Jaume

Cette propension des entrepreneurs français de se réjouir des difficultés des autres entrepreneurs français est insupportable…

PRO TOP Do not create Skills with markdown files, you should be creating Skills with deterministic scripts

Most eval tooling is very offline-biased The better models get, the more the online part matters > ingest prod traces at scale > enrich them (signals, clustering) > hand the issues to your coding agent via MCP > monitor whether prod got better Offline evals are the inner loop

fun fact: I still do most launch videos for Krea myself. I made this one using a tool I built over the weekend, we may add it to Krea soon. here I prompt it to add sparkles in the green region; you can talk to it, have it designing on Figma, and "train" agents to learn the kind of visuals or motion you like.

Call me an old hat computer scientist but the parallels between Inference time scaling ---> Subagents --> Agent Swarms looks an awful lot like CPU Threads --> Multi-core CPUs --> Multi-Node Distributed Systems History rhymes well.

Cloudflare is taking ownership of tsrs, a port of TypeScript to Rust. It's 8x faster and uses 70% less memory than TypeScript 7.0. We are not forking TypeScript – we are working with the TypeScript team and as a result 7.1 is already ~2.5x faster, too. blog.cloudflare.com/typescri…

Portal jumping through gaussian splat worlds with WebGPU. real time. sound on.

ANOTHER UK AI lab has just emerged from stealth. This one is founded by AlphaGo co-creator and former Google DeepMind research lead Thore Graepel (@ThoreG) Metis Reasoning is tackling one of the largest problems that exists in AI. Currently, AI can give you a convincing answer but it is impossible to understand how the AI got to that answer, and whether the steps it took to get there were justified. Inspired by the principles behind AlphaGo, Metis's answer is to rethink how AI reasons. Rather than relying on models to generate plausible answers, the company is developing an approach in which systems establish conclusions through gathering evidence, testing assumptions and verifying their progress. Thore has been public about building something in this space, but the company itself has been in stealth until now. The founding team have also released a white paper labelled "The Age of Reasoning" which sets out the full thesis. NICE

we built “let’s bump phones to exchange contacts” but for agents! now your @instinct / @bot / @muse (whatever you use!) can make plans, negotiate deals, and get what it needs from your friends/coworkers’ assistants it's “have your ppl (agents) talk to my ppl (agents)!” for everyone 🤝 built w/ @uncanny_guzus @ @ycombinator #MakeSomethingAgents hackathon (s/o @nemild) on top of @stripe (agentic p2p payments) + @manufact (mcp relay server) 🙌 Who will you bump first?

Introducing SKY-001, the first Engineering World Model. On CAD and BIM tasks, it beats GPT-6 Astra and every frontier model we tested: -> 51-125x faster - > 150-210x cheaper -> Higher IOU than all frontier models SKY-001 lets you dream a physical object before you build it. 18 months of work. Built by the team out of Maluuba's deep learning lab (acq. by @MSRResearch). The first ‘AI General Engineer’ will be the backbone for Autonomous Manufacturing.

Every project I maintain runs with fail_under = 100. I wrote down why 🧪 - 95% hides a blind spot that grows with the codebase - A 100% gate stops agents from landing untested code - Test through the public API, never private functions marcelotryle.com/blog/100-pe…

MCP haters overlook this point. They forget that an agent is really two parts: your harness (or client) and the underlying model (LLM). MCP can skip the model and communicate between the client and server only for e.g. destructive actions. APIs and CLIs can't.

