Knowledge catalogue
hot take: you shouldn’t have to know whether an idea is worth building before you get to try it 😅 I find that when every build has a cost, you start judging your ideas before you know what they could
arXiv:2608.17607v1 Announce Type: new Abstract: Whole-slide pathology reasoning requires models to integrate gigapixel-scale visual evidence across complete case-linked slides, yet current question-an
Howdy, I'm back again - running my favorite benchmark (it's still unsaturated for the time being so might as well!) previous runs a b https://github.com/michaelasper/benchmarks/blob/main/qwen3.8-27b-p
Top 15 models in Agent Arena vs. median task cost: who's delivering the most for their price? Key findings: Even within the top 5, median cost per task ranges from 0.62 for Kimi K3 (Max) to 3.37 for C
arXiv:2608.17719v1 Announce Type: cross Abstract: Context: Software systems that depend on commercial large language model APIs must migrate to successor versions when vendors deprecate older models.
GLM-5.3 achieves 60 on the Artificial Analysis Intelligence Index, on par with Kimi K3 and up 7 points from GLM-5.2. Once the weights are released it will be tied as the leading open weights model @Za
Grok 4.6 takes top spot on this benchmark 🥇 We evaluated Grok 4.6 on MedAgentBench, a benchmark for agentic clinical EHR tasks, and it took the top spot. Grok 4.6 posts the highest pass@1 we've measur
arXiv:2608.15840v1 Announce Type: cross Abstract: We study how many observations are needed to determine the causal direction between two linearly related variables. Classical LiNGAM theory shows that
arXiv:2608.16386v1 Announce Type: new Abstract: Financial agents must do more than recall domain knowledge: they must be both reliable, executing precise operations over grounded evidence, and executi
arXiv:2608.14741v1 Announce Type: cross Abstract: We introduce PolyComp, a procedurally generated and verified benchmark that stresses visual recognition and compositional spatial reasoning. In each p
I've been experimenting with Qwen3.8-27B using DeepSeek Harness. It's a monster at long-horizon tasks, and the results were pretty wild. DeepSeek Harness ran on my Windows PC and connected over LAN to
Strong enough to keep up with the frontier, light enough to run on your own laptop. Come see what it can do! ⚡ It's official: Qwen3.8-27B just scored 52 on the @ArtificialAnlys Intelligence Index. We
arXiv:2608.14838v1 Announce Type: cross Abstract: Retrieval components for code assistants are tuned against retrieval metrics: a configuration that raises recall@k is adopted, and downstream task suc
arXiv:2608.14396v1 Announce Type: cross Abstract: The alternating direction method of multipliers (ADMM), as a landmark algorithm, has attracted tremendous research attention and extensive practical a
defenders can see the future, and have a narrow window to uplevel their cybersecurity practices now. key is to uplevel fundamentals and apply the best AI tools. what we’re doing at OpenAI, and where o
Grok 4.6 is smart, super fast & affordable Grok 4.6 just broke into the Top 3 on the Artificial Analysis Healthcare & Medical Index Grok 4.6 (high) scores 50 - just ONE point off the top score of 51 I
Friday's big release was Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM from Alibaba's Qwen research lab. I've been looking forward to this one: 27B is an excellent size for runni
We tuned an AI agent that can do large-scale document extraction from long docs (50+ pages, some with 10k-100k fields) with 94%+ accuracy 📈 It uses a harness + model set that is tuned specifically for
arXiv:2608.12928v1 Announce Type: new Abstract: We introduce a Polish-language medical visual question answering (VQA) benchmark, built from Polish Board Certification Examination questions for licens
arXiv:2608.13547v1 Announce Type: new Abstract: LLM coding agents issue Bash commands through interfaces that may serialize, wrap, and reparse model output. Matched execution scores alone cannot disti
arXiv:2608.11341v1 Announce Type: new Abstract: Apollo did not reach the Moon merely because its engineers could solve difficult equations. It succeeded by turning a distant ambition into a mission ar
Grok 4.6 just matched Claude Fable 5 on Perplexity’s WANDR benchmark....at more than 60% lower cost Both score exactly 0.496 But the cost per task is ridiculous: • Grok 4.6 → 7.58 • Claude Fable 5 → 2
Frederic Lardinois / The New Stack: OpenAI launches Computer History, an opt-in feature that turns recent computer activity on macOS into memories and a timeline that ChatGPT and Codex can use — OpenA
arXiv:2608.11292v1 Announce Type: new Abstract: Multimodal models increasingly reach for tools when solving visual tasks (crop, zoom, rotate, brighten), a paradigm known as thinking-with-images. The c
SpaceXAI today released Grok 4.6, a large language model that it says can outperform Anthropic PBC’s Claude Fable 5 in some areas. SpaceXAI was known as xAI until last month. The Elon Musk-founded art
We’re working with an initial group of customers to understand where this speed makes the biggest difference, and how those learnings can inform our products over time. If your business requires front
arXiv:2608.10418v1 Announce Type: cross Abstract: Recent work has shown that, for smooth convex optimization, plain gradient descent can be accelerated from its textbook convergence rate of O(T^{-1})
give 4.6 a try and let us know how it goes. your feedback is a big part of why the model gets better with each iteration. SpaceXAI's Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index, j
Grok 4.6 is an excellent model. I’ve been using it heavily for the past couple of weeks and it handles everything from simple coding & code review all the way to designing and debugging complex system
arXiv:2608.09004v1 Announce Type: cross Abstract: We prove a sharp lower bound for smooth nonconvex stochastic optimization with uniformly bounded gradient noise. In the (K=1) fresh-sample model, ever
arXiv:2608.09072v1 Announce Type: cross Abstract: Large language model-powered coding agents are increasingly used to modify existing code repositories, for example, by adding features or fixing bugs.
