Thomas Capelle
@capetorch
Chilean 🇨🇱 living in France. I build DL models and pipelines. ML Engineer at @weights_biases cargobike ♥🚴
🤖 Just dropped our new AI Agents course and it's completely free! If you've been curious about building AI systems that can actually *do* things autonomously - like handle complex workflows, work with teams of agents, or tackle real-world problems - this might be worth checking…
this is one of my favorite charts; IEA expert solar forecasts each year vs actual installations people are really bad at wrapping their heads around exponentials
Good taste here.
Fun fact: our model is called Kimi, but our company is Moonshot — named after Pink Floyd's The Dark Side of the Moon. We're a team of scientists who love rock (Radiohead, Pink Floyd) and film (Tarantino, Kubrick). A big reason I joined was because the taste just felt right.
Back from holidays and no open model from the not so openai yet...
Whereas the aforementioned vernacular of jurisprudential discourse, hereinafter referred to as ‘legalese,’ serves as an indispensable mechanism for the perpetuation and preservation of billable hour optimization strategies; and whereas the deliberate obfuscation of otherwise…
Meet your butler for the weekend: A Unitree G2 Pro robot dog delivering snacks while you hack. And if your agent protocol project (MCP, A2A, ADK) is the best? You’ll take it home. WeaveHacks is here: ⚡ July 12–13, SF 💰 $15K+ in prizes 🐕 Robot dog grand prize Join us! 👇
buuuncha stuff merged into verifiers main highlights: - retired janky dynamic batching in vLLM - full vLLM args via AsyncLLMEngine - full response objs in env/trainer - better async rollouts/rewards - lots of bug fixes - non-optional deps are now just these: v0.1.2 real soon :)
Mediterranean food markets, I am on holidays in northern Italy, a super small town, went to the market today, it looks like this (actually better).
This is what the ideal grocery store looks like. Minimally processed (NOVA Group 1) food only (no "edible food-like substances"), organic, local, fresh. Food should not be more complex than this, yet I don't believe this exists.
Merchandise: Shoggoth GF: watermark-4.creator-spring.com/listing/shoggo… Shoggoth RLHF: watermark-4.creator-spring.com/listing/new-rl…
my full talk from AIE world’s fair is out now :)
🆕 Training Agentic Reasoners today's feature is @willccbb's triumphant return to the AIE stage RL track - now as part of @PrimeIntellect! A lot of agent builders are basically doing "RL by hand". He concisely explains current RL algorithms in one slide (!) but then argues…
Big fan of Scott’s eval guide. I like that it’s highly interactive (“choose your own adventure”), and that it distills a lot of good wisdom from the experts. I also personally love that it exemplifies using LLMs to make sense of and succinctly communicate unstructured text info…
How I built this: - @sh_reya's DocETL to help find relevant quotes / tips from my favourite eval guides / chapters / case studies across different key dimensions (defining eval requirements, dataset building, scoring, etc.) and prompt iteration - Claude Code to synthesize the…
The @altryne @thursdai_pod & crew weekly section on open source models has a very global view and were early on qwen, deepseek, etc. A quick LLM analysis on their model mentions over the last 9 months below.
Holiday time in Como Italy. Slack kicked me out due to okta/migration, it's a proper holiday. :)
