Pinned Tweet
ASTRA GETS DISTILLED | ECONOMICS OF INTELLIGENCE | MISTRAL LARGE 4 nitter.cf/i/broadcasts/1RKZzBava…
Mistral CRO Marjorie Janiewicz explains why you don't need hundreds of billions of dollars to compete with OpenAI and Anthropic:
"It's possible because of the open-weight approach and the design choices we make."
"It does prove the point that you don't need billions of dollars to build a frontier model that’s as strong as closed-source models."
"Customers are starting to worry about their bills, the token price. When Mistral builds those models at a fraction of the cost, customers see that value back in cost efficiency."
"We released a time series model for financial services. It's very small, very agile, and it outperforms all models in time series context when it’s adapted to our customer’s data."
@MistralAI
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
Mistral CRO Marjorie Janiewicz predicts AI chatbots are reaching their ROI ceiling, and says companies will need to stop renting intelligence and start owning it:
"We just raised a fundraise very recently. Imagine that ML4 was developed without even touching that fundraise. Expect more from us on coding, more on cyber."
"What will remain true a year from now is that the best alternative to ensure business continuity around AI is to go open."
"The other topic is ROI from general purpose AI and from chatbots is going to meet its limits. We need to empower larger enterprises to transform their core business, and for that you need control and you need customization."
@MistralAI
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
Kino co-founder @lukeigel explains why taste is becoming more important than editing skills as Claude learns to turn conversations into video edits:
"Its like Fight Club. Number one rule of taste, you never talk about taste club."
"When we made that launch video, tons of those moments we only said in Granola, and then we wrote very sparse notes."
"Claude inside of Kino read those Granola notes, laid out those timelines, and left them as comments, very similar to comments in Figma or Google Docs. Then either we or an agent would solve those comments, and you’re just debugging a video from that point."
"We were on the rooftop of our Chinatown office. Donald suggests, what if the video opens with me just chucking the hard drive off the roof? Granola picked that up, we told Kino to try it, and we're like, this works."
@trykino
Kino co-founder @lukeigel explains why Claude Code could wipe out AI video editing startups, but video still needs its own GitHub or Figma:
"Is your startup just going to be washed away by Claude Code next year with a basic harness? My startup for copywriting gets washed away by GPT 3.5. Translation, many forms of coding."
"There's been a massive Opus 5.5 moment in the past four weeks. My co-creator Donald supplied a K-pop song about Claude, and it spent 12 hours orchestrating different models to make something come together."
"We're trying to do the Zen thing where we don’t get in the way of agents like Claude or GPT. We're trying to do all the things that are so crucially missing."
"Infrastructure where the footage on this hard drive is known by your computer and your agent is something everyone takes for granted in coding because GitHub and Git are amazing. We take it for granted in Figma. No one’s done this for video editing."
"Video editing is extremely fragmented. A whole layer of software just for storage, physical devices just for storage, tons of editors just for editing, and then software just for review, like Frame .io."
@trykino
SuperMe CEO @onecaseman predicts AI agents will start seeking advice from other AI agents that represent real human experts:
"A big change happened toward the end of last year, agents started to be able to go multi-turn in a way that was predictable and didn't compound small errors into bad experiences."
"You can think of all of our profiles as individual agents representing that person’s accumulated knowledge."
"When George Washington was born, that's a fact. For a lot of business decisions, there isn't a fact-based answer. There's a bunch of perspectives."
"Whether it's you as the human or the LLM operating a component, if they have more informed data from multiple sources, they're gonna make a more informed decision."
"Right now, mostly humans are pulling us into a Claude session. Hey Claude, look at what Casey and Sophia said about this, then decide what to do. Longer term, a lot more agents are gonna be talking to other agents."
@SuperMeAI
You shouldn't need the best network to get the best insights.
SuperMe turns operators' work into profiles that work through your problem. Now in group chats, Claude and ChatGPT.
@ludo_antonov and I raised a $20M Series A led by @kirstenagreen from @ForerunnerVC with participation from @GreylockVC.
Read more about the changes here: superme.ai/blog/series-a
Eve CEO @jayanthmadhesw1 explains why getting justice in America often depends on whether your case is profitable enough for a lawyer:
"Our dream is to make it so that law isn't just for the rich. A set of facts should win on its merits and merits alone."
"We want to represent all 300 million people in America."
"When you get harassed at work or you get fired, the lawyer representing you doesn’t have to do the math on, is it worth my time to spend three hours helping this person when I might only make $100. So many legal tasks in America just go unseen."
