Heading to the AI Summit next week? Catch my talk on engineering the context layer to build better AI Solutions. I have 4 big lessons for everyone building. Catch the session on the Finance Stage: The Context Layer: Your Shortcut to AI-Driven Alpha Dec 10, 2:25 PM - 2:50 PM

@burkov Groundbreaking, yes. But let’s strip away the marketing gloss. This paper is indeed a technical marvel, but characterizing the dominance of tree-based methods as an "embarrassment" misses the point of why trees ruled in the first place. Here is a more nuanced take on where

Excited that my team’s new doc is out! It’s a pitch for a more pragmatic direction for interpretability. The reasoning: recent work (science of misalignment, eval awareness, lie deception, trawling) has been most successful when it’s focused on problems, not new methods.

An agentic alternative to GraphRAG. We built a Metadata Search Tool to solve reference traversal without the rigid complexity of static graphs. The result? Agents resolve complex queries in fewer steps with higher accuracy. 🧵 1/4

Did a quick post on GSMA Open-Telco LLM Benchmarks and the Telelogs dataset. I compared the AT&T and Huawei approaches for Root Cause Analysis: https://t.co/kxg7DDdn1V

Have some fun, check out the @kaggle Santa Tree Packing Challenge

Introducing "The Eiffel Tower Llama"!🗼 Remember Golden Gate Claude? Unfortunately Anthropic's viral demo was shut down after 24h, and key technical details remained hidden. So we recreated it, uncovering key insights on steering LLMs using SAEs⚒️ Full blog post + live demo 👇

LLM-as-a-Judge doesn't scale. After 100+ test cases, you're stuck with: → Binary scoring that treats 95% correct the same as 5%. → Judge prompts that drift every update. → Zero insight into WHY something failed. So we built something better. 🧵

Check out the blost post to learn more: https://t.co/GTysk1Mv5r Great work from: @RulinShao @AkariAsai @shannonzshen @hamishivi @varsha_kishore_ @JingmingZhuo @xinranz3 @IAmSamFin @david_sontag @CoachMurray47 @sewon__min @pdasigi @soldni @faeze_brh @scottyih @tongshuangwu

I am excited about RL with Evolving Rubrics Check out DR Tulu from @allen_ai

@ankrgyl If you do this it’s likely because customers are asking for these generic metrics so it’s hard to resist You could go against the grain and say we help you discover your applications failures, not show you generic metrics And show specific errors in an ad. Maybe this helps

My full practice run:

Nano Banana Pro is Amazing This infographic is crazy good and took less than a minute to create. (I have been going through the DR Tulu paper - video soon) Prepare to be overloaded with "useful" infographics (Like all Gen AI, it can hallucinate, so check over everything

Ever use AI to extract structured data from a 400-page financial document, only to spend more time verifying the output than you saved? Template-based tools break on complex docs. LLMs hallucinate. You're stuck manually checking everything anyway. We built something better. 🧵

Stop model shopping. Start learning from users. Feedback & Annotation now in @ContextualAI • Real-time capture • Intelligent categorization • Dashboards for faster fixes Blog post for details: https://t.co/lHueBcRZo2 Try it live: https://t.co/Dc73itsVuo

How to transform UMAP from a black box into a glass box: by using a special type of deep network, we can now compute exact linear equivalents that reveal which features drive each point's position in the embedding. @ArcadiaScience [1/8]

What happens when you build AI for one of the most demanding users in enterprise software? Claimwise is a legal tech company that serves patent attorneys—professionals with advanced degrees in science & law—who need to verify every single AI-generated result before using it in

Ever wondered how massive AI models like GPT-4 and Mixtral work under the hood? In this video, we build a "Mixture of Experts" (MoE) model completely from scratch using PyTorch. This step-by-step tutorial is perfect for anyone in the AI field looking to gain a deep, intuitive

Side effect of blocking Chinese firms from buying the best NVIDIA cards: top models are now explicitly being trained to work well on older/cheaper GPUs. The new SoTA model from @Kimi_Moonshot uses plain old BF16 ops (after dequant from INT4); no need for expensive FP4 support.

Had a great crowd for talking through the evolution from static RAG to multi-agentic retrieval. Talked about BM25, Big Vector, and Multi-Agents. Well run conference with great attendees, Thanks @wandb for #FullyConnected2025 in London.