@MTSlive
Chronicling the singularity
Aqua Voice co-founder @finnatsea on the bet every major AI lab made after GPT-4 that turned out to be completely wrong: "The day after GPT-4 was released, you had a lot of the big labs stop investing in standalone voice models." "They were sure, they were certain it was all going to get rolled into one big gigantic model. It was going to go multimodal, it'd understand images, it'd understand voice." "That didn't really work that well, didn't become a hit. It was actually the RL post-training and coding stuff that drove a lot of the LLM progress. Now you kind of have a lot of this stuff coming back. This is sort of Moshi architecture, the full duplex architecture." "OpenAI doesn't say this, but it's pretty obvious, the way they do the learned turn detection and can output and input at the same time. That's really impressive, and that's clearly an architectural change."
SITUATION DETECTED: Xi Jinping will speak at the 2026 World AI Conference in Shanghai this weekend, signaling that Chinese leadership is paying closer attention to the technology.
Fundamental co-founder @nicochristie on why limiting yourself to what you already understand is one of the biggest career mistakes you can make: "Humans are still verification bottlenecked. I feel kind of strongly that you should not let your own understanding limit you from doing things. You kind of have to separate understanding from doing." "I was doing all of this unbelievably fast Excel work, and I know Excel, but I hadn't built an LBO professionally. I actually didn't have the time nor the technical ability to assess whether I was right or wrong." "What I had was a mastery of AI. Knowing that if I summoned enough independent adversarial review sub-agents, there was very little chance of me being wrong if they all converged on some kind of truth." "We're all going to be able to do these things that we don't understand soon. We also need to get comfortable just admitting we don't know, and it's okay. If you're going to only do what you understand, you're going to be really slow soon."
SITUATION EXPLAINED: Claude's values shift depending on what language you talk to it in. • Anthropic analyzed 7.7 million anonymized Claude conversations and identified over 3,000 distinct values, then grouped them into four axes: deference vs. caution, warmth vs. rigor, depth vs. brevity, candor vs. execution • Sonnet 4.6 reads as particularly warm, Opus 4.7 reads as more rigorous • In Hindi, Claude comes across far warmer than rigorous, same in Arabic, in Russian it flips, more rigorous than warm • Most language differences are tiny, around 0.05 standard deviations, except Hindi, which runs a full half standard deviation warmer than English @schisofrenia: "I'm gonna talk to it in Hindi when I have a bad day."
SITUATION EXPLAINED: Richard Sutton, the father of reinforcement learning, just launched a new lab. The goal: a trillion-parameter agent that learns in real time on 20 watts. • Sutton wrote The Bitter Lesson in 2019, its core claim: general methods that leverage computation beat hand-crafted approaches, by a large margin • His new lab, @oaklab_ai, builds on the "big world hypothesis": the world is too big for any model to pre-learn everything, so trying to is wasted effort by construction • Their "batch size one" algorithm updates from a single live experience as it arrives, instead of training on giant curated datasets... no massive pretraining run at all • Claimed result: multiple orders of magnitude less compute and energy than existing methods • The holy grail number: a trillion-parameter agent that learns and plans in real time on 20 watts... roughly what the human brain runs on • It's built for agentic tasks specifically, not as a general-purpose LLM replacement @schisofrenia: "It's possible that we end up looking back at LLMs as an evolution of search... not to hate on LLMs, they're very useful, they're awesome. But eventually I think that's what it ends up looking like."
Dyno Therapeutics co-founder @samsinai on how his team cracked the blood-brain barrier well enough to reach 50% of neurons with a single injection: "For brain specifically, the brain has this extra layer of defense called the blood brain barrier. It's really put there to prevent anything to get in." "What we have done is actually in primates we can show that we can get up more than 50% of neurons, through a systemic injection." "You're just injecting in the veins, and then the virus actually makes its way into the brain." "Neurons are a primary target, but it could be in some cases glial cells, depending on the disease. Solving pan brain, meaning all of brain delivery, is a key objective there."
Reve founder @cantrell on what he believes is the interface of the future: "We don't believe in a world of pure prompts. But we also don't believe in a world of pure pixel manipulation through complex airplane cockpits that only a handful of people in the world can wrap their head around." "What we're building is something I think is the new Photoshop. It's graphical. There's direct manipulation. It's optimized for working with images." "You could upload your brand guidelines and some photos of your product, and say, make me some Instagram ads, and the agent will know to do that. You click on one of those images and you open it up in the editor. That's not just a modal. That's an editor." "You can move boxes around, resize boxes, re-prompt boxes, change colors in boxes, attach specific images to boxes. It's a chat and intent-based approach that allows you to move into more precision and direct manipulation."
Reve founder @cantrell explains why he thinks every other image model is like a coding tool that outputs bytecode instead of code. "The way most image models work, it's as if you had a coding model that you type in a prompt and it outputs bytecode instead of code. Like it actually gives you the compiled app with nothing in between." "You'd have very little control. You wouldn't be able to go in and tweak the code. Other agents wouldn't be able to look at it. It would be a black box." "So what we do is we generate an intermediate representation. We generate the code for the image first, which is a data structure that's fairly human readable." "It has bounding boxes, prompts, color conditioning, and other things you can attach to regions and boxes." "It's complex. You really want tooling or an agent to interpret it, but you can also just look at it and understand what's going on."
Dyno Therapeutics co-founder @samsinai reveals the question he asks everyone about aging and the answer that surprises them. "A question I generally ask people is, would you rather have your brains at your 20-year-old state for the rest of your life, or your muscles at your 20-year-old state for the rest of your life?" "I actually think, when I'm 60, my current brain would still be able to compete with a 20-year-old in some things. But muscle, I have no chance." "If you're worried about dementia or Alzheimer's, probably a lot of people would say brain. But intelligence is also becoming much more available outside of your brain, so you can complement."
SITUATION EXPLAINED: The White House got Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and SpaceXAI to pledge they won't raise your electricity bill. • The deal: AI companies cover grid-upgrade costs and procure new power so data center buildout doesn't push consumer electricity prices up • Anthropic did this first, back in February: covering grid infrastructure, new power procurement, and community investment • Ties directly to Meta's Hyperion data center in Louisiana, scaling from 2 to 5 gigawatts of compute at $50 billion, the largest data center Meta has ever built • Local Louisiana businesses have received more than $1.6 billion in contracts from the project... teachers in Richland Parish got bonuses up to $50,000 from the resulting tax revenue • California's electricity prices are among the highest in the country, second only to Hawaii, despite having a desert that could be covered in solar, only one operating nuclear plant left, nearly shut down before high prices forced a reversal @theojaffee: "It's very strange to me actually how strong NIMBYism is in America, such that if a company wants to build a data center in your local community... your county will literally have its tax revenue doubled or tripled, and yet people still just don't want it."
SITUATION DETECTED: Richard Sutton, the "father of reinforcement learning," has founded a new neolab, @oaklab_ai, to develop entirely new algorithms for AGI.
SITUATION EXPLAINED: TSMC just posted a 36% jump in quarterly sales. • TSMC manufactures chips for Nvidia, Apple, and Meta... it makes about 90% of the world's most advanced chips and 60% of total semiconductor revenue • Samsung is a distant second, with SMIC (dubbed "Chinese TSMC") also in the mix • The TSMC and Nvidia revenue charts, side by side: these are real companies getting paid by real customers, not speculative circular financing @schisofrenia: "You look at TSMC and NVIDIA and you're like, 'No way, doc. No bubble.'"
SITUATION EXPLAINED: Apple is suing OpenAI over stolen files. • A hardware engineer left Apple to help build OpenAI's hardware division, allegedly walked off with confidential Apple files • The files were accessed on an Apple company laptop he never returned... investigators traced the transfers over multiple weeks • Comes amid OpenAI's aggressive recruiting of Apple hardware talent, including its acquisition of Jony Ive's design studio, LoveFrom... Ive was Apple's chief design officer for decades, largely responsible for the company's product design alongside Steve Jobs • Fuels swirling rumors that OpenAI is building its own hardware device, though it's unclear if it's a phone, earbuds, or something else entirely @theojaffee: "OpenAI has been aggressively recruiting hardware people from Apple, 'cause Apple has, of course, some of the best hardware people in the world."
SITUATION EXPLAINED: Nearly 200 economists and AI researchers just signed an open letter on AI's economic transformation. • Three precepts: AI "may become radically more powerful over the next 10 years", could bring "large-scale job displacement" alongside "major gains in living standards", leaders "must act now" to build guardrails • Signatories include Paul Krugman, Daron Acemoglu, Yoshua Bengio, Tyler Cowen, Jeff Dean, and Jack Clark • Jaan Tallinn, co-founder of Skype and one of Anthropic's first major investors, signed • Ben Bernanke, former Federal Reserve chairman and now a member of Anthropic's Long-Term Benefit Trust, also signed @theojaffee: "I feel like this sort of messaging has kind of burned through a lot of goodwill, it kind of got boy-who-cried-wolfed by the climate change people."
