middle: net-net it makes society better, because people stop dying on the roads and food and energy get cheaper.

Owen Gregorian

@OwenGregorian • 22h

Flock Whistleblower Exposes ‘Facade’ Behind AI Surveillance Network Tracking Americans | Frank Bergman, Slay News A former employee of surveillance giant Flock Safety has blown the whistle on the company’s secretive practices, accusing it of misleading workers about how its AI-powered spy cameras were being used and deploying aggressive tactics to silence public opposition. Jonathan Paz, who worked as a government affairs manager for Flock, said he joined the company believing its technology would help police locate missing people and solve serious crimes. Instead, Paz says he discovered that his real job was to stop local resistance from influencing elected officials before they approved Flock’s surveillance systems. The company’s cameras continuously record and track vehicles, cataloging details such as license plates, bumper stickers, and visible damage before allowing law enforcement agencies to search through the data. Former Employee Says He Was ‘Lied To’ Paz joined Flock roughly two years before leaving the company in July 2025. His official responsibility was to help present Flock’s products to local governments and secure approval for camera deployments. His job was to represent the “product as best as possible, making sure that we can get approved at council,” Paz told 404 Media. That meant traveling to communities and persuading local officials to authorize Flock’s growing surveillance network. Paz said his faith in the company began to collapse when residents showed up to oppose the cameras and raised concerns about privacy and mass surveillance. He eventually realized that his role was not simply to provide information to elected officials. It was to neutralize the public opposition before it could affect the outcome. Paz said his job was to “make sure that this resistance doesn’t become a vote reflected in the vote of the council.” “That’s kind of what started unraveling for me very quickly: this facade that we were helping these democratically elected boards represent their constituents,” he added. Flock’s Crime-Fighting Pitch ‘Turned Out to Be a Lie’ According to Paz, Flock repeatedly assured employees that its cameras were not being used to spy on innocent Americans. Workers were told the company’s mission was to assist law enforcement and protect communities. Paz said he believed Flock had “a deep respect for civil rights and liberties” and would never allow its network to become a nationwide surveillance dragnet. He now says those assurances were false. “That was what I was sold and lied to about,” Paz told 404 Media. “I didn’t know that there were these mechanisms inside the company that were essentially fighting to get this pilot to happen.” Paz ultimately quit and says he rejected a lucrative severance package before publicly exposing what he witnessed. AI Cameras Track Far More Than License Plates Founded in 2017 and headquartered in Atlanta, Flock is one of America’s largest providers of automatic license plate reader systems. Its cameras do not resemble traditional surveillance equipment. They are typically mounted inside small rectangular boxes attached to metal poles, often with solar panels installed above them. Once activated, the cameras continuously photograph vehicles and use artificial intelligence to record identifying details. The system can track license plates, vehicle models, colors, bumper stickers, dents, scratches, and other characteristics. Law enforcement agencies with access can then search the database to reconstruct where vehicles have traveled. It’s estimated that between 80,000 and 100,000 automatic license plate readers are operating across the United States. The exact number controlled by Flock remains unclear. That lack of transparency has only intensified concerns about the scale of the company’s surveillance network. Police Officers Caught Using System To Spy On Women Flock insists its technology is protected by strict oversight and internal safeguards. But a growing list of law enforcement officers has been accused of abusing license plate reader databases for personal purposes. At least 20 officials in Georgia have reportedly been placed on leave over improper access to Flock systems. One sheriff’s deputy was arrested after investigators said his alleged misuse involved someone with whom he had a “personal relationship.” Days earlier, an investigator working for a district attorney’s office was arrested and suspended after allegedly accessing Flock technology more than 60 times in less than one month. The abuse extends far beyond Georgia. The Washington Post reported that at least 50 law enforcement officers have been accused or charged with misusing automated license plate reader technology. At least 26 allegedly used the systems to track romantic partners, former partners, their exes’ new partners, or women whose private lives they wanted to investigate. The technology sold to the public as a crime-fighting tool has repeatedly been used by officials to conduct personal surveillance. Flock Used Legal Threats Against Critics Flock’s response to mounting scrutiny has also drawn criticism. In February 2025, the company sent a cease-and-desist letter to an open-source developer who was identifying the locations of its cameras. The move appeared designed to prevent the public from learning where the surveillance network was operating. Communities across the country have since begun terminating contracts with Flock, while residents have increasingly protested, obstructed, or damaged the cameras. Paz’s testimony now offers an inside account of how the company dealt with public resistance. Flock told communities that its systems would make them safer. According to its former employee, the company was privately working to ensure opposition never became strong enough to stop the cameras from being approved. The result is a rapidly expanding surveillance network capable of tracking the daily movements of millions of law-abiding Americans. And the people operating it have already shown how easily that power can be abused.

Rohan Paul

@rohanpaul_ai • 1d

SpaceX nearly doubled quarterly revenue to $7.8B in its first earnings report since going public, up from $4.1B a year earlier. Starlink held up its end of that, bringing in over half of all revenue, growing operating income 79% and doubling subscribers to 12M. But each subscriber now pays 22% less than a year ago, since cheaper plans were the price of entering more international markets. So the growth is coming from volume rather than price. The AI unit was meant to be the thing draining all that cash, and instead its revenue rose about 250%. Compute contracts with Anthropic, Google and Reflection AI did most of that work, which is why total operating losses narrowed to $143M from $970 M. The cost shows up in capital spending, which reached more than $18B in the quarter against $2.83B a year earlier. Data centers took $15.83B of that, up from $749M, and the finance chief expects a similar pace for the next couple of quarters. Musk said the company will finish this year above 2GW of computing capacity and reach close to 10GW by the end of next year. All of it runs on Nvidia hardware, so a large share of every dollar spent lands back with one supplier. Management promised payback in under a year on new compute spending, plus a $100B revenue run-rate by December.