"Veterans get denied disability claims on one blood test, even though they've been having those symptoms forever. You can’t wait six months for an appeal."
"Now Eve is able to take that case on, which before was economically unfeasible. We help lawyers give you a second chance."
Huge milestone: Eve has helped 350,000 people take on a system built to wear them down.
Take Mike Dickerson. A four-car pileup, two hip surgeries, then an insurer betting he’d take a lowball offer.
He didn’t. His lawyers @GeigerGroup fought for years and, with Eve’s help, got him 15x the first offer.
Firms on Eve have recovered $3.5B+ for their clients, and we’re just getting started.
Next milestone: 1 million people helped.
SITUATION EXPLAINED: Google DeepMind researchers say AI has barely changed the economy because AI agents still can't work together like human teams.
• Google DeepMind's AGI economics director Alex Imas, researcher Andrew Koh and co-founder Shane Legg published the essay
• Better individual skills like coding or writing shift wages and jobs gradually, but better teamwork between AI agents could change them suddenly
• Today's models are weak in the same places, so running more copies adds capacity but doesn't cover each other's blind spots
• In a quiz of 100 people, the researchers found that pairs matched to cover each other's weak spots beat random pairs by 32%
• For now, the essay predicts falling pay for workers whose skills AI already matches, and rising pay for specialists and the managers who coordinate teams
• The paper maps 4 stages of AI, and at the "super cooperative" stage, whole industries could consolidate into a few AI-only firms and wages could drop sharply
• Legg told the Financial Times he's "comfortable" with a 50% chance of what he calls "minimal" AGI by 2028
@theojaffee: "The fully automated firm is when I think we get things that look really crazy, like really wild sci-fi, real superintelligence, real x-risk if we do it wrong."
I joined @GoogleDeepMind to develop a research program on the economics of AGI. Today, together with Andrew Koh (@andrewjkoh) and Shane Legg (@ShaneLegg), we are incredibly excited to release The Economics of Intelligence, the first in a series of projects from our group on how AGI will change the economy.
How do capability improvements transform the economy, and when? In this project, we define thresholds of economic capabilities, e.g., minimal AGI, based on how they transform the economy. We then track how the economy–market structure, employment, and wages–changes during the transition across capability regimes.
We have seen exponential increases in what AI agents can do but the broader economy seems almost remarkably unchanged. Part of the reason is that the solo AI agent is the wrong economic object; economic transformation is an equilibrium outcome where organizations compete. Looking at how intelligence functions within organizations, transformation will depend on when two key capability thresholds are crossed: competence (gains on individual skills, e.g., coding, math) and coordination (gains on ability of agents to cooperate and work together).
We show that what economic transformation looks like depends on whether AI becomes supercompetent, supercooperative, or both—the regime of artificial superintelligence (ASI). Tracking this transition shows why the real visible transformation may come from improvements in coordination—a capability most benchmarks don’t yet track. Past a certain threshold, AI capabilities can break the Coasian ceiling on firm size and lead to an economy that looks drastically different than the one we see today.
Two main takeaways:
1. To understand the path of transformation, we should be tracking agentic coordination more closely.
2. The impact of AI will likely show up in wages before market structure or employment.
We are really excited about this paper. Excited to hear people’s thoughts.
Link: institute.deepmind.com/essay…
Databricks VP @nikitabase argues Claude Code is just the beginning, and the next wave of AI software factories will require an entirely new infrastructure stack:
"Claude Code was all the rage, still all the rage. But now we're talking about AI factories. We're moving from a tool that helps you generate code better to a platform where you can ship software as a factory."
"But the place you ship software to is still some sort of infrastructure. The AI factory of the future will work a lot better on this agentic infrastructure."
"State, such as database state, is paramount, and it’s probably one of the harder things to solve."
@databricks @neondatabase
Base director @antoniogm predicts AI agents will become the biggest source of revenue for online publishers within 10 years:
"Programmable money is very nice until the bots start programming money, and then it's like, holy shit."
"At Spindl, we ran a campaign and saw no page views but all these on-chain conversions. We thought they’re gonna fire us. And the client's like, bots are humans too, we’re happy."
"Crypto has speed run all the crazy stuff that's gonna launch now, and you’re seeing it manifest in regular AI. Crypto was the canary in the cage."
"Your bot scrapes 20 sites, and right now those sites don't get paid. The bot pays half a cent for the content."
"Within 10 years, the AI line item on most publisher sites is gonna be the biggest one. Not humans."