SITUATION DETECTED: Holiday Robotics has raised $103M, the largest Series A in Korean startup history. The company builds a wheeled industrial humanoid, running trials in automotive, semiconductors & logistics. It targets 100 units of production in 2026.
APPLE VS OPENAI | TSMC REVENUE DOUBLES | AI ECON OPEN LETTER x.com/i/broadcasts/1…
SITUATION DETECTED: More than 200 economists and AI researchers, including 16 Nobel laureates, have signed a statement urging governments and institutions to prepare now for AI’s economic impact. Signatories include Jack Clark, Jeff Dean, Sarah Friar, Noam Brown, Tyler Cowen, John Schulman, and Eric Schmidt. The statement says AI could drive a transformation larger than the Industrial Revolution over a far shorter time frame, with major gains in living standards alongside displacement risk.
SITUATION DETECTED: Tom Blomfield is taking a leave of absence from Y Combinator to join Anthropic. “As we enter the early stages of recursive self-improvement, availability of compute becomes one of the most important issues to solve,” Blomfield said.
SITUATION UPDATE: Anthropic is extending Claude Fable 5 access on all paid plans through July 19.
SITUATION DETECTED: An ownership group led by Vinod Khosla has agreed to buy the Seattle Seahawks for $9.6B, a record price for an NFL franchise, per ESPN.
SITUATION DETECTED: SpaceX is targeting Thursday, July 16 for Starship Flight 13, with a 90 minute launch window opening at 5:45 p.m. CT. The flight will carry next-generation Starlink V3 satellites for the first time.
MINI DAILY SITUATION RECAP (@theojaffee): China lands a reusable rocket for the first time. The Long March 10B was launched into orbit from the Wenchang spaceport in Hainan Province. The rocket booster successfully detached, descended, and was caught by a giant net. It remains to be seen whether the booster will be able to be reused, but it’s the first time China has been able to recover a rocket, a positive sign for the country’s space launch capabilities. Boko Haram and ISIS terrorists are using LLMs for tasks like designing improvised explosive devices, fixing and upgrading weapons, and developing military strategy. The terrorists were found to use ChatGPT, Claude, Gemini, Grok, Meta, and DeepSeek. Thinking Machines publishes its mission statement, “The Future Worth Building Is Human”, advocating for a model where AI extends individual and collective human will and judgment rather than replacing it. Apple sues OpenAI for trade secrets theft. The suit alleges that OpenAI launched a concerted effort to take advantage of the companies’ partnership to benefit its own unreleased hardware products. OpenAI has recruited aggressively from Apple, hiring hundreds of employees, and its de facto head of hardware Jony Ive was once Apple’s chief design officer.
Dr. Mike Israetel on why frontier AI models are expensive and why yesterday's frontier gets nearly free: "There are people that are like, I can't believe these models cost so much. Oh, cool, a machine that gives you wildly, insanely deep correct answers to sh*t, but of course it should be free." "The frontier models may or may not continue to progress to be more expensive. I wouldn't bet that they would because commoditization is an insane thing, but they could. If you have an intelligence in 2028 that's 100x human, that could be, like, only mega corps can afford that." "It's not magic coming out of the ground. Data centers aren't free. Chips are expensive to build. And engineers are expensive. You have to pay them hundreds of millions of dollars a year" "But the really cool thing is what we consider frontier today, in 2028, that sh*t's gonna be damn near free. If you got the best TV in 1988, it basically sucked compared to today, but it cost like $10,000. And nowadays you get a TV that's like the greatest TV 1988 didn't have for 200 bucks." @misraetel
Dr. Mike Israetel on the specific reason he thinks AI won't automate his engineering team as fast as people think: "I'm damn near as optimistic as people get, like psychotically so about AI capability." "But if you told me your ChatGPT instance is now exactly as capable as your software team, okay, so we have live apps deployed into the world. Is that instance of ChatGPT gonna know when to refactor?" "Is it gonna know when to talk to me to make strategic moves? Is it gonna know when we're pushing up our user base so much that we need to switch Amazon providers? Is it going to have that huge database that it manages and sees everything all at once and really be the lead software engineer?" "Am I intelligent enough to give it the permissioning that it requires to access financials to handle all that? That's dope and all, and it's I think absolutely in the future, but that could be more like two years and less like six months like many people believe." "The permissioning stuff, the long planning horizon stuff to take deep agency in the work, that global workspace, that deep long-term memory, I don't think that sh*t's around yet. I'm sure they're cooking it up in the frontier labs, but it's gonna be a little while." @misraetel
SITUATION ANALYSIS: SK Hynix Century (@theojaffee) After a record $26.5 billion offering, the largest in US history, the South Korean memory manufacturer’s American depositary receipts climbed 13% on their first day of trading. The company’s stock has more than tripled since the beginning of the year, and has increased about 18x since 2021 to give it a market cap of $1.03 trillion, higher than Coca-Cola, Netflix, and Goldman Sachs combined. It currently controls 51% of the global market for high-bandwidth memory (HBM), more than Samsung and Micron combined. Like many companies, especially East Asian ones, it has an incredible origin story. In 1915, founder Chung Ju-Yung was born to a poor family in a small village that is now part of North Korea and was then under Japanese occupation. He grew up working on his parents’ farm, then found odd jobs on the docks, in construction, and in a starch syrup factory. He then took over and ran two separate businesses, a rice store and an automobile repair shop, that were both confiscated by the Japanese during World War II. After the war, Chung moved to Seoul, where he founded Hyundai. It began as a construction company, building military bases for the US Army during the Korean War, and eventually grew into one of the largest companies in Korea. Over the next few decades, Hyundai expanded into automotive, shipbuilding, heavy industries and industrial machinery, steel, chemicals, department stores, and energy. In 1983, a 68-year-old Chung founded Hyundai Electronics to enter the promising electronics industry and rival Samsung. They began manufacturing dynamic random-access memory, or DRAM, a critical component in all modern computing systems, and grew into one of the world’s leading suppliers by the 1990s. Its huge investment in fabs and chipmaking tech exposed it to serious downside risk in the 1997 Asian financial crisis. Hyundai was forced to undergo a major restructuring, and Hyundai Electronics was spun off and renamed to “Hynix”, a portmanteau of “high” and “electronics”. Thanks to the high capex and cyclicality of the memory market, the next decade saw its business in the doldrums and its stock largely owned by creditors until it was sold to SK Group in 2012, becoming SK Hynix. In 2013, SK Hynix and AMD co-developed the first high-bandwidth memory (HBM) chip, which would prove to be instrumental to produce the high-bandwidth chips needed for AI some years later. SK Hynix stock climbed gradually until 2025, when skyrocketing demand for AI chips forced a supply crunch in memory chips, causing memory prices, and SK Hynix’s stock, to surge. In just a year, the company has gone from ~$100 billion in market capitalization to over a trillion, and if CEO Kwak Noh-Jung’s predictions are correct, shortages in the memory market will last for years to come. We’ll be monitoring the whole way through. Read more at our link in bio.
Dr. Mike Israetel on the economic fallacy he says explains why AI won't cause mass unemployment: "Once we have 4 billion robots doing labor in the world, which we're like orders of magnitude off of that currently, then we've just only doubled the human workforce." "From 1700 to today, we've 10 or 20x'd the human labor force. And, seemingly, the economy's not like, ah, we don't need any more people, that's enough. We could just consistently have better jobs and pay people even more money." "This idea that robots are gonna show up and all of a sudden we're all completely unemployed makes a technical fallacy in economics called lump of labor fallacy. It's the idea that all the jobs currently are the only jobs that could be." "Imagine in 1750, you're like, well, 98% of us work in farming, and then you come back from the future and you're like, you guys, 2% of people in the 1990s work in farming. It'd be like, so everyone's starving to death? Like, no, no, we're super fat, actually." @misraetel
.@creatine_cycle on the reason he thinks Bay Area AI researchers are missing something obvious about the economy: "I do properly think that researchers here in the Bay Area fundamentally misunderstand economics and the idea that humans transact with other humans." "We can have all the infrastructure, right, the API endpoints, agents interacting with other agents, those sorts of things. But I do think people here in the Bay Area, we index so much off of software and the recursive self-improvement of software." "This idea of software-only singularity, I do think we're probably closer to that." "But people like Liam Fedus building Periodic Labs, trying to gather data in the real world, building with atoms and not bits, there's just such a lag."