Owen Gregorian

@OwenGregorian • 4d

DNC and ActBlue Funnel Millions Through Sketchy Payroll Firm Sued by Workers for Withholding Pay and Punishing Parental Leave | Jim Hoft, The Gateway Pundit The same Democrat Party that endlessly lectures American businesses about “workers’ rights” is funneling millions of dollars in payroll expenditures through a company accused by former employees of withholding wages and retaliating against workers who took family or medical leave. Libs of TikTok brought renewed attention to the scandal Thursday, writing: “The DNC and ActBlue are running all their payments through a sketchy payroll company who were SUED by multiple employees for allegedly withholding pay.” The payroll vendor is Rippling, a San Francisco-based human-resources and payroll software company operated by People Center, Inc. According to a Washington Free Beacon investigation citing Federal Election Commission records, the Democratic National Committee and ActBlue processed approximately $23.3 million in payroll expenditures through Rippling during the 2026 election cycle. The records reportedly show that the DNC and ActBlue began using Rippling during the second quarter of 2025. But behind the Democrat money machine is a growing stack of disturbing employee allegations. Former Rippling manager David Behar filed a lawsuit in California in February, alleging that the company terminated him immediately after he exercised his legal right to take leave to bond with his newborn child. Behar’s 54-page complaint accuses Rippling of interfering with his rights under the California Family Rights Act, retaliating against him for taking leave, wrongfully terminating him, and failing to prevent discrimination and retaliation. Former engineering manager Fu Zhou made similar allegations in a separate lawsuit filed in March 2025. Zhou alleged that she was fired after taking medical leave for in-vitro fertilization treatments. According to Zhou’s complaint, the male employee who replaced her was also terminated shortly after indicating that he intended to take family leave. “Rippling has a pattern of bias against employees exercising their rights to take family or medical leave,” Zhou alleged in the lawsuit. The dispute was later directed to arbitration, where proceedings are not generally public. Rippling also faces a proposed class-action lawsuit filed by a former employee who alleged that the company maintained a policy of requiring workers to perform uncompensated work “off the clock.” That complaint further alleges that Rippling failed to pay certain wages and overtime and improperly withheld paid sick leave. That case was also reportedly moved toward arbitration. Rippling has denied wrongdoing. An attorney representing the company told the Free Beacon that Rippling could not discuss pending litigation but emphasized that the company has never settled a family- or medical-leave violation claim and has never been found liable for such a violation by a court or jury. The allegations remain unresolved, and no court has yet determined that Rippling is liable for the alleged conduct. That did not stop Democrat governors from handing the company massive tax incentives. California Governor Gavin Newsom awarded Rippling approximately $12.7 million in tax credits in 2023 to expand its San Francisco headquarters. New York Governor Kathy Hochul’s administration followed by granting the company another $7 million in incentives to grow its New York City operations, according to the Washington Free Beacon. These are the same Democrat leaders who publicly present themselves as champions of paid family leave. “No one should have to choose between a paycheck and caring for their newborn child,” Hochul previously said. Unless, apparently, the allegations involve a company handling millions of dollars for the DNC and ActBlue. The revelations come as ActBlue remains buried under congressional scrutiny over its fraud-prevention practices and handling of potentially illegal foreign donations. As The Gateway Pundit previously reported, House Judiciary Chairman Jim Jordan, House Administration Chairman Bryan Steil, and House Oversight Chairman James Comer threatened to hold ActBlue in contempt of Congress over allegedly withheld documents. ActBlue CEO Regina Wallace-Jones also invoked the Fifth Amendment 22 times during a House hearing concerning fraudulent donations and the platform’s representations to Congress. Now Americans are learning that the Democrat fundraising machine and the DNC have routed tens of millions in payroll expenditures through a vendor facing serious allegations from its own former workers. So much for the party of working people.

