Really nice report. Follow up: why has the fall in AI prices been so fast? When you plot the price decline against cumulative R&D investment rather than time, you get the elasticity of price declines to R&D investment. By this margin, AI is not unusual – its price elasticity to R&D investment is squarely in the middle of Epoch's considered technologies. So the AI price fall is historically unprecedented because we've dumped money into AI R&D at a historically unprecedented rate – and that R&D has paid off at a very average rate.

i've been surprised at the response to Jev, but it makes sense in retrospect. sure it's just a classifier but it's a zero shot classifier with frontier-ish intelligence. i'm surprised someone hadn't built it before. i wonder what other old ML ideas are also worth rescuing

In 1995, I would have stood here explaining why you needed a website. In hindsight, a website was never an advantage but a requirement. The firms that went without lost their customers to the firms that didn't. AI will play out the same way. Intelligence will not differentiate you, because everyone will buy it off the same shelf. The advantage that holds comes from what you build out of the residue of every customer interaction: context on your customers, and on your relationship with them, that no competitor can reconstruct.

You're growin' up fast, kid. I remember when I could fit you on my laptop and bounce you on my knee. Could never get you to sleep back then. Still nobody can. Gonna get you a nice ZK prover for your birthday. Happy 11th.

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. images.nvidia.com/pdf/Open-W…

On one side the frontier models from @AnthropicAI and @OpenAI On the other near-frontier models like Moonshot AI in China While Anthropic and OpenAI approach the $1 trillion valuation, Moonshot AI current valuation stands at $30 billion Since DeepSeek several months ago, and now Moonshot AI, the U.S. lead at the frontier of AI is narrowing, even as the technology becomes central to national security, economic power, and geopolitical influence As I wrote previously, AI labs will have to invest either in distribution and applications, the way Anthropic is positioning itself with Claude, or in infrastructure (data centers, chips, and so on), the way OpenAI and SpaceXAI are

I love Satya's essays, because they push you toward deeper reflection Everyone is focused on the model, the benchmarks, the subscription etc, but the paradox is this: to benefit from this intelligence born of compute and algorithms, we have to feed it what makes us unique → Context, vision, methods, preferences, decisions, corrections, and the rest The model then, is only one part of what it takes to reach a genuinely relevant result. the real capital lies in the layer that retains memory, coordinates the tools, and accumulates knowledge over the course of working with the model. And it's the same for companies and individuals. you have to own that layer, protect your footprint, and be able to switch models without starting from scratch - don't be depend over one model! That's why the Pi Agent harness is the most important piece of software I've used over the last six months. to give a bit more context: models remain important, and they serve as the engine in a car but Pi Agent is everything else: the wheel, the GPS, the dashboard, the chassis Models and engines need to be easily interchangeable as the structure of your car should never make you lose sight of the destination, nor the experience of the journey

The two best models on earth are still gpt and claude and they've been climbing the same curve for years, trading the lead every few months. there's no clear winner, but we start to see open-source competitors: deepseek, qwen, kimi glm may will commoditized the model layer and is getting competed toward zero margin. so the only thing that matters now is where you go to escape it, and anthropic and openai split in two opposite directions. anthropic climbs UP on top of the model into the application layer: code, design, cowork, tag and the rest. the goal is to stop selling a chatbot and start selling the labor, own the workflow the filesystem the inbox the slack channel where work actually happens and let the model underneath become a swappable part. they subsidize that layer hard, doubling cowork limits, handing you more value than the sub costs (not generosity) a growth engine built on habit. openai goes the exact opposite way DOWN under the model into infrastructure, $500B on stargate, $1.4T floated over eight years, custom broadcom silicon, gigawatts of texas power, owning the compute floor so nobody can match inference at scale. so, one owns everything above the model the other everything below it and the model gets squeezed from both sides, the model alone is no longer a moat but it stays strategic for the narrative and the ipo that's coming, and the sharpest tell is elon, xAI folded into spacex and turned into a compute landlord renting colossus to anthropic, google, cursor instead of fighting on models. whether that was real foresight or just weak grok adoption dressed up for an ipo, he landed in the right structural spot before almost anyone. so yes keep reading the benchmarks to always use the best model, but who leads ai in ten years is still wide open and it won't be decided by the models alone