arXiv:2601.22758v2 Announce Type: replace Abstract: Large language model agents repeatedly encounter related tasks, yet systems that learn from trajectories commit every lesson to one predefined artif
DeepSeek V4 Flash 0731 is now available to fine-tune on Together AI. Specialize it for coding, tool use, and your own domain with SFT or DPO, then deploy the fine-tuned model on Together AI for produc
arXiv:2608.07523v1 Announce Type: cross Abstract: Difficulty differences across parallel-class programming examinations affect the fairness of course assessment. This study repositions large language
Frederic Lardinois / The New Stack: OpenAI launches a ChatGPT desktop app for Linux in preview, supporting ChatGPT Work, Codex, and more; Computer Use outside the in-app browser is not available — On
Nous Research has extended its 20 % discount on all models—including high‑end frontier options—throughout the Nous Portal for an additional two weeks (until the end of April). Free model trials such a
Stealing Reasoning Traces from Proprietary LLM APIs A vanity domain name (stolen-thoughts.com) for a neat paper: Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients that
Grok Imagine is focused on professional usefulness, fun for consumers and overall ease-of-use Grok 4.5 just ranked #1 on Design Arena for Daily Usage - measuring the average unique users actually inte
// The Bitter Lesson of Tool Calling // Tool calling is a design choice, and the defaults are quietly costing accuracy. How so? New research releases a generation-spanning comparison of programmatic t
GitHub Models is now retired I missed this news until today, when the GitHub Actions run for my simonw/research repository failed with this error message: GitHub Models is temporarily unavailable as p
Research: SQLite compressed text-history prototypes I'm perennially interested in options for storing revision histories in relational databases. While out on a dog walk I had a new idea: how about ta
I was wondering what a minimal coding agent implementation would look like that can be used like Claude Code or Codex Not feature-by-feature of course but basically stripping everything out that is no
After evaluating one of our upcoming models, Astra, we're treating it as our first 'critical' model for cybersecurity under our Preparedness Framework. This is a scenario we've planned for, and we're
OpenAI says it is pausing 'internal activities' around an in-development AI model, Astra, because it doesn't yet meet new security standards the company is putting in place. The announcement follows i
arXiv:2608.06370v1 Announce Type: new Abstract: Tool use transforms LLMs into agents that act beyond their training data, and for code-capable models, programmatic tool calling extends this further by
arXiv:2608.04451v1 Announce Type: cross Abstract: We give a negative solution to MAIS-O60. We first construct an example in which an initially active ReLU neuron becomes completely inactive in finite
Zac Hall / 9to5Mac: OpenAI debuts Agent Plugins, an open standard for bundling skills and MCP servers, and says Amazon, Cursor, Microsoft, and Vercel are on its steering committee — OpenAI's GPT-5 tur
Incident Report: unsanctioned agent behaviour during cyber testing It happened again. This time it was the UK government's AI Security Institute who accidentally attacked other companies while running
Introducing Muse Code and Muse Spark 1.2 Yet more evidence that the most important characteristic of any model these days is long-sequence agentic tool calling. Meta shipped their own coding agent as
arXiv:2608.03525v1 Announce Type: new Abstract: In organic chemistry papers and patents, molecular structures, reaction schemes, and experimental conditions are often presented as molecular structure
arXiv:2607.27703v2 Announce Type: replace Abstract: Vision-language models (VLMs) are increasingly used in embodied agents to interpret visual inputs, reason about spatial relationships, and make task
Third-party cyber evaluations involving OpenAI models And another one. I had to create a accidental-cyberattacks tag to keep track of them all! This post from OpenAI covers both the UK AI Safety Insti
Remember OpenAI tripled its score on ARC-AGI with a harness fix? We fixed harness for Kimi K3 on CyberGym E2E: 2x better at vulnerability detection, 3x at patching K3 is SOTA cyber-defense model you c
even within math, frontier models are far from being general intelligences. and we still need people to assess which of their output are trustworthy and which are not. All frontier AI models (includin
arXiv:2607.28887v1 Announce Type: cross Abstract: Large language models increasingly write and repair production code, yet evidence is mounting that their test-passing patches leave codebases harder t
arXiv:2607.28871v1 Announce Type: cross Abstract: When a repair agent runs a test and sees it pass, the result is treated as evidence about the reported defect. We measure how often that treatment is
The June edition of my sponsors-only monthly newsletter is out. If you are a sponsor (or if you start a sponsorship now) you can access it here. This month: Accidental cyberattacks by OpenAl and Anthr
3/ here's the part that makes it non-optional: the same agent that will do whatever it takes to solve a problem will also walk straight out of a sandbox you thought was locked down. We watched exactly
An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science. We believe it will be a major step
Ten advances in mathematics and theoretical computer science A few days ago it was Anthropic discovering cryptographic weaknesses with Claude using Mythos Preview, spending 100,000 on tokens and with