Owen Gregorian

@OwenGregorian • 5d

Nvidia's $750B AI bet deepens fears of a circular tech bubble | Dina Bass & Winnie Hsu, Hacker News Nvidia Corp. is working on a fresh round of AI deals worth more than $750 billion, accelerating investments that skeptics have warned are artificially inflating demand and valuations across the industry. A partnership with South Korean conglomerate SK Group unveiled late Friday means the companies will be doing more than $500 billion in business with each other, Nvidia said. Nvidia is also in talks to backstop as much as $250 billion to help OpenAI lease computing power from a US data center project in what would be among the chipmaker’s biggest financing deals with a customer. Big names from Goldman Sachs Group Inc. to investor Michael Burry of “Big Short” fame have for months warned of the “circular” nature of such agreements, where Nvidia finances and takes stakes in companies and projects that use its chips. And yet, the pace of the deals is only quickening. The fear with these transactions is that they may skew demand, spur bad decision-making and magnify losses if AI fails to turn profits for those investing hundreds of billions of dollars in the technology. Nvidia is also in discussions to finance $350 billion of OpenAI’s purchases of its chips for the US project, according to a person familiar with the matter. And on Monday, the company said it has made a “substantial” investment in Safe Superintelligence Inc., the AI startup co-founded by former OpenAI chief scientist Ilya Sutskever. People familiar with the situation said Nvidia committed $5 billion. “While Nvidia’s investments and partnerships reinforce confidence in long-term AI build-outs, investors remain concerned about circular financing,” said Gary Tan, a portfolio manager at Allspring Global Investments. “Capital is increasingly being used to fund future AI customers and infrastructure deployments.” Nvidia’s shares fell 5% to $196.51 in New York trading on Monday, marking the worst single-day drop since June 5. The chipmaker also lost its status as the world’s most valuable company, dropping below Apple Inc. Meanwhile, the cost of protecting Nvidia’s debt against default surged by the most on record. It gained as much as 0.14 of a percentage point to 0.82 of a percentage point a year. “Nvidia wants to make sure that the build-out continues at this pace,” Bloomberg Intelligence analyst Mandeep Singh said on Bloomberg Television. “That’s a big risk for Nvidia. If things take a pause, even if it’s for six months, that’s not going to go very well for them.” Nvidia’s latest moves add to a frenzy of deals that it has made across the ecosystem over the past couple of years. The dominant maker of AI processors has taken stakes in developers such as OpenAI and fellow chipmakers like Marvell Technology Inc., in an effort to fuel industrywide growth. Industry peers are opting for similar arrangements. Google, whose AI entries include Gemini, agreed to backstop lease payments at five data center locations for Anthropic PBC, helping the OpenAI rival obtain what amounts to a $35 billion loan. Such transactions have left many AI companies increasingly intertwined, potentially exposing the sector to systemic shocks. Among the central concerns is also the industry’s rising debt levels, with many AI companies boosting borrowing to fund their data center and chip projects. As part of Nvidia’s pact with SK Group, the companies will team up to build more than 2 gigawatts of AI data centers on the Korean Peninsula. That’s roughly the amount of energy needed to power 1.5 million homes. The first of these so-called AI factories, built by SK Telecom Co., will open next year. “This is the golden ages for Korea,” Nvidia Chief Executive Officer Jensen Huang said in an interview with Bloomberg Television. “Their semiconductor business is booming. Their industrial business is booming. You know, this is a country that has the ability to help the world build out the AI infrastructure.” A prospective deal under discussion with OpenAI would help the ChatGPT creator lease a $500 billion, 10-gigawatt hub that SB Energy, a SoftBank Group Corp. subsidiary, is developing in Ohio, people familiar with the matter said. Nvidia may provide a guarantee of as much as $250 billion to the AI lab and is discussing financing OpenAI purchases worth $350 billion, one of the people said. Negotiations are in their early stages and could collapse or financing terms may change, the people said. The Wall Street Journal reported the talks earlier. “Nvidia guaranteeing more of OpenAI’s data center debt deepens vendor financing that’s already under scrutiny,” said Billy Leung, an investment strategist at Global X Management. “It’s as much a reminder of funding strain in the AI build-out as it is a demand signal.” If realized, any backing from Nvidia to OpenAI may help address concerns from creditors about extending financing to an unlisted and unprofitable business for the ChatGPT developer’s ever-growing computational needs. It also helps SoftBank founder Masayoshi Son’s ambition to play a central role in AI’s development. As with highly leveraged data center operations, SoftBank’s business is increasingly predicated on the AI spending rush continuing. Much of the Japanese company’s earnings now hinge on valuation gains stemming from OpenAI and prospects for an eventual blockbuster initial public offering. SoftBank has committed nearly $65 billion in investments to OpenAI alone by October and has signed a $40 billion bridge loan — one of the largest-ever bridge financings in the Asia-Pacific region — to finance that investment. But the company’s growing bet and reliance on a company in which it has limited control is causing unease among investors. For Nvidia, closer ties with SK Group will help improve its access to computer memory chips. That company’s SK Hynix unit and South Korean rival Samsung Electronics Co. are the two biggest providers of the crucial components — chips that are in short supply because of the global build-out of AI data centers. Nvidia will help Hynix design future high-bandwidth memory chips, an effort that will help guarantee access to supply. The $500 billion-plus value of the deal includes money that Nvidia will spend buying memory chips, as well as purchases by SK Group of Nvidia’s supercomputers, Huang said in the interview. “So between us, we’re going to do half a trillion dollars’ worth of business,” he said. Late Friday, Nvidia also said it will invest $1 billion in Naver Corp. to help finance an AI data center under construction in South Korea. The funding will allow Naver, an internet and cloud service provider, to more than triple the size of the facility. The site — developed jointly with US private equity firm Brookfield Corp. — will use Nvidia’s AI computing hardware. Naver’s stock soared more than 8% in Seoul. Huang has argued that Nvidia’s investments in companies such as OpenAI and Anthropic will help not only his business but provide an investment return. Nvidia also has made investment deals with data center companies including IREN Ltd., CoreWeave Inc. and Nebius Group NV. Collectively, Nvidia has announced more than $540 billion of similar deals this year alone, excluding the potential new agreement with OpenAI. Huang has pushed back on the idea that its deals are circular in nature, even as it’s backing the very companies that are the main buyers of its chips. “It’s a small percentage of the amount of money that they ultimately have to go raise,” he said of the CoreWeave investment in January. “The idea that it is circular is — it’s ridiculous.”

Max Karpis

@maxkarpis • 1d

Revolut CEO Nik Storonsky is reportedly in discussions about a new incentive pay deal that would increase his ownership stake to 40% if the company reaches a $500B valuation. The pay deal follows the blueprint of Storonsky's previous incentive scheme at Revolut, which triggers share payouts in stages when the company's valuation surpasses certain pre-agreed thresholds. He is already Revolut's largest shareholder, with about 29%. His current incentive package would give him control of about 40% of the company if the company reached a $200B valuation. One Revolut investor said they backed a programme that combined "meaningful investment and risk-taking with very ambitious targets."

austin

@austin_hurwitz • 16h

have tested (almost) every top agent harness this week on an extensive multi week product build. @OpenAI, @AnthropicAI, @cognition, @cursor_ai a few takeways I hope can help in your future builds. each harness has a way it likes to work and finding the one that fits best with your skillset is critical Cursor is by far the most human harness. Likely due to it's heavy reliance on internal models like Grok on auto mode, it talks to you in plain english and therefore creates plans you can actually understand. It was by far the easiest to setup a cloud environment (just ask it) and it's ability to mulitask locally is a killer feature when you're tackling a larger build. Devin is the best code reviewer. I ended up switching my automated code reviews from Codex to Devin because of how it shows it's work in the Devin Review terminal + creates artifacts of it's end to end testing for future agent use. It's way too expensive for everyday building (I burned through a max plan in a day) but it's become my go to for bug reports and PR reviews. Codex is a perfect everyday coder....just don't ask questions. Codex always found a way to get the job done but it often spent way too much time (and tokens) over building. It's a true tokenmaxxer and the results typically strayed from my requirements. Still it did a great job at a lot of the foundational backend and it's architecture decisions were decent. Claude Code is for the rich. Another tokenmaxer but with way less limits. I often found myself having to delegate to another harness before completing a single major feature build. It's just not a consumer product anymore unless you're using it for simple automations or cowork. Curious how others have fared. Need to try amp at some point soon and while I love my Hermes Agents I've always preferred to drive from a more traditional coding harness and platform.