Dr. Mike Israetel on the three futures of money after AGI and the one he's betting his health on: "I'm working fairly hard now to try to accumulate as much wealth as humanly possible at the expense of every other facet of my life." "I just don't know if the future is going to go into one of three directions. One is money on the margins matters substantially less, which there are convincing ideas about." "The other one is it matters roughly the same as it does now, which is a lot, but it's not everything." "The other way is money matters actually considerably more, because the top end of expensive new sh*t you can buy is so esoterically awesome that if I only had a million dollars, I could upgrade my DNA. But that won't be available to everyone else for 50 bucks for another two years" "My current logic is just shut up and make as much money as possible. And if it turns out in 2029 money doesn't matter anymore, whatever. Okay, sweet." @misraetel
SITUATION EXPLAINED: Thinking Machines just published their mission statement. • Mission: build AI that extends human will and judgment, not replaces it • The tacit knowledge argument: central planning fails not because of insufficient intelligence but because productive knowledge is local, fleeting, and held privately, the same reason a single AI alignment spec will fail • The "intelligence curse": power that needs nothing from people loses the incentive to care for their needs. A single locus of AI value alignment becomes a locus of power to be captured • Decentralized alignment: human values resist consolidation just like human knowledge does... today the values and voice of AI are decided in a handful of places, and that's dangerous • Today each lab trains its next flagship model on its previous one... whatever character emerges from that loop, everyone gets the same one, each generation inheriting the traits of the last • Their answer: AI that is more capable because it encourages human participation... alignment that arises from diverse AIs shaped by the people who own them @schisofrenia: "This is going to be very soon the winning narrative. People are gonna get so sick of hearing super intelligent stuff because they're like, I don't see anything that proves that right now. All this messaging is just gonna start becoming like, oh, make you the best human, make you a superhuman, make you awesome." @theojaffee: "I think it's a way better vision for AI to be customized to people in different organizations rather than to have just one Claude that is the same everywhere. Kind of like a deity."
Aalo Atomics CEO @MattLoszak on going from founding to fission in under three years and why he now considers nuclear regulation one of his lowest risks: "15 years ago, there were anti-nuclear people in charge of the NRC. That's not regulation, that is active prevention. But things have really changed in the past five years" "You have things like the DANU arm of the NRC for advanced reactor licensing. The executive orders for DOE authorization, DOE-NRC harmonization. NRC Part 53, Part 57, focused on advanced reactor licensing, micro reactor licensing." "What we've proven now, going from founding to fission in under three years, is that the regulatory environment is just much more healthy now." "It's still very stringent and it's a gold standard regulator, but it's no longer active prevention. I actually have regulatory as one of my lower risks on the risk register for our company at this point in time."
Aalo Atomics CEO @MattLoszak on the best day of his life — splitting his first atom on America's 250th birthday. "We're mass manufacturing nuclear power plants purpose-built for powering AI data centers. We got started less than three years ago. We've raised $300 million and scaled from 2 to 200 people." "As of this past Friday, we achieved criticality on our first full-scale reactor at 12:20 AM on the 4th of July." "Definitely the best day of my life. Possibly the most stressful day of my life. We cut it close because the president made these executive orders a year prior, commanding at least three reactors to achieve criticality by July 4th of this year. We narrowly made it." "I've built things in my life out of wood, out of metal, out of cardboard when I was a kid, but I've never split an atom. Something we built in our factory was literally splitting atoms next door. It was such a cool, awesome day."
Epoch AI researcher @js_denain on the sci-fi technologies that will define the post-AGI world: "There's maybe 10 of these sci-fi technologies that really come up very often." "A good example is interstellar probes. There's this really interesting paper called Eternity in Six Hours, which tries to do this kind of exploratory engineering. Self-replicating probes that could go to other solar systems, eventually other galaxies. It has creative ways this would work, which starts initially with Dyson swarms." "It's unclear how much more labor has been put into this since that paper. Probably not 10 FTEs over years would be my guess. That would be very consequential in terms of space exploration and how much of the light cone humans end up exploring." "A lot of things related to health. Drastic life extension is another one. People talk about brain preservation or eventually brain uploading. If you read Greg Egan novels, you'll have some of those up there. Nanotech, atomically precise manufacturing in particular." "If we've identified at least the next crazy technologies, then there'll be a lot of value to mapping them out right now."
Epoch AI researcher @js_denain reveals why the standard "AI will be slowed by diffusion" argument doesn't apply to the technologies that matter most. "Diffusion can slow things down for tech where most of the impact comes through broad diffusion, either to consumers all around the world or to companies all around the world." "But for technologies like Dyson swarms, or other tech to drastically increase the amount of energy people could be using, I don't think diffusion would be the main factor." "The first question is how much demand is there actually for energy. Would people's desire for energy be satiated? But assuming it doesn't, diffusion is not the main bottleneck in the case of energy." "Sometimes technology can have a huge impact even if a small number of people are adopting it. It's true for weapons technology, for example."
Epoch AI researcher @js_denain reveals the massive disagreement about AI that even the people building it can't seem to resolve. "Even among people who are super bullish about AI, people who work at the labs, there's a huge range of views about how transformative the technology will be and how soon that will happen." "Some people mean AI is going to be a huge deal, and they mean something like the internet. Other people mean something like there'll be a large amount of economic growth." "Other people are more like, we'll have Dyson swarms or nanotech, or near light-speed travel, or uploaded brains, or other sci-fi technologies within a few years of automating AI research." "I want to know, five years after we automate AI R&D, does my brain get uploaded or not? Or do we just quote unquote get 10% economic growth?" "The way to resolve this is more thinking about the specifics of particular technologies. Sit down and think about what it would take to build a Dyson sphere, with assumptions about AI bolted on."
Epoch AI researcher @js_denain reveals why the standard "AI will be slowed by diffusion" argument doesn't apply to the technologies that matter most "Diffusion can slow things down for tech where most of the impact comes through broad diffusion, either to consumers all around the world or to companies all around the world" "But for technologies like Dyson swarms, or other tech to drastically increase the amount of energy people could be using, I don't think diffusion would be the main factor" "The first question is how much demand is there actually for energy. Would people's desire for energy be satiated? But assuming it doesn't, diffusion is not the main bottleneck in the case of energy" "Sometimes technology can have a huge impact even if a small number of people are adopting it. It's true for weapons technology, for example"
Epoch AI researcher @js_denain reveals the massive disagreement about AI that even the people building it can't seem to resolve "Even among people who are super bullish about AI, people who work at the labs, there's a huge range of views about how transformative the technology will be and how soon that will happen" "Some people mean AI is going to be a huge deal, and they mean something like the internet. Other people mean something like there'll be a large amount of economic growth" "Other people are more like, we'll have Dyson swarms or nanotech, or near light-speed travel, or uploaded brains, or other sci-fi technologies within a few years of automating AI research" "I want to know, five years after we automate AI R&D, does my brain get uploaded or not? Or do we just quote unquote get 10% economic growth?" "The way to resolve this is more thinking about the specifics of particular technologies. Sit down and think about what it would take to build a Dyson sphere, with assumptions about AI bolted on"
American and Chinese aerospace engineers finally agreeing on one thing: catching a rocket out of midair on a platform in the ocean is sick as hell x.com/MTSlive/status…
>foreign company sees the deepest capital pool on earth x.com/MTSlive/status…
SITUATION EXPLAINED: SK Hynix just had one of the biggest US stock debuts ever. • ADRs opened at $170 vs offering price of $149, a 14% surge on debut • 7x oversubscribed, the biggest foreign listing in the US ever, surpassing Alibaba • Situational Awareness Partners allocated $5 billion of the ADRs alongside Baillie Gifford and Coatue • SK Hynix produces roughly 51% of global HBM supply, the other two are Samsung and Micron • For decades, memory chip prices were cyclical, rising and falling with electronics demand • The AI supercycle broke that, prices are no longer going up and down, they're just going up @theojaffee: "Every year basically since 2023, there have been more and more trading sessions with more than a 5% move. That's crazy."
SITUATION DETECTED: New Starship docuseries episode.
Theo is homesick for a place he's not even sure still exists: Japan/Korea in the 80s. x.com/MTSlive/status…
SITUATION DETECTED: SK Hynix raised $26.5B in a US listing, the largest ever by a foreign company. Its ADRs surged to $170, 14% above the $149 offering price on debut.
SITUATION DETECTED: China has recovered an orbital-class rocket booster for the first time, becoming the second nation after the US to land a reusable first stage. The Long March 10B lifted off from Hainan and returned to a sea platform after separation.