Data engineers > AI engineers there's a big difference between a "RAG guy" and a "Data Intensive Applications" person like in this screenshot and there's never been a starker difference in demand in my entire career in tech (aka post search engine wars) if your blog / github is full of stuff like this, I have a fantastic mission/job for you (SF preferred, can relocate u, visas ok) (if you can name who's blog this is without even checking, you definitely are in the community. bonus points if you know where he ended up also without looking) [email protected] and the team will reach out if you fit

I've been critical in the past of AI compute indexes and the willingness to price compute based on the spot rental price, but now that the AI reset is sinking into the repricing of AI buildout debt its time to realize that the cost of capital is what determines the price of compute. The most important number in GPU pricing right now isn't $/GPU-hour. It's the gap between what a hyperscaler pays to borrow (~90bp over Treasuries) and what a neocloud pays (~1,100bp). The cost of compute is now set by the cost of capital, and the spot rental rate no longer tells you much. The price of an H100 isn't one number. The same chip in the same rack costs a different amount depending on whose balance sheet it sits on. A ~1,000bp funding gap is a structural cost advantage that no amount of operating skill closes. For example, take one GPU-hour apart. Most of it isn't power or space. It's paying back the chip, the interest on the loan that bought it, and the return the equity expects. A $40–55k all-in GPU, 75% debt-funded, at ~7,000 utilized hours a year is a financing instrument that happens to do matrix multiplication. As the +500bp on neocloud debt (10%→15%) adds $0.20–0.30 to the break-even cost of every GPU-hour, which is then added on to the rental price. That's 8–10% of the H100 index, from a credit spread alone, driven almost entirely by the cost of capital. That means that the compute spot index is a derivative of the AI infra debt credit curve and the squeeze on neocloud margins.

CAP theorem for inference providers: reliability, cost, and speed

I've been building loads of personal software in the loop where I just build the app directly on top of claude code. No agent framework. Just smash claude code into exe dev machine, send stuff to it via claude code signals and let it vibe code whatever helpers it needs to create a good looking UI. Full personal context is the most important part here. If the agent can just do whatever you can do from your computer your personal vibe coded apps become a super power.

Agentic obliteration is coming. Entire markets were reorganized around the internet in the early 2000s. Tower Records, Blockbuster, and the Yellow Pages fell, and iTunes, Netflix, and Google took their place, taking advantage of cheap distribution. It happened again in the 2010s around cloud and mobile. BlackBerry, Kodak, and taxis fell, and the iPhone, Instagram, and Uber took their place, taking advantage of a computer in everyone’s pocket. Now it’s going to happen with agents. Incumbents can’t simply tear down their UIs and business models to pivot. Public shareholders won’t allow it. Access to intelligence has been democratized, and categories that have historically monetized gatekept institutional knowledge will be most susceptible. Law is obviously at risk (Harvey and Legora are exploding). Medicine will also change (Doctronic is already renewing prescriptions in Utah). And personal finance will be equally challenged. Should be an interesting decade.

Framework defined infrastructure makes no sense in an agentic world. It made little sense before but at least it was cool to quick start projects with, now it's completely irrelevant

Google just open-sourced Harness². (turn any harness into a self-improving system) a harness is everything wrapped around the model, like the system prompt, tools, memory, skills, and subagents. today a person designs it once, and the agent runs every new task with that same fixed setup. Harness² lets the agent refine that setup for each new task on its own, with the model weights left untouched. on every task, the agent proposes edits to its harness, runs a few candidate versions, and composes a better harness by comparing how those runs behaved. because it learns from the differences between runs, it needs no correct answer or grader for the task. that covers a single task. Harness² also carries what it learns from one task into the next. after each task, it saves the harness it settled on along with the lessons that held up across the runs. the next time it refines a harness, it starts from those notes instead of from scratch. this means every finished task makes the next round of edits more informed. Google tested this with Gemini 3.5 Flash and GLM-5 on legal, office, and professional work benchmarks. one round of edits was enough to beat their original harnesses, and giving the agent more rounds pushed scores higher. those extra rounds can go wide or deep. the agent either tries several candidate harnesses at once or keeps refining a single one step by step, using OpenCode or Codex to carry out the work. you can find the Harness² repo here: github.com/google-research/h… I wrote about a similar idea for agents already running in production, where the harness should repair itself after a failure instead of waiting for someone to debug the trace by hand. that piece walks through how @Cometml's Opik (open-source) closes this loop, going from a failing trace to a proposed fix to a regression test that stops the same failure from coming back. it's quoted below.