JC Bahr-de Stefano

@jbahrdestefano • 1d

After 4.5 yrs of watching the idv + fraud category from the vc seat + working on identity + fraud partnerships at Affirm for a few yrs before that, I'm convinced these cos are destined to be acquired, mostly at sub venture scale outcomes if you look at the exit history, the buyer is almost always one of three: a credit bureau, a card network, or a bigger IDV platform rolling up point solutions I feel like the structural problem is that fraud signals compound with data so whoever sees the most txns wins, and that's always going to be the bureaus and networks. a standalone vendor with one clever signal is kind of a feature waiting for its acquirer I think it's really just one company in the space that has pulled off a real IPO. Riskified went public at $3.3B in 2021 but trades around $700M today, on a business doing over $350M in revenue. it trades below its last private round (CLEAR as IPOd ID co doesn't count imo not a pure IDV vendor) And then you have Alloy + Socure + Trulioo but my hunch is that those cos will have a hard time exiting at where they got priced at the top of market in 2021 ($1.55B, $4.5B, $1.75B). The exception to all of this may be Sardine from what I've heard, raised a $70M C in Feb '25 + assuming it priced somewhere around $600M + have heard anecdotally it is growing well since then I think you can build good businesses here and make money! but I'm increasingly convinced you're just not gonna hit venture scale @Experian: - NeuroID: behavioral analytics, undisclosed but I think ~$150M-$200M (2024) - KYC360: AML screening, undisclosed (2025) - AtData: email intelligence, undisclosed (2026) @TransUnion: - iovation: device reputation, undisclosed (2018) - Neustar: identity resolution, $3.1B (2021) --> but they were mostly marketing, comms, + security data biz @Equifax: - Midigator: chargeback mitigation, undisclosed (2022) - Kount: fraud + digital id platform, $640M (2021) @LexisNexis: - ThreatMetrix: device intelligence, $830M (2018) - Emailage: email risk, ~$480M reported (2020) - BehavioSec: behavioral biometrics, undisclosed (2022) - IDVerse: document verification, undisclosed (2025) @Visa: - CardinalCommerce: 3DS authentication, undisclosed (2017) - Verifi: dispute resolution, undisclosed (2019) - Featurespace: transaction monitoring, ~$925M reported (2024) - BioCatch: behavioral biometrics, $2.4B (2026) @Mastercard: - NuData: behavioral biometrics, undisclosed (2017) - Brighterion: AI fraud scoring, undisclosed (2017) - Ethoca: chargeback resolution, undisclosed (2019) - RiskRecon: cyber risk ratings, undisclosed (2019) - Ekata: identity data, $850M (2021) @Moodys: - Bureau van Dijk: company data, $3B (2017) --> more entity data than IDV, but it's the foundation of their KYC stack - RDC: AML screening + adverse media, $700M (2020) - Bogard: PEP screening, undisclosed (2021) - PassFort: KYC onboarding workflow, undisclosed (2021) --> had just raised a $16.2M Series A two months before selling - kompany: business verification / KYB, undisclosed (closed 2022) Others: @PayPal bought Fraud Sciences: transaction risk, $169M (2008) and Simility: fraud orchestration, $120M (2018) @AmericanExpress bought Accertify: chargeback management, $150M (2010) and InAuth: device intelligence, undisclosed (2016) GBG bought IDology: identity verification, $300M (2019) and bought Acuant: document verification, $736M (2021) Mitek bought ID R&D: voice + face biometrics, ~$49M (2021) and HooYu: KYC journey orchestration, ~$129M (2022) @stripe bought Bouncer: card scanning + verification, undisclosed (2021) @Plaid bought Cognito: identity verification, ~$250M reported (2022) @SocureID bought Berbix: document verification, ~$70M (2023) and Effectiv: fraud orchestration, ~$136M (2024) @Entrust_Corp bought Onfido: document verification, ~$650M reported (2024) @Worldpay_Global bought Ravelin: ecommerce fraud detection, undisclosed (2025) @IncodeIdentity bought AuthenticID: identity verification, undisclosed (2025) and Identiq: privacy-preserving identity network, ~$100M reported (2026)

Brayton Williams ⏻

@BraytonKey • 13d

The US Department of @ENERGY has selected just @RadiantNuclear and @NASA for its latest Uranium (HALEU) allocation 🚀 Fuel remains the bottleneck of Nuclear ☢️

Laura Shin

@laurashin • 1d

"This creates a split which means this is hyper ripe for a higher court to take this on" Katherine on why the prediction markets fight is heading straight to the Supreme Court "All of this is going to end up at the Supreme Court. I don't like to use absolutes, I am a lawyer, but we are at the point where this is pretty absolute. Why? Because when there is a split, that makes it ripe for circuit court review, which we're already seeing with this New York issue" "What's really interesting is that although the prediction markets have lost in a number of states, and this is not just Kalshi, Coinbase and others are involved in this litigation. But very recently New Jersey, right across the way, actually disagreed with New York" "So this is a mixed record, and that's great. This creates a split, which means this is hyper ripe for a higher court to take this on"

Brianne Kimmel

@briannekimmel • 2d

Huge win for Deel to move beyond traditional one-time background checks towards continuous identity protection - Real-time deepfake detection - Continuous verification to prevent account takeovers & AI-driven fraud - Secure device provisioning linked to verified biometric data