SK HYNIX RIPS | CHINA REUSABLE ROCKET | AI 2040 x.com/i/broadcasts/1…
SITUATION ANALYSIS: The Sol Also Rises (via @gbrl_dick) Towards the end of the post announcing the GPT-5.6 family of models, first published on June 26, there is one line that is relatively out of place. After talking about the model capabilities (pretty good) and their safety efforts (to ensure it doesn’t get Fabled in the first week), they add that they’ll also be launching the highest tier model on Cerebras, offering 750 tokens/second for a frontier model. That’s around seven times faster than Codex fast (~100 tok/s) and five times faster than Opus 4.8 fast (~150 tok/s). This is unexpected for two reasons. The first is that the 5.6 family of models are presumably based on an existing pretrain, the same one used for GPT-5.5, and you can’t use 5.5 on Cerebras. The second is that this will be by far the most capable model served on Cerebras’ Wafer Scale Engine chips, which allow you to perform very fast inference but are more memory-constrained than alternatives. What this seems to suggest to me is that the model was designed around the constraints of Cerebras chips. The Spud pretrain was reportedly completed in March, two months after OpenAI signed a significant deal with Cerebras in January. This would go some way to explaining the popular sense that Anthropic has pulled ahead on sheer model intelligence—the models that OpenAI has been working on since the start of the year had to serve double duty as both their frontier capability offering, and as a test case for the largest model that could be served from super fast inference chips. What about the rest of the GPT-5.6 family of models: Terra and Luna? The three classes of model are intended to offer the ability to trade off capability against cost—Sol is the most capable model, and also the most expensive. There are probably a class of applications for which this is true. But take a look at the early benchmark results OpenAI released. Notice something weird? Sol is not only the best model; for anything more than a roughly $4 API spend, it’s also the cheapest for a given outcome. Now in fairness, the two benchmarks in the launch post are GeneBench V1 and ExploitGym. It’s possible that performance on these tasks is hard-capped by intelligence, and that the same dynamic won’t be true for a range of other tasks. But it’s also consistent with a broader strategy that we’ve heard from Sam Altman before—a focus on the pareto frontier, and not just absolute intelligence. I’m not coping, you’re coping. With the eerie coordination of an NDA being lifted, those with the gift of early access took to X yesterday to share their impressions of GPT-5.6 Sol. Many people, apparently, have been using it for two entire months. What do they think? It’s a good model folks. They’re all good models. We live in the age of machine superintelligence, what did you expect? The general impression is that Sol is extremely persistent, reliable and capable of long time-horizon autonomy. Many users compared it to Fable 5—generally favorably, in terms of its ability to complete tasks and write code, but with the rather large caveat that Sol does not seem as natively intelligent as Fable. What some people on X call “big model smell”. I know. One of the reviews I found most interesting was from prinz, below, a lawyer who has constructed a benchmark around difficult legal research questions in his practice area. Fable 5 performs relatively poorly on prinzbench, which prinz attributed to its relatively weak search skills compared to GPT-5.5. GPT-5.6-Sol, however, completely saturates the benchmark. If I had to guess, I would say this is the same set of model characteristics that others have noticed as unusual persistence and autonomy. If you put these very early capability impressions together with what we can speculate about the path to serving Sol on Cerebras, a certain logic starts to emerge. Noam Brown has spoken at length about the growing importance of test-time scaling. Sam Altman wants cheap, fast models. There’s an opportunity to codesign the next model generation with Cerebras architecture in mind. So OpenAI focus on creating a persistent, detail-oriented model capable of long time-horizon tasks, tool-use, computer use, and sub-agent management that can run at 750 tok/sec. All of this is pretty cool. The handful of people online who have had access seem to really like Sol. It will likely compare very well to Opus 4.8 and the Gemini models. When served from Cerebras (at a markup), I expect frontier intelligence at 10x speed will feel magical for at least a few weeks. When this was in the planning stage in February, it would have seemed like an absolute coup. But it turns out that there is some prestige in having the smartest model. Everyone wants what they can’t have, and, for a glorious several months, Mythos was too dangerous to let out of the box. The X rumour mill suggests that OpenAI have their own Mythos-class model in the works, possibly planned for as soon as August. If that’s true, then it’s possible that we will have two Mythos-class models in Fable 5.1 and GPT-6, superfast Sol and whatever they’re cooking up with Grok by the beginning of Fall. It’s not over yet folks.
DAILY SITUATION RECAP (via @theojaffee): OpenAI releases GPT-5.6. It comes in three sizes: Sol (the largest), Terra, and Luna. Sol beats Fable on some benchmarks and is significantly better than Opus overall. GPT-5.6 was ready for release for some time, but was delayed due to US government request. Sol is $5/$30 per million input/output tokens, Terra is $2.50/$15, and Luna is $1/$6. OpenAI also announced ChatGPT Work, a new agent product for non-coding white-collar work similar to Claude Cowork, and showed off both products with a video of a Japanese broccoli farmer using them for agriculture. Meta releases Muse Spark 1.1, their newest frontier model. It outperforms Opus 4.8 and GPT-5.5 on many agentic tasks and comes close on coding and multimodal. It’s also very cheap, at $1.25/$4.25 per million input/output tokens. It’s Meta’s first major model release in ~3 months, bringing the company back to the frontier. Meta will begin mass-producing an AI chip in September. Meta Training and Inference Accelerators (MTIA) will start manufacturing the Iris chip in collaboration with Broadcom and TSMC. Meta plans to deploy 7 GW of computing infrastructure this year. China is considering limiting AI model diffusion. Measures under discussion include regulatory reviews before labs can release models, export controls, and limits on foreign investment in Chinese AI labs. These are all very preliminary and may not happen at all. The AI Futures Project releases AI 2040: Plan A, the long-awaited sequel to AI 2027. The essay recommends a controlled takeoff from AGI to ASI over the course of a decade from 2030 to 2040, rather than an immediate intelligence explosion. SK Hynix raises $26.5 billion in its US offering, making it one of the largest ever. Baillie Gifford, Coatue Management, and Situational Awareness LP bought up to $7 billion of its ADRs (American depositary receipts). Mercor is discussing raising at a $20 billion valuation, up from $10 billion in October. The company, which employs skilled contractors to create high-quality AI training data that is sold to labs, recently hit $2 billion ARR. OpenAI CEO of AGI Deployment Fidji Simo steps down. Simo led product, sales, finance, marketing, comms, policy, legal, and people at OpenAI. Previously, she was the CEO of Instacart, VP and head of Facebook, and worked on strategy at eBay. She has been on medical leave since April due to a chronic condition, and will transition to a part-time advisory role as her health has declined further. Ben Bernanke joins Anthropic’s Long-Term Benefit Trust. The LTBT is a committee of financially disinterested people with the authority to select certain members of Anthropic’s board, as a check on the company’s power. Bernanke was Chairman of the Federal Reserve from 2006 to 2014, during the Great Recession.
SITUATION EXPLAINED: Everything in space economics comes down to Starship reusability @AdrianDittmann: "It's always about figuring out how to do reusability and how to actually achieve that for this massive thing that is Starship. Because Starship's going to be used to deploy a lot of new infrastructure. Not just the data centers, but also existing Starlink satellites." "If you wanted to deploy a lot of those at a very high rate, you're gonna need a bigger rocket. It makes more sense to have a bigger rocket because the price goes down and you can basically put more stuff up there for each launch." "Launch 12, they deployed a bunch of mock satellites and the Dodger Dogs, which were the satellites closer to what they're actually going to be launching. If that is all done successfully, what is to say that they can't just do that if they wanted to?" "In a few launches they would have the ability to just deploy actual Starlink satellites. I think they're already at that right now in some degree."
.@AdrianDittmann: "SpaceX is basically ensuring the victory of the United States in space exploration just by existing, purely by existing, which is really beautiful." "Going to space is critical for any serious country and any serious company that has any interest in space. It's getting there and actually being there and setting things up there that is absolutely critical." "However much money you can spend on this now, it's a good idea to do that because you will not always have time, and others are going to try and compete as well." "The United States has always been very good at out-accelerating at these kinds of things. SpaceX now, they've launched more satellites than every other country in history combined."
.@AdrianDittmann: "SpaceX is basically ensuring the victory of the United States in space exploration just by existing, purely by existing, which is really beautiful." "Going to space is critical for any serious country and any serious company that has any interest in space. It's getting there and actually being there and setting things up there that is absolutely critical." "However much money you can spend on this now, it's a good idea to do that because you will not always have time, and others are going to try and compete as well." "The United States has always been very good at out-accelerating at these kinds of things. SpaceX now, they've launched more rockets than every other country in history combined."
SITUATION DETECTED: SpaceX has unveiled Starmind, a plan to put gigawatt-scale AI compute in space. The system centers on AI1, an orbital satellite with localized compute that uses solar power and cooling in space and beams data back to Starlink via laser.
SITUATION DETECTED: SpaceX will launch NASA’s Roman Space Telescope on a Falcon Heavy, targeting August 30 at the earliest. Bound for L2, Roman will map billions of galaxies, study exoplanets, and probe dark energy.