@base
Base director @antoniogm predicts that within a year, an e-commerce brand will make $1 million a month without a website, with AI agents doing the shopping:
"Agent commerce is gonna have an inflection point where Muse, Instinct, etc, are gonna be using hooks to enable you to actually buy things. This is gonna be an aggressive prediction."
"There’s gonna be some D2C brand or e-commerce thing that goes totally viral that doesn’t have a website."
"Or it has a website that's literally just a splash card. You're not supposed to see this. The AI agent is. Have them go here."
"I predict that's gonna happen in a year. They might even run Instagram ads, but they're not gonna have a website, and they're gonna make a million dollars MRR."
@base
Dream co-founder @AndyElder11 explains how AI is making designing real houses as easy as playing The Sims:
"We've built in 73 unique types of architectural and interior design styles."
"You can upload something from Instagram that you liked, and we'll find real products, real materials, the backsplash, the flooring, the light fixtures, even the furniture, and design a full concept that’s regulatorily sound and structurally sound."
"I grew up playing the Sims. We all had this feeling of, wait, why is there not something just as easy for the real world? As easily as you drew walls in the Sims, you can now do that, except there’s a layer of intelligence behind it."
@dreamhq
Today @arjunmathur_ and I are introducing Dream.
We’re announcing $17M in seed funding from @a16z , @svangel , @xyz_vc , and @ziggcap.
@dreamhq is developing a new class of models trained to reason about physical space: its geometry, its constraints, and the relationships between objects and the spaces that contain them. With these models, AI can make a real impact in the $13 trillion architecture, engineering, and construction (AEC) industry for the first time.
Dream co-founder @AndyElder11 reveals how his startup is bringing AI to a $13 trillion industry still relying on software from the 1990s:
"We are creating built world models. These are models that actually design the physical world around us."
"The $13 trillion architecture, engineering, and construction industry has really been left behind."
"The building that you're recording this podcast from was likely designed using software from the late ’90s, and it probably took 12 to 18 months to design. The models we're developing can do that in a fraction of the time and a fraction of the cost."
@dreamhq
Today @arjunmathur_ and I are introducing Dream.
We’re announcing $17M in seed funding from @a16z , @svangel , @xyz_vc , and @ziggcap.
@dreamhq is developing a new class of models trained to reason about physical space: its geometry, its constraints, and the relationships between objects and the spaces that contain them. With these models, AI can make a real impact in the $13 trillion architecture, engineering, and construction (AEC) industry for the first time.
Foundry Robotics CEO @adarshkulkarni explains why robots that work 99% of the time still aren't reliable enough to build satellites:
"Something that's 90% reliable fails one out of 10 times. Something that is 99% reliable fails one out of 100 times. Something that is 99.99% reliable fails one out of every 10,000 times."
"So the difference between 90% and 99.99% is 1,000 times fewer errors. That’s why it’s so hard."
"When you're doing manufacturing for critical industries, when you're building drones or batteries or satellites, you really have to be pretty good. You gotta have the nines."
@FoundryRobotics
Today, @FoundryRobotics is announcing our $100M Series A, led by @MarlinspikeVC with continued support from @khoslaventures.
Together we’re securing America's ability to defend, compete, and innovate by building critical hardware at home with our new 130,000-square-foot factory in Emeryville, California.
We’re already shipping to paying customers across energy, telecom, autonomous vehicles, and defense. The next era of American manufacturing starts here.
Time to build.
Foundry Robotics CEO @adarshkulkarni reveals why the factories of the future will look more like data centers than traditional factories:
"When you come into Emeryville and see the Foundry facility, it's not like a traditional factory."
"It is somewhere between a warehouse and a data center and a factory. There’s GPUs under each work cell. There’s fiber lines coming from the ceiling. There’s a mini data center off to the side. Robotic work cells doing bimanual work, mobile robotics moving parts between them, humans collaborating with them."
"I am pretty sure I'm the only founder in Silicon Valley who's gonna brag about his screwdriver. We are building a fundamentally new fastening end effector that is highly reliable, robotics-oriented, and dramatically cheaper than what exists off the shelf."
"Everybody else is talking about how they're going to use a complex robot to hold a drill from Home Depot. Doesn't really make sense"
"By being a factory company, we get to solve these meaty engineering problems that aren't necessarily companies on their own, and roll it into the factory as the product."
@FoundryRobotics
Today, @FoundryRobotics is announcing our $100M Series A, led by @MarlinspikeVC with continued support from @khoslaventures.