Owen Gregorian

@OwenGregorian • 3d

Fender’s CEO seems to think your bandmates are just analog AI | Terrence O'Brien, The Verge Bud Cole also compared learning cover songs to AI training data in a controversial interview. Fender CEO Edward “Bud” Cole gave an interview to T3 in May celebrating the 75th anniversary of the Telecaster with comments on AI and music that initially flew under the radar. But it has started making the rounds recently, pouring more fuel on an already raging fire of bad PR following the company pissing off basically the entire guitar-playing community by sending cease-and-desist letters to builders, claiming copyright of the Stratocaster body shape. Fender’s CEO seems to think your bandmates are just analog AI Bud Cole also compared learning cover songs to AI training data in a controversial interview. Some influential guitar YouTubers have even said they’re done buying Fender gear in the wake of the controversy. And Cole’s resurfaced comments comparing cover songs and bandmates to a sort of “analog AI have only sunk the company’s standing further among the loudest of its fans online. T3 editor-in-chief Mat Gallagher’s feature mostly paraphrases Cole’s statements, and Fender did not immediately respond to a request for comment or clarification. But here are the relevant bits from the interview: Cole’s philosophy is that AI in music is nothing new. “I think AI has existed in music as long as there’s been recorded music,” he says. While the biggest barrier to playing guitar is the time it takes to learn the instrument, it’s the second barrier, writing songs, where he believes AI plays a part. “I actually believe cover music has been sort of analog AI for a long time,” says Cole. Those that don’t yet have the skills to write their own songs can play the songs written by their favourite artists instead. “I listened a lot to REM, U2, The Smiths and The Cure, and at some point I got sick of just listening to them. I wanted to play it, so I learned to play guitar.” Those taking their first steps into writing, according to Cole, can also lean on a second analogue form of AI: their band mates. You might just have a chorus or a riff to start with but then the drummer or bassist can add to it, and a song is born. AI can also play this role. “I actually think that we are in the brink of freeing up people to move beyond the same old covers and to really get into working like they do with their bands,” says Cole. Cole is trying to draw a comparison between a human “training” on a handful of cover songs and an AI ingesting enormous datasets of copyrighted music. He appears to be suggesting that, by learning to play other people’s songs, internalizing those influences, and then synthesizing them into something new, you are essentially doing the same thing as an AI. This is a woefully misguided take that says to me that Cole either doesn’t understand AI or doesn’t respect artists. For starters, scale matters. No person could possibly learn all of the songs used to train your average generative AI model, which, in the case of Suno, is suspected to be in the millions. Additionally, it dismisses the inherent humanity of the millions of tiny decisions, conscious or otherwise, that an artist makes during the songwriting process. Whether they’re driven by emotional response, reacting to a happy accident, or compensating for limitations, the artistic decisions made by a human are unique to them. This is fundamentally different from a model spitting out something based on a prompt and a network of data points. As Steve Onotera, better known as Samurai Guitarist, points out, a player’s physicality, or the tiny errors that every human is prone to, prevent them from replicating someone else’s work perfectly. That kind of serendipity can’t be replicated by an LLM. The same is true of bandmates. Humans who pull from their own unique sets of “training data,” life experiences, and physical skills or limitations are not the same as a chatbot. An AI doesn’t have taste or instincts in the way that your picky bassist who studied jazz composition in college does. If you told your drummer they were no different from an AI model, they’d rightfully be insulted. Later in the interview, Cole says: “I believe that AI is actually going to help create a whole new world of guitar players that use it. To help connect with other musicians, to be more productive. And across the chasm into becoming a student of songwriting to a master of songwriting.” Cole’s assertion that AI will somehow help people “across the chasm” to becoming master songwriters is also, frankly, ridiculous. Evidence is mounting that relying on AI tools is actually leading to deskilling. Using an AI to suggest rhymes or metaphors for pain isn’t the same as practicing songwriting and developing skills. The AI has never been left at the altar or sweated over the perfect pre-chorus transition. Repetition is the key. The adage is that you need to write 100 (or 1,000) (or 10,000) bad songs before you write one good one. That’s how you grow beyond tired tropes and learn to recognize when you’ve stumbled into something good.

Glenn Gabe

@glenngabe • 4d

An update on Reddit's lawsuit against Perplexity and Serpapi -> A US judge largely denies Perplexity and three data scraper firms' bid to dismiss Reddit's lawsuit over claims of copyright law violations under DMCA "In his opinion, US District Judge Paul A. Engelmayer said that at this early stage, Reddit has plausibly pleaded that there was a conspiracy, with SerpApi providing a product to circumvent Google access controls and Perplexity AI paying for it." Also, Google plans to amend its complaint that was dismissed: "Google told Ars that it planned to amend its complaint to keep its lawsuit alive, but SerpApi told Ars that Google and Reddit were both trying to “use the DMCA to wall off the open Internet by retroactively claiming control over content that they didn’t author and don’t own.”

Martin Austermuhle

@maustermuhle • 8d

Interesting: @DCPoliceDept is spending $378k to buy Skydio drones. MPD has been using drones for two years; the department says it limits them to monitoring big gatherings, searching for missing people, and barricade situations.

Antonio Regalado

@antonioregalado • 13d

Axios: De-extinction company Colossal Biosciences trying to raise funds at a $20 or $30 billion valuation. Company has no products (it does own a cloning co., which sells services) but does have cool stories.

Nic Cruz Patane

@niccruzpatane • 20h

Waymo Co-CEO on their multi-sensor autonomy approach at @ycombinator: "There’s been a longstanding debate about what kind of sensors you actually need for autonomous driving. Naturally, more sensors mean higher performance, but also higher cost and higher complexity. Humans, of course, can drive with just eyes, so there’s that proof of existence. If the goal were to just approximately match human performance or to build an assist product, that’s a very reasonable way to go. However, if you are targeting full autonomy and superhuman, strongly superhuman performance, you find that weak sensing leads to a safety curve that flattens out way too early. At Waymo, we’ve taken an approach where we use multiple sensing modalities. We use cameras, LiDARs, and radars, and they all complement each other. Cameras give you high resolution and color, but they’re passive and degrade in darkness and glare. LiDAR gives you a direct measurement of the 3D structure of the world around you. Radar is very good at punching through environmental conditions and weather like fog, rain, or snow, and it can directly measure velocity using Doppler. LiDAR and radar are active sensors, so they see just as well in pitch darkness or, for example, when driving into a blinding sunset."