RT @levelsio: 🇪🇺 Chat Control has passed 😔 They can and will now legally scan any person's messages, emails and photos you send without a…

Open-source Chinese models are getting very smart and I think we are moving toward a time where models become a commodity. That is why Claude is trying to build an Apple-like ecosystem, keeping the application layer inside Claude The real bottleneck will be chips and hardware. There will still be opportunities to build applications on top of models and sell very specific services to very specific niches. In any case, open-source Chinese models will probably be adopted quite fast as the gap between US AI labs and Chinese AI labs keeps getting smaller.

We're living through a paradox On one side, you've got the tech people, completely hooked on the latest models, switching the day a new one drops, paying whatever it costs to chase the last 10% of performance and they treat a marginal edge like oxygen (i'm in that category). On the other side, you've got the entire generl public, they've heard of AI, they know it exists, some open the free model (GPT/Claude) once in a while, most never touch anything at all, and they genuinely believe they're in the same game as the first group. Here's what people miss: the gap between those two populations isn't closing, it's widening every single week, and it's not a gap in tools, everyone can technically access this, it's a gap in what you do with it and how early you get there. And this is the part that actually matters, the access to information that defines where the opportunities are, the arbitrage, the edge, that's tilting hard toward the people running the best models, they see patterns earlier, they synthesize faster, they ask better questions and get better answers, and they compound that advantage daily while everyone else is still deciding whether AI is worth twenty bucks a month So the real divide of this decade won't be who has access to AI, almost everyone will, it'll be who understood early that the model you use is the ceiling on what you can see.

In pure benchmarks yes. but in multi-step agentic workflow + routing, it becomes very interesting (especially with the 1M context).

It will be interesting to see if they downgraded Fable's pre-shutdown performance. In recent weeks, GPT has performed better than Claude. Let's see what happens with GPT 5.6

Best econ paper I've read in months about AI, the one from Jones & Tonetti. The core idea is very simple Automate 99% of a job and the last 1% you couldn't automate becomes 100% of the work, the constraint never disappears when you automate around it, it just moves and quietly becomes everything the paper answers the question 'we automated so much, so where's the growth' but we're measuring the wrong thing think about building a house, you can prefab the walls, the roof, the windows in a factory, fast, perfect, ready in days, but if the permit takes six months to come through, the house takes six months, all that speed upstream doesn't matter, you're still standing there waiting on the one piece nobody could rush. so the whole thing still moves at the pace of whatever you couldn't rush token usage will continue to increase and bottlenecks will gradually disappear. productivity will increase, just probably not as fast as people expect so at the end, the real edge goes to people who can either remove those bottlenecks or keep enough context in their head to work around them

Nice piece from @desh_saurabh, the trust-by-proof (zkVM) vs trust-by-policy (canton) framing is exactly why institutions hesitate today, and my bet is they tip toward trust-by-proof over the coming weeks and months If a significant part of thesis lies in ZKsync not having competition as Ethereum institution infrastructure, I would recommend having a second look. ZKsync might be the loudest right now, but they're far from the only ones, and most of them are making progresses quietly time will tell

Tdlr: with Claude Tag we're moving away from fragmented, private workflows and toward transparent, async collaboration. Self-closing loop where intelligent agents independently triage a problem, map out the steps, and ship the solution Claude has again, crushed a lot of VC-funded startups. They are building the best model in the world, while keeping the value within their own ecosystem. This is important to reach the $1T in revenue that Dario Amodei is aiming for.

A new player enter the game: @SakanaAILabs Raised $400M and now competes with Fable & Mythos across several benchmarks... what's interesting is that Fugu is not a frontier model, but an orchestration layer. A model decides which other models should work, how they should collaborate, and how their outputs should be combined. On one side, you have frontier models doing everything they can to build the smartest models in the world. On the other, orchestration layers that make use of all these models, get them to work together, and achieve excellent results with far fewer resources.

friends don’t let friends write slop docs

RT @levelsio: This is how you do a great startup video Humble and chill and a real cool backstory