Together we’re securing America's ability to defend, compete, and innovate by building critical hardware at home with our new 130,000-square-foot factory in Emeryville, California.
We’re already shipping to paying customers across energy, telecom, autonomous vehicles, and defense. The next era of American manufacturing starts here.
Time to build.
Foundry Robotics CEO @adarshkulkarni explains why America needs to build entirely new factories around AI instead of just putting robots in old ones:
"You can’t just retrofit old factories. You have to build net new factories. We need more capacity than exists today."
"And if you're gonna build factories from scratch, you might as well figure out the most efficient way to do it that includes robotics, AI, and everything in between. Let's zoom out. The factory's the product."
"We have a lot of customers in defense, energy, telecom, infrastructure, industrials. People need to build more things, and we need capacity that doesn't exist."
"On the other side, robots are starting to show up. The foundation model labs are pushing the frontier every single day."
"Our job is to take the base models and deploy them in really important settings, like manufacturing. Laundry folding is great. I want one. So does everybody else. But we have to do the work to go from a laundry folding model to a satellite building model and deploy it at scale."
@FoundryRobotics
Today, @FoundryRobotics is announcing our $100M Series A, led by @MarlinspikeVC with continued support from @khoslaventures.
Together we’re securing America's ability to defend, compete, and innovate by building critical hardware at home with our new 130,000-square-foot factory in Emeryville, California.
We’re already shipping to paying customers across energy, telecom, autonomous vehicles, and defense. The next era of American manufacturing starts here.
Time to build.
Mistral researcher @Sagar_Vaze argues Mistral can catch OpenAI and Anthropic despite starting years later and having far less compute:
"What it would have taken for us to get to the frontier in 2023 was us to be founded in 2020 or 2019."
"This is really the key difference between ourselves, a very new lab, and some of the more established players. We just started a few years later."
"The models we've released today are state-of-the-art in a bunch of key capabilities, finance, law, engineering, some multimodal tasks."
"On agentic coding, we’re the best OSS model. Catching up with OpenAI and Anthropic is absolutely on the cards. That's absolutely the vision of Mistral."
"We don't want open weights and closed weights to have some delta in performance in the long term."
"The thing we need to get there is compute. We’ve raised 6 billion, dollars or euros, a lot of money by traditional standards. But compared to what these other labs have raised, it’s a drop in the bucket."
"We’ve convinced ourselves that we can do more with less, and the next step is to do much more with much more."
@MistralAI
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
Mistral researcher @Sagar_Vaze reveals how Mistral Large 4 matches GLM 5.3 on a coding benchmark while using 50% fewer output tokens:
"It's our first flagship OSS release in a while. It's a trillion parameter model trained on 1E25 FLOPs."
"This FLOP budget and parameter scale is a key milestone in any AI lab's development. We had to build out our own data centers. This model was trained from scratch on our own compute, which we built just outside of Paris."
"It's by far the best agentic coding model we've released, both in our internal testing and with third party evals. On the AA coding index, it's exactly tied with GLM 5.3 but needs 50% less output tokens."
"On these critical domains for our customers, cyber defense, finance, law, we're not only best OSS, but in many cases, the best model, period."
@MistralAI
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
Mistral researcher @Sagar_Vaze argues some of the most serious recent AI cyberattacks came from closed models, not open-source ones:
"If you look at the most significant malicious attacks that have happened recently, these have actually been perpetrated not by open weights models, but by closed weights models which have either been jailbroken or, in some cases, during training, escaped their sandbox. In practice, that’s how many of the attacks have been perpetrated."
"Defenders trying to shore up their own defenses are having their requests refused by these same providers, because cutting-edge defensive requests look quite similar to offensive ones."
"People trying to defend from malicious attacks, often from closed weights models, find they have no recourse but to use an open weights model."
"It's critical that there are open weights models of sufficient strength that people can scalably defend their infrastructure."
@MistralAI
Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity.
💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis.
🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others.
Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it.
Read more from our team and customers 👇:
AI safety researcher @s_r_constantin explains why AI can write code but still struggles to come up with genuinely good ideas:
"Writing is bad. Generally, I don't see hallucinations. I don't see syntax errors in code."
"Often I do see bad ideas if I’m trying to idea generate. I’ve been trying to make a font and they’re terrible."
"Vibe coder websites tend to have a bunch of client-side bugs, where you reload, it goes away, because it hasn't tested all the different possible user paths."
"This is a unanimous thing I hear. What’s the next project? They give you stuff you wouldn’t wanna do."