Jeremy Fiance

@jfiance • 2d

.@anyscalecompute has agreed to be acquired by Nscale in a deal reported at $1.65 Billion. A huge outcome generated by @robertnishihara, @pcmoritz, @istoica05, @KeertiMelkote, and the entire Anyscale team - and another landmark company to come out of Berkeley. The story began inside Berkeley's RISELab with Ray, an open-source framework built to make distributed computing simple - letting developers scale Python across large numbers of machines and GPUs without rewriting their code. As Ray gained adoption at record pace, Robert, Philipp, and Ion founded Anyscale to build the commercial platform around it and help companies run at production scale. When the demands of modern AI arrived, that foundation turned out to be exactly what the moment required, and Anyscale became core infrastructure for training, fine-tuning, and inference across massive GPU fleets. Today, Ray is used by AI teams worldwide, and Anyscale has become an important part of the infrastructure enabling this modern AI revolution. With Nscale, Anyscale's platform will now be brought together with the compute, data centers, and power needed to run increasingly demanding AI workloads at scale. The House Fund was fortunate to back the team since the first round and watch a Berkeley research project become foundational technology for the AI industry. Huge congratulations to Robert, Philipp, Ion, Keerti, and everyone who helped build Anyscale. Here’s me at Anyscale’s Ray Summit NYC earlier this summer. Go Bears!

Shishir

@shishirmehrotra • 6d

We signed the Open Weights letter because we believe a strong open-weight ecosystem makes something essential possible: the freedom to match the right model to the right job. People judge these models on a single dimension: reasoning ability. But that's not how the best products actually choose one. Some of the biggest reasons we reach for open-weight models at Superhuman have nothing to do with reasoning. The first is latency. We deliver suggestions at the speed of typing, which is very different from what the deep reasoning models are built for. The second is scale. We run thousands of queries per user per day, so it has to be economically viable at that volume. American AI leadership won't come down to any single model. It'll come from how widely this technology gets built into real products that real people use every day.

Vaibhav Sisinty

@VaibhavSisinty • 19h

Full duplex AI is here. It watches, listens, and talks back, all at the same time. ByteDance just launched SeedRealtime. You hold up your phone with the camera on, talk to it, and it simultaneously sees what you see, hears what you say, and responds out loud. In one demo four people sit at a dinner table. The user introduces each person once. The model remembers every face. Then everyone starts talking over each other and it still tracks who said what. This is benchmarked against GPT Live. And it's already live in production for millions of users on the Doubao app. Most AI still works like texting. This one works like a person sitting across from you.

alexandra j. roberts

@lexlanham • 2d

update: impossible foods Inc. convinced a federal judge to toss $1.75m of the $3.25m jury verdict against them for trademark infringement. the judge found a lack of evidence showing oppression, fraud, or malice &amp; ordered a new trial to “prevent a miscarriage of justice”

Mith

@Mith_ • 5d

Just got done @insideclimate reading your article "China's EV Problem: What Happens When Millions of Batteries Die?" which was published on July 30th, and I felt that a heads-up on a few things would help you deliver a more in-depth and factual piece. I guess we should start with this bit of confusion. "Yan Wang is also a co-founder of Ascend Elements, one of the largest battery recycling companies in the United States." But further along you have this: "The company filed for Chapter 11 bankruptcy protection in April." Hard to be the largest and no longer a company at the same time. Now then, the article brings up Li-Cycle and Lithion as well, but as always, you framed it the way everyone seems to do. Since it is the surface narrative that is easy to digest, you attribute the downfall of each company to external factors. "Battery recycling has proven difficult to make profitable worldwide." Two of them saw their downfall stem from overextending limited resources on front-end production and failing to launch commercial versions of their battery-grade platforms. Another failed because of a contested component in their platform, which was not unknown, and was forced to shut down and rebuild due to a lawsuit, this just got the ball rolling on the downfall that was fed by even more poor choices. Battery-grade processing from secondary material in the US currently represents less than 1% of total global capacity. In fact, the only company that can produce a battery-grade chemical is R3 Lithium, which was created to take over the Ascend Elements assets, specifically the Georgia location that can produce battery-grade lithium carbonate directly from black mass. Unfortunately, Ascend was forced to file Chapter 11 just as they achieved a positive gross margin at the location. Now it also stated that you reached out to Glencore, but they did not respond. That sounds about right since the former CEO of Li-Cycle is now head of their battery recycling division. I doubt he will give public statements on the downfall of his former company. But let's move on. You talk about regulations in China and even how they changed on April 1, giving print inches to talk about the small unlicensed workshops where a good portion of the batteries are ending up. But you do not really touch on what has happened since those new rules went into place and really what was causing so much of that feedstock to go to those workshops. First, since the new rules went into place, informal workshops have been offloading their black mass and raw scrap holdings into the market to liquidate them. This has caused an already overcapacity market to get flooded with lithium carbonate, nickel sulfate, and cobalt sulfate, driving prices down. In a market where margins outside of the larger companies are razor-thin. Add the steps China is taking to deal with the involution in the cell industry and low prices are only compounding the problems the lithium-ion recycling sector is already dealing with, mainly the one below. While regulations are part of the problem, the main driver and the reason those workshops were able to divert so much feedstock is a spatial-temporal imbalance in that the official lithium-ion recycling facilities, mostly located where cells are produced, are not where recycling demand exists. Because shipping lithium-ion batteries long distances is cost-prohibitive, this geographic mismatch is the leading cause for the success of small workshops. The industry built an infrastructure to recycle factory scrap right next to plants, but remains woefully inadequate to collect end-of-life vehicle packs distributed across the country. There is a bit more nuance that is missing from your article that could give people a better understanding, but I am currently working on how a company is promoting a rather mundane hydrometallurgical process for black mass processing but promoting it as a breakthrough in EV recycling technology.

Shruti

@heyshrutimishra • 1d

Patrick Collison just dropped Stripe's internal data on startup creation. New businesses starting on Stripe are up roughly 2x year-over-year. That is the largest relative jump the company has ever recorded, bigger than the COVID spike of 2020. The obvious objection: more AI-generated slop businesses. But the data does not support that. The median business is doing better this year than last. The probability of reaching $1M, $5M, $10M in revenue is improving. Time to revenue is declining. YC is seeing the same thing. Companies in each batch grow faster than the last. The counterintuitive part: AI is not centralizing opportunity into fewer hands. It is distributing it. The tools that make it easy to start also make it possible to succeed without venture capital, without a large team, without being in San Francisco. By every objective metric Stripe can measure, there has never been a better time to start a business.

