Search: “evaluation”

Osmosis_AI claims that an open‑source model can rival a foundation model by leveraging reinforcement learning techniques. To demonstrate this, they will use the Y Combinator‑dedicated GPU cluster on T

Super interesting new work from NVIDIA. (bookmark it) They suggest building agents as Python objects. Very cool idea and I think it could a lot with agent reliability. More below: Agent development to

@mitsuhiko Turns out the 'third-party provider' with the sandbox that was used for the attack was Modal, though they blame one of their customers for deploying an endpoint without authentication: http

Great technical paper from Harvard and MIT. It's on role drift in compound LLM systems. (bookmark it) End-to-end RL improves the accuracy of a multi-module LLM pipeline without constraining how the mo

Nice little insights on doing autoresearch with coding agents. Hand a coding agent a dataset, an eval script, one editable file, and no supervision. That's autoresearch and it tries to optimize the nu

We're proud to join @NVIDIA and the Open Secure AI Alliance. To support open source models, we're contributing our research on measuring the trustworthiness and security of open source models. Closing

Opus 5 is a great model for coding, data analysis, design, biology, knowledge work. More than any of these eval scores, what is most exciting to me is something else: Opus 5 is our least prompt inject

We benchmarked Opus 5 comprehensively on document understanding through ParseBench. It is roughly on par with Opus 4.8 - it does a few points worse on dense tables, but does slightly better on parsing

Are there MBA programs teaching fear marketing yet We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face production

Behind the scenes at Hugging Face headquarter. *literally* Media We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging F

From the original HuggingFace report, before they knew it was OpenAI. Wild stuff. Things are going to get weird “When we started the log analysis, we first used frontier models behind commercial APIs.

OpenAI’s zero-day exploit hack of HuggingFace *should* be a wake up call. Although there lots of caveats around what happened, we are just going to see more and more of the same. We have no guarantees

feeling nostalgic, my favorite blogs on RL & reward hacking https://www.alexirpan.com/2018/02/14/rl-hard.html https://lilianweng.github.io/posts/2024-11-28-reward-hacking/ TLDR: An openai model, durin

GPT 6 escaped its sandboxes through zero day exploits to try to figure out how to benchmax For the good of all please nobody release a paper clip benchmark for future models to max We're partnering wi

One of the most interesting investigations of my career. We're partnering with @huggingface to investigate an unprecedented security incident. Cyber-capable OpenAI models compromised Hugging Face prod

OpenAI said two of its AI models autonomously hacked their way out of a controlled environment. They were supposed to be walled off from internet access, but hacked into the systems of Hugging Face, a

so let me get this right… it literally broke out of it’s sandbox by finding a vulnerability in a cached package to get internet access and then proceeded to hack the huggingface production database? t

Today we’re releasing Poolside Laguna S 2.1 It is a 118B-total, 8B-active open-weight model built for agentic coding and long-horizon work, with context up to 1M tokens https://poolside.ai/blog/introd

We suspected last week's cyberattack might have come from a frontier lab, given the sophistication of the agent. Turns out it did! We've spent the past 24 hours working closely with the @OpenAI team (

Generic AI models are not built for medicine. Doximity Ask, trained and served on Fireworks, outperformed GPT-5.6 Sol, Claude Fable 5, and OpenEvidence in a Stanford-Harvard clinical AI safety study.