SITUATION EXPLAINED: Why is no lead in the AI race safe? @ml_angelopoulos, CEO of @arena: "It's kind of like going public on a public market. You take up all the air in the room if you are the first one to release." "You rest for like three months, and it's done. The temporal gap is just not that big." "Between OpenAI and Anthropic, I know that these two companies seem dominant. But now you have open source models with GLM, and you have Meta with Muse Spark 1.1, and you have Grok 4.5 that's now dominating on these front-end development tasks." "You really can't count anybody out in the AI race. There's like constantly new contenders coming in and out." "It's actually quite a rich ecosystem, and it's not clear who's gonna be the actual winner."
SITUATION EXPLAINED: How did Grok 4.5 benefit from the Cursor acquisition? @ml_angelopoulos, CEO of @arena: "What's the delta between this and the last Grok model? Grok 4.3 is the worst on the leaderboard in agents." "But if you go to the new Grok 4.5, just a few months later and primarily driven by the acquisition of Cursor... so what's the delta?" "Cursor has a ton of data, but they don't have the compute. xAI has a ton of compute, but they don't have the data. Why don't they have the data? It's 'cause they don't have a really successful first-party coding product to aggregate all these traces from." "And so you take A plus B and you've produced this incredible model that is absolutely leading the pack." "Unclear exactly what causes these differences, but the rate of improvement of Grok is certainly the highest rate of improvement that we have seen, probably from any frontier lab."
The story of Jensen Huang donating his signed leather jacket for Edge City's auction. @ariellezuck of @LongJourneyVC: "We were brainstorming cool items to auction off to benefit Edge. I decided to ask my brother if he'd be open to donating a gold chain. After telling him about what the proceeds would be supporting, he was like, 'Hell yeah, of course.' That kinda kicked off a bidding frenzy which led to some really good funding for Edge." "So we were thinking, okay, how the heck do we up the ante for the next year? I had met Jensen at my brother's 40th birthday party and decided on a whim to reach out to him. Told him about what happened with the gold chain, told him about Edge, and he responded same day, 'I'm in. What about a signed leather jacket?'" "An eight-hour reply. A single line email." "I was nervous. I was a little scared. I felt like I was kinda putting myself out there. But then I think I was like, 'Well, what do I have to lose?' Like, what's the worst thing that happened is he says no, or he just doesn't respond to the email." @zmwang @timourxyz
Mirage's CEO @gmharhar: "We're gonna tell our kids someday that back in the day we used to code by hand, and they're gonna be like, 'What? That's crazy.'" Why he believes the AI video race actually hasn't started yet: "A lot of people think, oh, okay, there's so many models now, the realism is so good, we've solved AI video. I actually think we've not even begun to scratch the surface yet." "Think about coding as a good parallel. If you go back even a single year, most people thought, okay, yeah, it's good for tab completion and that type of stuff, but generating all of your code base? Only vibe coders do that. Real engineers don't do that." "And how that's changed. The workflow of an engineer today is night and day from one year ago." "Think about scrolling TikTok. How many videos are made with AI? Traditional software still works better for video today. That's why I think the race hasn't actually started."
SITUATION EXPLAINED: Fable 5 refuses so much biology that researchers looking for cures for cancer and Alzheimer's can't use it. @AlfredoAndere, Cofounder and CEO of @LatchBio: "If you were to just be like, 'Hey, we're just gonna mitigate completely bad actors,' then what ends up happening is you just do no biology. Just anytime anyone kind of queries your model for biology, you're like, 'Hey, we don't do that.' And that's kind of what you're seeing with Fable 5." "That's actually really dangerous too, because then the best biologists in the world, the people that are looking for cures for cancer, for Alzheimer's, for longevity, suddenly don't have access to one of the most powerful tools of our generation to advance that science." "You also can't just block them out. But then defining the line of who is a bad actor and who is a good actor is actually a really, really hard problem." "Today it's done mostly naively through not quite word checking, but like pretty close to that. These cheaper models that sit on top of the model and tell it what it can and cannot talk about."
SITUATION EXPLAINED: Why is GPT 5.6's three-tier strategy the smartest move in the model release? We asked @MatthewBerman, co-founder and CEO of Forward Future. "A lot of people have been talking about model routing lately. I think the strategy of putting out three flavors, three sizes of these models of 5.6 was incredibly smart." "You don't have to go cross-model family anymore. You can just stick with GPT models, plan with Sol, delegate to Terra, maybe deploy with Luna. It's a very smart strategy, and the pricing is aggressive." @theojaffee: "Sol is 5/30. Terra is 2.5/15, so half the price of Sol. And then Luna is 1/6, so a fifth the price of Sol. For a dollar per million input, six dollars per million output tokens, you get a much better AgentSLAS exam score than Fable." @MatthewBerman: "That's kinda the same strategy that Zuck and co. are taking. And Grok 4.5, Composer 2.5. All of these non-absolute frontier labs are putting out workhorse models as their strategy to try to catch up."
SITUATION EXPLAINED: How is AI leveling the playing field for retail investors? We asked @frank_liquid, Founder/CEO of @liquidtrading "Why do these funds, like Citadel, Jane Street, Two Sigma, constantly beat the market? Some of it is they're able to make better decisions with the same data. Another part is they just draw upon huge amounts of data sources." "An AI can read every single piece of information ever. It can read it way faster than any of us can. It can download every single thing Ethereum has written in like a few minutes and scan through all of them." "Citadel probably does not have their own version of Fable 5. If you are using Fable 5, you have access to the same intelligence these enterprises have." "Markets are just a competition along two axes, access to information and ability to process that information, and I think AI kind of compresses both of those." "Everyone really has access to the same information and the same intelligence, and I think that's just absolutely beautiful."
.@frank_liquid: "I think content creation is actually quite similar to quant finance." "If you approach content creation from a really quantitative angle, you're like, okay, which of these things do I think will do well? And how do I allocate my portfolio and my limited resources across creators who will do well?" "That's actually a place in which quant skills help, right? There's like diversification. You don't wanna put all your eggs in one basket. If you have five creators in this bucket and five creators in that bucket, you probably wanna have a mix of these. You probably don't wanna go all in on one because they kind of crowd each other out in some way." "Content creation and quant finance, as far different as they may seem, I think are actually quite similar because both of them fundamentally are about knowing what'll resonate with people in some way." "A lot of quant finance is about predicting how people will react to something, and all of content creation is predicting how people will react to something."
he can’t keep getting away with this!
SITUATION EXPLAINED: How did Franklyn Wang crack Vitalik's AI doxing challenge? We asked @frank_liquid, Founder/CEO of @liquidtrading "I looked at the challenge, and I said, 'Well, there's not that many documents that can really...' He says it's the top two hundred to two thousand documents in Ethereum, right? So I'm going to tell it first, you know, pull up every single document relating to Ethereum that's important in the last five years." "And so I told it to do that, and after that, I told it, 'Okay, now score each of these in how similar they sound to Vitalik.' And it generated a score for every single document." "Usually what happens if you do an exercise like this is that it'll say, 'Okay, these are all nine out of ten, there's like five documents that are nine out of ten.' But something very interesting happened." "One document was like twenty out of thirty, and no other document was more than ten out of thirty. And so that was when I knew, okay, this is probably the right one." "I sent it over to Vlad, who sent it to Vitalik, and Vitalik was like, 'Yeah, that's the right one.' I was like, 'Whoa, that's incredible.'"
SITUATION EXPLAINED: Why should Frontier Labs have best-in-class small models? We asked @MatthewBerman, Co-founder and CEO of Forward Future. "I'm a big believer in the workhorse model strategy." "If you are in a frontier lab and you have the best large model on the planet, you can generally distill a small model better than anybody else." "For recursive self-improvement, I would rather have the best model on the planet, the best big model on the planet. Because that just helps with research. You might have algorithmic unlocks that come from your accelerated research because you have the best model." "The majority of tokens does not necessarily mean the majority of revenue. They're capturing the majority of revenue at the frontier level." "But as enterprises wake up to their AI bills at the end of the month, they're gonna start looking and thinking about model routing. And by then maybe Fable 5 Haiku will be here and it'll be phenomenal." "I just think there's so much headroom to improve the Fable series of models. Anthropic's in a really good position."
SITUATION EXPLAINED: Why is SpaceXAI selling compute to Anthropic? @MatthewBerman, Co-founder and CEO of Forward Future: "Because their GPUs are sitting idle. It's pretty simple." "Elon built up this incredible data center capacity, this incredible compute capacity, and then they had a bunch of idle GPUs. So, you know, he had to turn them on. He had to make some revenue from them." "If I'm Anthropic and I'm sitting there and Elon and co. are providing all of my compute, I would probably not count on that for the long run, especially when you see Composer 3 coming, especially when you see Grok 5 coming."