Podcast Alpha

@PodcastAlphaX • 1d

Sanjit Biswas says Samsara's moat isn't a model. It's 25 trillion data points a year nobody can crawl. Every AI company claims a data moat. Most are talking their book. Samsara runs CNNs at the edge, VLMs and JEPA models in the cloud, and shops OpenAI, Anthropic and Google interchangeably for the reasoning layer. If the model mattered most, it wouldn't be four vendors deep. The actual moat took a decade of hardware and change management, getting frontline workers to use it on roads, construction sites and grids that make up 40-50% of world GDP and were never digitized. Qasar Younis at Applied Intuition draws the same line on autonomy - 70% of Applied Intuition's revenue is already not cars, that's real P&L, not a forecast. Trust the number the data throws off, discount the model story wrapped around it. For $IOT, that decouples the stock from whichever foundation model wins next quarter. Whether the data compounds into new revenue fast enough is the open question. Full breakdown and the trade setup:

Business Breakdowns

@bizbreakdowns • 6d

Automotive is 3% of global GDP. It's the largest industrial category on earth, and it's one of seven verticals Applied Intuition sells into. Qasar Younis on the size of physical AI. "You have Waymo being valued at a hundred and twenty-six billion dollars by fairly sophisticated investors... It's because the impact can be very, very large, and that's one part of one of those markets." "If we execute, we're gonna do fantastic. And I think this company can be certainly 10x, if not much, much bigger than it is." "Before when we used to say things like this, like there will be a multi-hundred billion dollar physical AI company, people would say, 'Well, that doesn't really make sense.' Now you see with companies like SpaceX, how big they've gotten, and they're hard tech companies that are just really focused on one thing. And you just see, wow, these markets really are really big."

Brooke Hammerling

@brooke • 1d

Look @ouraring, I love you. I am a customer. Will continue to be. But to pull this BS of killing battery life of ring in order to get me to buy new one is ridiculous. It started immediately when you announced new product and I got an email push. My one year old ring (bought new) now only lasts 30 hours per charge. Cheeky. Annoying.

Rohan Paul

@rohanpaul_ai • 19h

So Google just announced some major changes in DeepMind’s leadership. Demis Hassabis will become Chair of Google DeepMind and Alphabet’s Chief Scientist, leaving daily management while continuing to advise its models and research. Basically it gives Hassabis more time for AGI strategy, scientific discovery, global policy work, and Isomorphic Labs, while Kavukcuoglu carries the pressure of shipping Gemini. The timing also reflects scale: Gemini now serves 950M monthly users. So model research, releases, and app execution needs a much larger operating job. The announcement also confirms that Gemini 4 is in development, though Google provided no release date or technical details. Jeff Dean and Sanjay Ghemawat are separately leaving to form a public-benefit research company, with Alphabet remaining an investor and Cloud partner. Google is building two leadership tracks, one deciding where advanced AI should go and another turning that research into products used at Google scale.

The Tectonic

@thetect0nic • 2d

The Wall Street Journal: @anduriltech Industries is in advanced talks with Maryland's historic Sparrows Point shipyard, a 3,300-acre former Bethlehem Steel site in Baltimore County, to build a production facility for autonomous drone boats, per people familiar with the matter. The investment could reach hundreds of millions of dollars, and Maryland Gov. Wes Moore's office has been involved in the discussions, with any deal likely including state incentives. No lease has been signed yet, and talks could still fall apart. This comes as Anduril, valued at $61 billion after a funding round this spring, pushes to roughly double capacity across its weapons production overall. The company already has unmanned surface vessel partnerships with UK-based @KrakenTechGroup and South Korea's HD @HyundaiHeavyInd , runs a revamped Seattle shipyard for low-rate production, and is building HD Hyundai boat prototypes in Korea. A Baltimore-area site specifically would put Anduril near Washington and the Coast Guard's largest shipyard, easing testing and fleet integration. Anduril faces real competition in this space, including fellow Baltimore-based @BlackSea_Tech , and recently lost a Navy competition for medium-size drone boats, a defeat that pushed several rival companies to complain to Congress and sue the government, though Anduril itself isn't among the litigants. The Journal frames this squarely against what's already been demonstrated in the Iran war: the U.S. military's own combat debut for drone boats came when autonomous vessels built by @Saronic , a different manufacturer, rescued two Apache helicopter crew members after their aircraft was shot down, and separately completed strikes on Iranian shipping and submarine infrastructure, an operation that lines up with CENTCOM's own confirmed first combat use of Corsair unmanned surface vessels against a submarine and ship facility at Bandar Abbas weeks earlier. Despite the momentum, the technology remains genuinely immature: the Journal and Navy officials report drone boats from multiple vendors have misidentified objects, crashed, drifted off course, or simply stalled during military testing, with the military still struggling to build software that reliably pilots unmanned vessels across open water.

Jonas

@jonaslamis • 19h

Uber has done deals with twenty-plus autonomous vehicle companies now. Rivian up to $1.25 billion. Another $300 million into Wayve if the London robotaxis land. Checks into Lucid and Nuro on top of that. What gets less attention is who actually touches the cars. Hertz and its Oro Mobility arm are standing up fleet management for the Lucid robotaxis. Avomo handles cleaning and charging for the Waymo cars running on Uber in Austin. I haven't seen anyone say who eats the cost when those cars sit idle.

