Clem

From my experience, ChatGPT has been trained to aggressively ask follow up questions, and I feel a single button tap might be a nice way to keep it going

A pattern I like for reducing friction with LLMs is to dynamically render quick answer options or CTA buttons in the UI. Makes the experience smoother especially in domain specific settings. Haven't seen this around much, do you know of other examples?

@bennetkrause yep I agree! But was pretty cool to see the model improve through self-judging. I used Gemini as the judge on the test-set to make things fairer

@clemhus Cool idea! Although some people prefer using entirely different models as judges due to bias within the same model family

RLAIF w/ TRL+GRPO on a single GPU. Co-locate vLLM with the trainer and reuse the same base model as a judge. Policy updates via LoRA; judge runs with base weights. No extra RM server, low latency.

@mr_besher thanks!

@clemhus very clever! hats off.

Does TRL expect/encourage this pattern? Any gotchas? @QGallouedec

For GPU/API-poor setups: reuse the same vLLM process that does GRPO rollouts to batch the judge calls. Simple + fast. Result: in a free-text medical diagnosis reasoning test, this setup on Llama-3.1-8B gave ~+14% accuracy vs. the baseline (48→62%, n=300, single seed)

having dinner with @garrytan wasn’t on my bingo card when I joined the bay 2 months ago amazing day at @ycombinator AI startup school thanks for the invite garry!

@ramealexandre Congrats !

Thanks for leading this post-training, with great engineering and research breakthroughs! Super achievement by the whole team.

this is insane

@arnie_hacker See you there !

Update: Going to SF this Summer Thanks to all those who helped <3

@jakub_lucki @schickkler Nice Jakub! Let’s see if we can organize something there

@schickkler I'm also going! Wanna do sth?

I'm going to ycombinator AI Startup School in Sf!!!!!!! Hmu if your going, especially if your from Switzerland/Europe

@arthur_bigt Thanks Arthur ;)

@clemhus Congratulations!

We won Meta's Llamacon Hackathon! Imagine engineers, PMs, or managers being able to query their org’s knowledge base in natural language: 🧠 “Who’s responsible for the normalization layer of the Llama 4 models?” 🔐 “Who’s an expert in our multi-step auth flow?” Had a blast hacking this up. Btw, we didn't use Cursor/Windsurf. I actually believe that planning ahead and understanding the exact structure of your code is key to deliver a project you own.

Was great to get invited at Meta HQ to discuss our experience working with the new Llama API. A unified way to use all models from any provider is the way for developers. Performance and cost rule!

Let him in! We gotta have a bigger Zurich delegation

@torayeff @MetaforDevs @cerebral_valley congrats !

Last week, I was in SF and, by a lucky chance, participated in the first LlamaCon Hackathon — and surprise: I won 3rd place and a $6,000 cash prize among top devs in Silicon Valley 🎉 Huge thanks to @MetaforDevs and @cerebral_valley for hosting this great event!

@gdb Love this. We hacked on something similar at Llamacon, mapping GitHub repos into a knowledge graph and layering LLMs on top. Super powerful direction.

Deep research for your codebase:

@MetaforDevs was fun cooking with the new Llama API 🦙

We're excited to announce the winners of the first LlamaCon Hackathon! These talented individuals and teams have demonstrated exceptional skill and creativity in their projects using Llama. 🥇 1st Prize: OrgLens An AI-powered expert matching system that connects you with the right professionals within your organization. By leveraging data from various sources, OrgLens creates a comprehensive knowledge graph and detailed profiles, streamlining expert matching. See their GitHub Repository: bit.ly/4k62WaO @KPJedrzejewski, @clemhus 🥈 2nd Prize: Compliance Wizards An AI-powered transaction analyzer designed to detect fraud and alert users. It uses Llama API’s multi-modality to assist fraud assessors in determining client involvement in criminal activities. See their GitHub Repository: bit.ly/3RRxiBS @SamDc73, @k_a__reem, @nicetomeetyu2, @sorhanft 🥉 3rd Prize: Llama CCTV Operator A Llama CCTV AI control room operator that identifies custom surveillance video events without model fine-tuning. It uses Llama 4’s multi-modal image understanding to assess and report predefined events. See their GitHub Repository: bit.ly/4d9UPrw @torayeff 🌟 Best Llama API Usage: Geo-ML This project uses Llama 4 Maverick and GemPy to generate 3D geological models, processing extensive geology reports into structured data for 3D representations. See their GitHub Repository: bit.ly/3GITT15 @WilliamJSDavis Please join us in congratulating these winners on their outstanding achievements! We're honored to have them as part of the Llama community. 🎉

@cerebral_valley @MetaforDevs Was amazing !

SF's LlamaCon Hackathon just ended — here are the winning teams from the 600 hackers that applied for a chance at $35,000 from @MetaforDevs (save for project inspiration later 🧵):