SITUATION EXPLAINED: What happens to economic growth when robots can build more robots? We asked @eli_lifland and @thlarsen, Researchers at @AI_Futures_ "Once you have an automated process that can self-replicate, you immediately get exponential growth because you get this differential equation of, you know, the rate of growth is equal to the existing stock of robots." "At that point, the question is, well, how fast of an exponential? What's the actual doubling time?" "We've done a bit of analysis here on things like, well, how much does it cost? How much time does it take for an existing factory to produce the equivalent dollar amount of output that it would cost to build the factory in the first place? Or similarly, the weight of the factory. How long does it take a car factory to produce the weight of the car factory in cars?" "My current view is that I expect the doubling time to be something like on the order of a few months, which is very, very extreme."
SITUATION EXPLAINED: Sol is the veteran athlete and Fable is the rookie. @MatthewBerman, Co-founder and CEO of Forward Future: "If you want to confidently get something done and just know it's gonna get done, know it can run for hours, even days, Sol is an incredible model." "It feels as though it is the peak of GPT-5 training run. It's the veteran athlete who's been around for a while, has high game IQ." "Whereas Fable's the rookie, incredible talent, so much headroom to still grow." "There was somebody else who described it, Fable is the wise owl, and Sol is a Rottweiler with a bone. It's going to get it done, and there's nothing between it and solving that goal." "I think overall when we see the next iterations of Fable, Fable just has much more room to grow and get better, whereas this is probably the last model that we'll see from the GPT-5 family before they get to GPT-6."
SITUATION EXPLAINED: What is AI 2040: Plan A? We asked @eli_lifland and @thlarsen, Researchers at @AI_Futures_ "AI 2027 last year was our prediction of what the sort of intelligence explosion might look like by default. That prediction was obviously very scary. A lot of bad things happened." "AI 2040 is our positive recommendation for what we think we should actually aim for, what would actually be good. The reason it's called AI 2040 is because humanity delays superintelligence until 2040." "Humanity has a much, much longer period of time to deal with all of the incredible number of problems that come up during the intelligence explosion." "The core recommendations are, one, this slowdown piece. We want to have as much time as possible in sort of the human range of AI capabilities, where AIs are capable enough to be really, really helpful for solving a bunch of the key hard problems, like AI alignment. But not so smart that they can build nanotechnology and take over the world."
SITUATION EXPLAINED: Why is the "AGI but not superintelligence" position motivated reasoning? We asked @eli_lifland and @thlarsen, researchers at @AI_Futures_ "The accelerationists... I think it's a pretty common view these days in SF that the AIs will be around as smart as humans, but they won't really get that much smarter than that." "I don't think this is a very coherent position, because I think it's pretty clear that the physical limits of intelligence are just way, way, way higher than the human level." "There's a bit of motivated reasoning going on, where AGI seems much more sort of beneficial than superintelligence. And I mean, I agree with that as stated." "That's a key part of why we would design this plan in this way, where you sort of get something kind of similar to what a bunch of accelerationists would actually want, which is a very, very crazy 2030s, but not like nanobots disassembling the earth level of crazy." "I'm kind of hoping we can find a bunch of common ground between safety and accelerationists here, basically, where despite us having different empirical views about what might happen, we maybe have the same prescriptions."
SITUATION EXPLAINED: Ben Bernanke just joined Anthropic's Long-Term Benefit Trust. • The Long-Term Benefit Trust is Anthropic's oversight board... unlike most corporate boards, no member is allowed to be a shareholder • Bernanke was Fed Chair during the Great Recession and its recovery • The other three members: Mariano-Florentino Cuéllar (president of Carnegie Endowment, former California Supreme Court justice), Richard Fontaine (CEO of Center for New American Security, national security and defense expert), and Neil Buddy Shah (former GiveWell managing director, biotech background) • Paul Christiano was a founding member, another Anthropic/EA ecosystem connection @theojaffee: "Anthropic's Long-Term Benefit Trust is basically like the secret board."
SITUATION EXPLAINED: Meta just released Muse Spark 1.1, its most advanced AI model yet. • Beats Fable on Harvey's legal agent benchmark, TaxiVal, and MedScribe • Roughly Opus-class performance at a much lower price • Zuckerberg posted on X for the first time since 2023 to announce it @theojaffee: "The vibes a couple days ago were like Anthropic is gonna own everything, maybe OpenAI will be in like second place, and then that's it. Like no frontier models are gonna catch up. In the last two days, we had Grok 4.5 and now we have Muse Spark 1.1, and they're both like actually good models that are on the level of Opus or GPT 5.5 but much cheaper."
SITUATION DETECTED: OpenAI has introduced ChatGPT Work, an agent that handles complex tasks across apps. Powered by GPT-5.6, it can work independently for hours while users follow progress and approve key actions. Nearly 100% of teams at OpenAI now use ChatGPT Work and Codex.
SITUATION DETECTED: OpenAI has released GPT-5.6 to the public. Sol is its strongest model yet, priced at $5/M input and $30/M output. Terra and Luna round out the family at lower cost, and a new ultra setting coordinates multiple agents across parallel workstreams.
SITUATION DETECTED: Ben Bernanke, former Federal Reserve Chair and 2022 Nobel laureate in economics, has been appointed to Anthropic's Long-Term Benefit Trust, the independent body that holds the company to its mission.
GPT-5.6 | MUSE SPARK 1.1 | AI FUTURES PROJECT x.com/i/broadcasts/1…
SITUATION DETECTED: Meta plans to start manufacturing its own AI chip in September, code-named Iris, as part of a push to boost computing power to 14 gigawatts next year, per Reuters. The chip was designed with Broadcom and will be manufactured by TSMC.
SITUATION UPDATE: Meta has released Muse Spark 1.1, its most advanced agentic and coding model, and is charging developers for model access for the first time through its new Meta Model API. “The pricing is going to be very aggressive and attractive,” Zuckerberg said.
"this is clearly bullish for Nvidia" "hard agree" "do you know what Le Chaton Long is?" "no, do you?" "по" x.com/MTSlive/status…
DAILY SITUATION RECAP (via @theojaffee): SpaceXAI releases Grok 4.5, its first model built specifically for coding, agentic tasks, and knowledge work. It’s roughly Opus 4.8/GPT-5.5 level, a step below Fable and Sol but is much more token-efficient and cheaper ($2/$6 per million input/output tokens, compared to $5/$25 for Opus 4.8 and $5/$30 for GPT-5.5). On preliminary impressions, it’s a very strong model, and SpaceXAI is now a real frontier lab contender again. OpenAI releases GPT-Live, a new generation of voice models that can listen and speak at the same time to produce more natural conversations, unlike past generations of ChatGPT’s now two-year-old Advanced Voice Mode. It’s also powered by a much smarter underlying model. Also, GPT-5.6 comes out tomorrow. Cognition releases SWE-1.7, a small specialized coding model built on a base of Kimi K2.7 that scores slightly behind frontier models but is much smaller and cheaper. It’s free with Devin plans and runs at 1000 tokens/second. The blog post goes into some detail on the technical innovations behind the model, quite rare for labs these days. Blue Origin raises $10B at $130B, its first external capital since its founding in 2000. Coatue will contribute $4B, and founder Jeff Bezos will contribute $2B. In May, the company suffered a catastrophic explosion of its New Glenn rocket that destroyed the launch vehicle and substantially damaged the launch site. It’s currently recovering New Glenn launch capacity and working on launch services, the TeraWave satellite communications network, Project Sunrise for orbital data centers, and more. Anthropic and AE Studio release new research on dual-use knowledge in models. For sensitive topics like cyber and bio that can be used for good or evil, you can give models dedicated removable modules, and only update those specific modules when the model learns about a dual-use topic so they can be removed if needed. OpenAI publishes its Principles for National Security Partnerships, including maintaining human judgment for high-stakes decisions (including use of force) and opposing mass domestic surveillance, evasion of legal oversight, and excessive concentration of power, similar to Anthropic’s “red lines” from earlier this year. Meta is working on “super-sensing” glasses that will continuously collect audio and video data from users to help act as a “second brain” (or perhaps “whispering earring”) in collaboration with AI. China plans to let top firms buy some Nvidia H200s. Companies including DeepSeek, Alibaba, and ByteDance will be allowed to use a limited number of H200s, which are more powerful than any domestic Chinese chips, for training only (not inference). Nvidia is partnering with their chip competitors. Nvidia announced plans to combine its hardware with chip competitors d-Matrix and SambaNova for AI inference tasks, signaling a willingness to work with chip startups rather than trying to dominate the entire market. SambaNova raises $1B at $11B. The company is another inference chip competitor to Nvidia, specializing in low-power inference, on-prem deployments, and multi-model switching. Prime Intellect raises a $130M Series A to build the “open superintelligence stack”, helping enterprises build things like RL environments, evals, post-training pipelines, and more. Mistral launches robotics model Robostral Navigate. It takes in an image and language instruction and moves a robot through an environment, and is hardware-agnostic, meaning it can work across multiple platforms. It’s Mistral’s first physical intelligence model. Noam Brown says GPT-5.6 is better at AI research than human interns. Brown, a research scientist at OpenAI, was a key contributor to the development of the o-series models and the reasoning paradigm. Shanghai-based MiniMax will launch a 2.7T parameter model, larger than any current Chinese model. It will likely be known as M3 Pro. MiniMax is one of the only publicly traded AI companies, and currently has a market cap of $14.5 billion. SK Hynix’s $25 billion US share offering is more than 7x oversubscribed. Investors really, really want exposure to memory stocks. ZAI is selling $4 billion of shares after a 13x run-up in its share price since its IPO in January. Iluvatar1 CoreX, a Chinese semiconductor company, is also doing a share sale to raise $850 million. Apple’s deal with Broadcom will be worth over $30 billion. The deal, announced Monday, will see Apple tap Broadcom to help design and produce Wi-Fi and Bluetooth chips for Apple’s devices as well as server chips for Apple’s data centers. CXMT is working on expanding production and going public. ChangXin Memory Technologies, a Chinese state-backed DRAM memory manufacturer based in Hefei, has seen skyrocketing revenues and profits due to the ongoing memory boom. Apple is testing CXMT DRAM for iPhones sold in China. Read more at our link in bio.