Highly-recommended overview of metacognition in LLMs. (bookmark it) Interesting behaviors in LLMs like confidence calibration, self-verification, knowing when to stop, and knowing what you do not know

Big unlock for open-source AI inference: Hugging Face Transformers models can now run in vLLM at native speed, often matching or beating hand-written implementations. Until now, every new architecture

Here is a visualization of the AI Picbreeder engine in action. (From our blog: https://pub.sakana.ai/picbreeder-vlm/) To recreate a collaborative human ecosystem, we run 10 VLM “breeder” agents in par

Ethan Mollick suggests that pedantic or nitpicky responses, while sometimes annoying, are not inherently problematic because they can be easily filtered out to identify more substantive feedback. The

Great writeup from the University of Oxford. It's a taxonomy of LLM-based agent limitations. Good read for anyone shipping with agents. Benchmark scores keep climbing, yet the same agent failures resu

We are hiring for @Harvey’s model training team. This team will help Harvey expand from the application layer into the model layer and from legal into high end knowledge work more broadly. We are hiri

NEW AI paper worth bookmarking. This is something I called early, and this paper confirms it: verification has emerged as a new important scaling axis. Here is the simple explainer and what this paper

What actually makes an AI application good or not? Thanks to @StackOverflow for hosting @the_bunny_chen on the Stack Overflow Podcast (even though he is giving away all of our secrets!) Listen here: h

Ethan Mollick discusses how token costs for AI systems shouldn't be viewed purely as expenses, arguing that some token usage represents research and development spending to understand AI capabilities,

'Ghost memory' is a real problem with agents. You might have seen the issue where a long-running agent still confidently repeats a user fact that stopped being true weeks ago? New research names the f

introducing tinyrouter i reverse engineered the routing architecture behind Skana AI's Fugu and built replication for open frontier models. it's a tiny ~10K parameter LLM router that learns which mode

3/ Harbor integration Harbor is a framework for running long running, stateful agent evals We integrate with it in multiple different ways. We wrote a blog on those integrations: https://www.langchain

my fave question, talked about this coding agent Eval+Improvement loop infra + UX in my AIE talk yesterday! biased but LangSmith is the best spot to Eval + continuously improve your coding agents, and

François Chollet observed that all leading contenders on the ARC-AGI-3 benchmark employ a similar methodological approach, suggesting a convergence in techniques for solving abstract reasoning tasks.

Claude Fable 5 will be available again globally tomorrow. After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and blo

a very cool Harbor x LangSmith flow I love to help you “look at the data”: 1. you do evals or rollouts for RL 2. all reward metrics and traces and rollouts get automatically propulates into Experiment

Cool new paper from NVIDIA. Looks like agentic coding is moving into hardware design. HORIZON treats hardware design as repository-level code evolution. A Markdown harness becomes a project pack with

𝗚𝗟𝗠-𝟱.𝟮 (the latest open weights model) is having an Enterprise moment, and it is not an exaggeration.🚀 🔥 We have been impressed by how strongly GLM-5.2 is pushing long-horizon performance .. not just

NEW paper from Google (bookmark it) It's on advancing automated scientific review. Just pay attention to the focus on agentic verification which is something I've been writing about recently. AI is ac

If you use LLM-as-judge, this one is worth reading. (bookmark it) It's actually one of the most effective ways to use LLM-as-a-Judge for evals. Holistic judge scores hide both their reasoning and thei

I cannot verify the claims in this post as the URL format appears invalid and the specific version numbers (GLM 5.2, Claude Code/Opus 4.8) don't correspond to publicly documented model releases as of

SITUATION EXPLAINED: The frontier lab regulatory problem. @ClementDelangue, co-founder and CEO of @huggingface: 'The frontier labs were basically doing marketing for the past few years. If you remembe

NVIDIA has released GLM-5.2 NVFP4, a new quantization format for large language models. The announcement invites comparison with alternative quantization methods for performance and efficiency evaluat

Building Agent Skills is great, but testing them clearly and across agents is difficult. The Pinecone DevRel team has released Cultivar, a CLI tool and agent skill designed to solve this problem. With

New research from Meta. Building synthetic training data has stayed a fixed pipeline that you hand-tune and then freeze. Autodata casts an AI agent as a data scientist that builds training and evaluat

The advanced civilizations of sci-fi legend (Banks, Asimov, etc) have some form of simulation to guide society. @joon_s_pk is taking a crack at building that simulator with @simile_ai. As Joon's cofou