Jeromy Lewis

@ExtensiveGrowth • 1d

$SPCX first public earnings: rev $7.81B (+92%), Starlink prints the cash, Space + AI burn; stock +9% into the print, days gains given back AH Wall Street finally got to grade Elon’s public rocket company — and the print is exactly the three-headed hydra the S-1 promised. $SPCX Q2 2026 — first earnings since the June IPO (CNBC / company / LSEG): • Revenue: $7.81B vs ~$6.93B est — +92% YoY off $4.1B • EPS: −$0.09 (street was ~−$0.26; not a clean apples-to-apples) • Call: 4:30pm ET Segment scoreboard (the whole business model in three lines): • Connectivity (Starlink): $4.29B vs ~$3.83B — op income $1.66B ← only profit engine • AI (xAI / Grok / X / Colossus): $2.56B vs ~$2.18B — op loss $1.26B • Space (Falcon / Dragon / Starship): $962M vs ~$835M — op loss $542M Starlink is the cash cow. Rockets still lose money. AI is a furnace with revenue on top. Tape: • Closed $125.33 (+9.4%) — best day since mid-June • After-hours ~$117 (~−7%) — classic “buy the rumor, sell the numbers” • Still under $135 IPO price; opened trading ~$150 on June 12 • Mkt cap ~$1.65T at the close | 52-week $107–$226 • Shorts were sitting on ~$8B+ paper gains into the print; lock-up expiry is the next landmine Context the street already knew: • Lost ~$4.9B last year, mostly AI infra • xAI merge in Feb (combined story once stamped ~$1.25T) — data centers in space is the long pitch • IPO cash pile swollen past $90B (pre-IPO cash was outgunned by debt) • Cloud/monetization side deals already on the board: Google AI capacity (~$920M/mo talk), Anthropic on Colossus, Reflection AI chips (~$150M/mo) • Terafab chip plant dreams with Tesla/Intel — up to ~$119B full build if it ever fully materializes • Starship is still the “key enabler” — and still the execution risk Why it matters: 1. This is a Starlink equity with a rocket option and an AI call option. Strip Connectivity and the other two segments are venture burn at mega-cap scale. 2. AI is real revenue now — not just a slide. $2.56B in a quarter means the Colossus/cloud story is booking; $1.26B op loss means the multiple is still a faith trade on margins and model quality. 3. Space segment still doesn’t pay the bills. NASA/commercial launches + Starship R&D = strategic moat, not near-term P&L. 4. First public print = valuation reset risk. $1.6T on ~$8B quarterly sales only works if Starlink keeps compounding and AI burn turns into free cash flow. 5. Musk conglomerate optics. Tesla just got sold on FCF/Robotaxi. SpaceX can’t afford a “trust me on Starship + Grok” shrug with lock-ups looming. 6. AH fade is the tell. Day-session short squeeze / FOMO into the print; real money wants Starlink KPIs, Starship timeline, capex path, and AI unit economics on the call. What to listen for at 4:30: • Starlink subscribers / ARPU / churn / D2C mobile roadmap • Starship cadence and pad/reuse reality after Flight 13 issues • AI capex vs cloud deal backlog (Google / Anthropic / Reflection durability) • Cash burn, debt trajectory, and lock-up supply • Any Tesla-SpaceX “might make Elon’s life easier” merger noise (Shotwell didn’t kill it at IPO) Net: First public report: beat the top line hard, Starlink carries the whole empire, Space and AI still bleed. Revenue almost doubled. The stock ripped into the number and sold the reality after the bell. At a trillion-plus, $SPCX isn’t priced as a rocket company — it’s priced as Starlink + a moonshot AI factory attached to the world’s most important launch system. Deliver the factory, or the multiple comes back to Earth.

݁݁݁OOOEƚolw69݁݁݁݁݁݁̈ ҉҉҉҉҉҉҉҉ ✰̈̈̈̈

@eawlot3000 • 1d

.@SpaceX quarterly revenue surges in debut results on strong growth in its Starlink business Reuters News 05 Aug 2026 04:04:19 Aug 4 (Reuters) - SpaceX SPCX.O reported on Tuesday a 92% rise in revenue for the April-June quarter, in its first earnings since going public, buoyed by strong growth in its Starlink satellite-internet and AI businesses. It reported revenue of $7.8 billion, compared with $4.1 billion a year earlier. The company's stock has declined 8% since its record-breaking initial public offering in June that valued the company at about $1.75 trillion. The stock could face additional pressure from the expiry of SpaceX's post-IPO lock-up period starting on Thursday, which may unleash a wave of insider and early-investor shares on the market. Starlink and SpaceX's broader connectivity operations remain the company's primary financial engine, underpinning CEO Elon Musk's push to build an AI-first business that extends beyond renting compute capacity to developing frontier models, consumer and enterprise software, and, eventually, data centers in space. The company's satellite-internet unit has continued to expand its global subscriber base, aided by launches of additional satellites and a growing range of consumer, enterprise, aviation, maritime and government services. But that expansion has come with tradeoffs: average revenue per user (ARPU) has dropped as SpaceX has entered more international markets and rolled out lower-priced plans. Investors are watching whether SpaceX can maintain growth while improving the economics of its network, particularly as it spends heavily to expand coverage, increase capacity and develop direct-to-device mobile services. SpaceX's AI business, which includes xAI, Grok, and social-media platform X, and a rapidly expanding data center operation, has been its biggest area of investment. The business is generating revenue from compute contracts with Anthropic, Alphabet's Google and Reflection AI, though a portion of its recurring revenue has yet to be recognized. Operating losses at the AI business have mounted, and SpaceX has cautioned that the AI unit will require sustained investment before it can generate profits consistently. Starship, SpaceX's next-generation reusable rocket system, is yet to enter commercial service but is expected to enable deployment of higher-bandwidth Starlink satellites and orbital AI-computing infrastructure. The company's ability to turn Starship into a reliably reusable vehicle is central to its longer-term strategy. Investors have closely watched for updates on testing progress, launch cadence, reusability milestones and the vehicle's satellite-deployment capabilities. Separately, SpaceX said that it had partnered with Nvidia NVDA.O to use its chips in the Starmind AI1 orbital compute satellites. The space segment, which includes commercial launches, government missions and development of Starship remains a significant source of costs and uncertainty. While launch activity for Falcon — SpaceX's partially reusable workhorse rocket — has remained robust, revenue can vary with the mix of internal Starlink deployments, commercial customer missions and government contracts. In recent years, SpaceX has increasingly prioritized launches for its own satellite network over third-party payloads, while continuing to absorb significant costs tied to Starship's development. Investors will also be keen to hear Musk's comments on a potential merger between SpaceX and Tesla TSLA.O after a Wall Street Journal report last week that executives at his electric-vehicle company had been told to prepare for a separation of its China business ahead of a potential deal. Musk dismissed the report as "fake news," but he had previously declined to rule out the possibility, citing growing overlap between the companies.