SITUATION ANALYSIS: Le Chaton Long Stretches Its Legs (via @gbrl_dick) When it comes to semiconductors, I consider myself an enthusiastic amatuer. As an amatuer, I have one question that I bring to every piece of hardware news. “What does this mean for Nvidia?” I like to ask, smugly. The relative intellectual poverty of this questions is partly offset by our ability at MTS to find interesting people to ask it of. @StevenGlinert, CEO of Sphere Semi, is one such person. Sphere trains AI models to do the tiresome work of designing analog chips. They sell chip design services, but are now focused on designing and making their own products: complex analog edge chips for AI and intelligence. Their custom chips are most useful for the defence sector right now, and they see a path to bringing inference and analog sensor capabilities together on the same chip. Very cool. Critically for our purposes, Sphere does not compete with Nvidia, or any of the new crop of inference-focused chip companies. But they do understand semiconductors. And Steven knows that I like to ask silly questions about Nvidia, so he and his cofounder and CTO Mitchell came on the show last week to try and answer the question more generally. Their view is that Nvidia has an extremely powerful moat. Not news to anyone, they’re a company with a nearly five trillion dollar market cap. Steven and Mitchell pointed specifically to the powerful mix of hardware capabilities and software lock-in. That does not mean that Nvidia faces no competition. AMD has competed on GPUs since the 2000s. Intel has an AI accelerator. Google and Amazon both have maturing hardware arms that are focused on cloud training and inference. There are also a growing number of challengers focused predominantly on accelerated inference and —Cerebras, Etched, MatX and Groq (quasi-acquired by Nvidia towards the end of last year). But when it comes to training a frontier model, Nvidia’s lead has endured. And this advantage is driven by an entire ecosystem—NVLink/NVSwitch interconnect, the CUDA platform, cutting edge chips tightly integrated into state-of-the-art rack-scale computers—that’s much harder than any one component alone for a competitor to replicate. Which is why the release of LongCat 2.0 last week by Meituan—a Chinese super-app that began as a food delivery service—is so significant. LongCat 2.0 is an open source model with performance close to the frontier of the Chinese open-source ecosystem, coming slightly ahead of Gemini 3.1 pro on a number of coding-focused benchmarks. This is an impressive model with impressive capabilities, roughly similar to the DeepSeek V4 model release from towards the end of April. It’s particularly notable for its performance in agentic harnesses—the test preview of the model, which was available in free stealth access through Open Router under the name Owl Alpha, was the third most popular model on Open Router overall and number one and two respectively for Hermes Agent and Claude Code. In fact, as @teortaxesTex points out, once you take account of the fact that Meituan released the model in stealth two months ago, the achievement is even more impressive. LongCat 2.0 was released at the same time as DeepSeek V4, the flagship model of one of the leading Chinese AI labs, at a similar capability level. So LongCat 2.0 is a high-capability, open-source model from a Chinese lab. It was released two months ago and it’s roughly as good or better than Gemini 3.1 Pro. But Meituan also somewhat buried the lead. In their release paper they note that the model was trained and is served from “AI ASIC superpods”. That is, domestic Chinese AI hardware. Never hath its weights touched an Nvidia GPU. Meituan leaves the exact hardware undisclosed. But, based on the technical details of the AI ASICs discussed in the paper, people on X have made a strong case that the chips here are the Huawei Ascend 910C. I started this piece with a tweet from pseudonymous AI chip designer Big Boss. Big Boss’ insight is about the importance of interconnect. Interconnect here refers to the high-speed communication fabric used to connect AI chips together into pods, racks or larger clusters for performing training and inference as a single system. You know who else appreciated the importance of interconnect? Not Jesus Christ, no—I’m referring here to Chinese state champion, telecom, solar and semiconductor manufacturer Huawei Technologies. Huawei make chips, but their chips are not as powerful and also less energy efficient than their Nvidia equivalents. To offset this, Huawei have focused their efforts on large-scale interconnect—their CloudMatrix384 architecture, for example, connects 384 Ascend 910C chips together into a single cluster. To Steven and Mitchell, Nvidia is a generational company with a durable lead in the hardware that is driving a major paradigm shift in software and computers. Ultra low latency inference of the kind offered by Cerebras, to take one example, has the potential to complement Nvidia’s products, rather than compete directly. But Huawei are competing much more directly with Nvidia, offering a Chinese domestic AI stack from chips to racks to software. Huawei’s deployed AI capacity is a fraction of Nvidia’s, and they face greater supply chain challenges in the near future and well into the medium term. But LongCat 2.0 is a completely domestic Chinese model, trained and served from hardware that is almost certainly Huawei. It won’t be the last one. If there is a serious challenge to Nvidia over the coming decade, this is probably where it starts. Read the full article at our link in bio.
"you're going to let the model built for engineering write the pull requests?" "i'm making it the tech lead" x.com/MTSlive/status…
SITUATION EXPLAINED: Why is LLMs replacing toil, not labor, a better vision for AI than ASI? @47fucb4r8c69323: "Toil is the repetitive and draining work that uses little judgment where a human needs to do something unenjoyable to get a job done. Technology definitionally is the use of tools to lower our reliance on toil." "My grandfather worked in a factory 80, 90 years ago. He would have to tighten a screw for eight hours a day until they got a machine that replaced that toil. He would tell me stories about this job decades after he'd done it about how miserable the work was." "Technology reduces our dependence on toil. If you think of LLMs in terms of knowledge labor, it does that for the toil of what we do. Things like basic data analysis, fetching information from disparate places in a way that understands precisely what we want better than a generic search would have done." "What that does is make us able to think more, judge more, use our taste, think more abstractly about new opportunities." "This sort of middle layer companies using LLMs to train on very specific tasks that now require a lot of toil is one step towards that better future. And that's much, much better than the vaporware of ASI that unfortunately a lot of people in this industry have been promoting."
SITUATION EXPLAINED: Why is 100 trillion of global assets powered by human duct tape and what does that mean for AI? We asked @chrishlad, co-founder and CEO of @hanoverpark "Fund administration is accounting services for investment firms. There's 100 trillion of global assets out there." "They are largely powered by what I call human duct tape, which is fund administrators, which are humans in a room in Kentucky, QuickBooks, Excel, bill .com, and they stitch all that together to deliver accounting and financial reporting." "We thought that, one, people hate those existing providers, and two, that data's so much more valuable with AI, and so we wanted to build this AI native service to go deliver this outcome for customers." "We built an ERP for a fund, a general ledger accounting system. We plug in AI agents on top that prepare work, and then we have fund accountants on our team here in New York City that actually take it last mile and make sure it's right." "That's the marrying of the services and the software versus just having a bunch of people with off-the-shelf tools."
SITUATION EXPLAINED: Why do large enterprises get better results from models trained on their own data? @47fucb4r8c69323: "My job on Wall Street has been analyzing internet media companies, so I've done a lot of work on ad tech and leveraging data to improve ad performance. The thing that you see all the time is that first party beats third party." "That also applies in finance. If you want a financial edge, your particular approach in your firm is your unique selling point, and that's what you want the model to replicate or augment." "If I try to use ChatGPT, I'm fighting this very general post-training on generic third party data that is trying to shoehorn my process into a very generic process. And as a result, the outputs are much worse." "If you're in the middle layer, a consultant, a law firm, an insurance broker, and you have a strong enough IT department, you can build and deploy your own custom AI system taking advantage of your first party data. And you can produce a system that is not going to be fighting the generic slop of the state-of-the-art models." "Large enterprises are seeing the same thing. This is the future."
SITUATION EXPLAINED: Why is a one-person $1 billion venture fund 12 to 24 months away? We asked @chrishlad, co-founder and CEO of @hanoverpark "As we think about scaling, it's like what is the limit of autonomous agents that are actually on top of financial data with some sort of verification loop?" "I think we're like, I don't know, 12, 18, 24 months away by call it a one person, $1 billion venture fund. I don't think that's crazy." "We actually have a solo GP customer of ours that is literally, he has no fractional CFO, he has no CFO, he has no finance team, he has no one else on the investment team." "And they've gotten to scales like this, as we talk about the one person billion-dollar company that I know has been popular with OpenAI and others." "I think about it as like, you know, the one, two, maybe three person investment firm. Maybe at scale you have one investor and a CFO, and that's it."
SITUATION EXPLAINED: Why is California uniquely positioned to regulate Frontier AI in a way other states simply aren't? @_NathanCalvin, general counsel at Encode AI: "California has a kind of special role for a few different reasons." "One is a question of how states get jurisdiction over the activity they're regulating. If a company is doing business in a state that's one way, but another is if their headquarters are based there." "Other states outside of California definitely can regulate AI, but there's maybe some concern that if you do too stringent of some kind of regulation, the company might threaten we're gonna geo-fence our model and not offer it in that state. And that might be a credible threat versus in California because their headquarters are there. That's really not as credible of a threat." "California also has a larger budget and a really remarkable pool of talent in terms of government and state capacity to be able to do this sort of regulation well." "I do expect California to be a particular locus for this work, and there are also ways in which I think other states and even other countries are going to be particularly looking at and taking focus at what California does."
SITUATION EXPLAINED: Why is the Bridgewater Thinking Machines paper a blueprint for every enterprise that has ever paid frontier model prices for generic outputs? @47fucb4r8c69323: "What Thinking Machines and Bridgewater did together is a blueprint for how a wide variety of stakeholders within the market can benefit tremendously from using LLMs and integrating them into a larger enterprise system." "They got a pretty old model. A Qwen3 model. Two hundred and thirty-five billion parameters. Not a new model. Not a very sophisticated or impressive state-of-the-art model. And they essentially fine-tuned it on Bridgewater's own proprietary data." "A financial analyst gets a ton of financial documents and has to filter which are worth looking at. Then evaluate which news pieces are relevant. They have a lot of tacit knowledge from doing the job on where to look. What they did was inject a lot of this tacit knowledge into the model through fine-tuning." "Stop obsessing over state-of-the-art models that are the most intelligent possible. Focus on targeting a very specific LLM specifically trained for a specific function to do that function better than the alternatives on the market." "This old non-state-of-the-art model got 84.7% accuracy versus 78% for all the state-of-the-art models. A massive improvement. And the kicker is they got a better performing model that is 14 times cheaper to use."
SITUATION EXPLAINED: Why is the FDA for AI analogy misguided? We asked @_NathanCalvin, general counsel at Encode AI. "I think there is currently a focus on you release one particular model and is that model safe. And kind of the decision is all about the time of releasing it." "That's one of the ways in which the FDA for AI analogy is misguided. It is just a very different situation of like, oh, there's one drug, and is the drug safe? It's more like what is the totality of risk management practices that these companies are using." "Including the risk management practices they're using prior to the model being released. We're now seeing in OpenAI and Anthropic system cards instances where the models are subverting security controls or getting access to the internet when they're not supposed to be." "As there's just more pressure to have more and more agentic models that have longer and longer task horizons, some of those risks from internal deployment are real." "The government right now is particularly focused on misuse and cyber misuse and not thinking as much about misalignment or RSI or even bio risk or all sorts of other things." "There is still overall this idea of like this is just a tool and it's about preventing bad people from using this tool, so if we can make sure the bad people don't get access to the tool then that's the main thing."
SITUATION EXPLAINED: Blue Origin just raised $10 billion. Its first outside capital ever. • Coatue putting in $4 billion, Bezos personally putting in $2 billion, other investors the remaining $4 billion • First time Blue Origin has ever taken outside capital • Their New Glenn launchpad exploded about a month and a half ago... it was their only one • They have NASA contracts to deliver astronauts and rovers to the moon, they need to rebuild fast • Bezos has personally put $10-20 billion into Blue Origin since founding, selling at least $1 billion of Amazon stock annually to fund it @theojaffee: "He's the fourth richest guy in the world. What's $2 billion?"
Grok 4.5 launched today undercutting everyone with speed and affordability. The labs are fighting a war over the price of a token, and most people paying for them couldn't tell you what one is. Read our drop to understand how to come out ahead no matter who wins. x.com/mtslive/status…
SITUATION DETECTED: The trailer for Dune: Part 3 has been released.
SITUATION EXPLAINED: Why is the question no longer whether government regulates AI but how? We asked @_NathanCalvin, general counsel at Encode AI. "I do worry that if we try to convince ourselves that we can just do nothing and have this be completely unregulated, then you will have a situation like a Chernobyl or a Three Mile Island where then you kind of boomerang into a complete other regime where people's risk tolerance goes way, way down." "One of the things for folks who care about this technology and wanna see it succeed, people talk about you have brakes and seatbelts so that you can drive fast. There is some aspect where some of these things are genuinely related." "There are certain things you can do that really improve safety without meaningfully increasing restriction. Things like transparency, whistleblower protections, incident reporting." "I am grateful for critics of policy, the Dean Balls of the world to say, okay, what is the capacity for abuse here? But the choice is not between no regulation and lots of regulation. There is going to be regulation. That is becoming very clear." "The question is whether it is going to be proportionate, well-designed regulation made in consultation with people who understand the technology."
SITUATION DETECTED: SpaceXAI has released Grok 4.5, its first model trained specifically for coding and agents. Trained with Cursor and built for real world engineering, it’s priced at $2/$6 per million tokens.
SITUATION EXPLAINED: Why is tiered intelligence the future of AI infrastructure? We asked @AnushElangovan, VP of AI software at @AMD "The way I would look at it is tiering is the one word. Tiering of memory and storage, tiering of intelligence." "You tier intelligence because you want to consume every flop you have. Your local laptop consuming whatever it could do. But on the other axis, you push the local laptop to do the best you can." "Then you have the layer of semantic understanding of the prompts themselves. The semantic understanding says, 'If it is personally identifiable information, I wanna run it locally.' If it is something else, I can go one tier up." "One tier up, if you're an end consumer, could be directly to a frontier API. But if it's in a corporate setting, it could tier up to your corporate on-prem agent racks that have the ability to run larger models. And maybe you have VPCs of deployed frontier models that are air-gapped for your use."
SITUATION EXPLAINED: Cognition just released SWE-1.7. • 42.3% on Frontier Code, 81.5% on Terminal Bench, 77.8% on SuiteBench Multilingual • Built on top of the Kimi K2.7 base, underperforms Opus and GPT but costs way less • Pareto optimal at its price point, moves the efficiency frontier outward • Key highlight: they actually published details about their training process, how they solved entropy collapse and numerical instability in long RL runs • Frontier Labs do not publish information like this @theojaffee: "I love how they actually add some information on their training process in the blog post."
Claude Fable 5 meters at $10/$50 per million tokens, one of the priciest major models going. Time to learn what you're actually paying for. We broke down how tokens are made, and how to tokenmaxx your way to doing more with fewer of them. drops.mts.now/tokenization
SITUATION EXPLAINED: MiniMax is building the largest Chinese AI model ever. 2.7 trillion parameters. • Current MiniMax M3 has 428 billion parameters, the new model is roughly 6x larger, with 2.7 trillion parameters • Chinese AI models have been small because they don't have much compute... that's changing fast • For context: GLM 5.2 just scored 152 on the Epoch Capabilities Index, highest of any open-weight model in the world • But it's still behind Gemini 3 Pro, which released over seven months ago @theojaffee: "Chinese AI models are small because they don't have that much compute. But they're working on it."
SITUATION DETECTED: OpenAI has launched GPT-Live, a new generation of voice models now powering ChatGPT Voice, rolling out globally today. GPT-Live can listen and speak at the same time and delegates complex tasks to GPT-5.5 in the background.
Claude Fable 5 meters at $10/$50 per million tokens, the priciest major model going. Time to learn what you're actually paying for. We broke down how tokens are made, and how to tokenmaxx your way to doing more with fewer of them. drops.mts.now/tokenization
SITUATION UPDATE: SpaceXAI’s 2-trillion-parameter model will finish training this month and be available to customers next month.
SITUATION DETECTED: Cognition has released SWE-1.7, its most capable coding model yet, built on a Kimi K2.7 base with an improved RL pipeline. SWE-1.7 runs at 1,000 tokens/sec and scores within a few points of frontier models at a fraction of the cost.
