Normal Science

trycua/cua

Scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation.https://cua.ai Give AI agents computers they can use.Cua provides open-source desktop automation, isolated cloud desktops, local macOS VMs, specialist decision models, and benchmarks for evaluating computer-use agents. Try Cua Fleets now at run.cua.ai Choose your path Cua Fleets: Provision a Linux desktop, run a command, and save a screenshot. CUA-S1: Explore small, spec...

cactus-compute/needle

Automation foundation model for tiny devices: 2-bit, 8-29 MB, tool calls, structured extraction and embeddings on phones, wearables, smart homes, robots, cars and microcontrollers.https://cactuscompute.com A foundation model for mobiles, wearables, robots, smart home, automotive and microcontrollers. The whole model is a single 8-29 MB binary built on our Simple Attention Network, and we trade general chat capacity to beat models 10x its size on mobile tool calls and match 2-3x bigger models on ...

asciimoo/hister

Your own search enginehttps://hister.org/Hister Your own search engine Hister is a private search engine for the pages you visit and the files you keep. It indexes their full contents so you can find information again from the web interface, terminal, or an AI assistant connected through MCP. Try the demo · Download Hister · Read the quickstart · Documentation Quickstart Download the binary for your platform from the latest release, then rename it to hister (hister.exe on Windows). On Linux or m...

cloudflare/quiche

🥧 Savoury implementation of the QUIC transport protocol and HTTP/3https://docs.quic.tech/quiche/ quiche is an implementation of the QUIC transport protocol and HTTP/3 as specified by the IETF. It provides a low level API for processing QUIC packets and handling connection state. The application is responsible for providing I/O (e.g. sockets handling) as well as an event loop with support for timers. For more information on how quiche came about and some insights into its design you can read a p...

Anthropic Looks At Some Of Its Alignment Problems

Anthropic has given us its assessment of four ‘recent cybersecurity incidents’ involving Claude that happened during cybersecurity evaluations, three of which were previously known. The report excludes the incident reported by UK AISI.There will also be a METR investigation of these incidents, which unlike the investigation done at OpenAI will be untimed.Table of ContentsOur Two Problems.First the Good News.We’d Just Like To Ask You a Few Questions.Internal Research Model On The Fence.Opus 4.7.O...

Sovereign: A Unified GPU Inference Substrate (Fractal Memory, Manifold Routing)

All of my Whitepapers. Contribute to CuppaTea1983/Sovereign development by creating an account on GitHub.

NationalSecurityAgency/ghidra

Ghidra is a software reverse engineering (SRE) framework Ghidra Software Reverse Engineering Framework Ghidra is a software reverse engineering (SRE) framework created and maintained by the National Security Agency Research Directorate. This framework includes a suite of full-featured, high-end software analysis tools that enable users to analyze compiled code on a variety of platforms including Windows, macOS, and Linux. Capabilities include disassembly, assembly, decompilation, graphing, and s...

JustVugg/colibri

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦https://justvugg.github.io/colibri Website · Discord · English · 简体中文 · 繁體中文 · Italiano Tiny engine, immense model. Run frontier MoE models — 744B to 2.8T parameters — on consumer and heterogeneous hardware, in pure C with zero engine dependencies, by treating storage, RAM, and VRAM as a single inference hierarchy (AI memory multitiering). Nine families run today: GL...

Making Community Governance Legible: A Semester with Bonfire

By Sohyeon Hwang Authors: Sohyeon Hwang (Princeton University), Andrés Monroy-Hernández (Princeton University) Most of us have spent the last two decades living inside platforms we have little control over. A handful of companies tune algorithms, moderate content, and set policies. Operating at “unfathomable” scales, this comes at a cost: it constrains diverse communities to a single set of rules, leaves little room for local values or context, and gives people almost no visibility — or voice — ...

multimodal-art-projection/YuE

YuE2: frontier music generation with symbolic planning, zero-shot covers, and agentic music editing.https://map-yue2.github.io/ Looking for the original YuE? Its code, documentation, and license are preserved on the YuE-v1 branch. YuE2: Unifying Symbolic and Audio Music Generation at Frontier Quality Compose in symbols. Create in sound. 🎧 Demos · 🗳️ Music Arena · 🤗 YuE2 · 🚀 Quick start · 🤖 Agent skill · 📊 Benchmarks · 🤗 MERT2 · 🤗 SheetSage2 · 🤗 WSB · 📦 Release · YuE2 brings frontier so...

Dynamically Scaled Activation Steering

Activation steering has emerged as a powerful method for guiding the behavior of generative models towards desired outcomes such as toxicity mitigation. However, most existing methods apply interventions uniformly across all inputs, degrading model performance when steering is unnecessary. We introduce Dynamically Scaled Activation Steering (DSAS), a method-agnostic steering framework that decouples when to steer from how to steer. DSAS adaptively modulates the strength of existing steering tran...

Noam Brown – Agent swarms, alignment, & recursive self-improvement

New episode with Noam Brown.We talk about multi-agent, Navier-Stokes, and what the current explosion of maths progress tells us about what happens once you automate AI research.And we also discuss how we will know if the models are actually aligned before we kick off RSI.Watch on YouTube; listen on Apple Podcasts or Spotify.SponsorsJane Street has been interested in AI for a lot longer than you’d think, and not just for trading. In 2011, a full year before AlexNet and over a decade before ChatGP...

OpenAI’s o1 correctly diagnosed 67% of ER patients vs. 50-55% by triage doctors

Researchers say results mark a really ‘profound change in technology that will reshape medicine’

Accelerating Gemma 4: faster inference with multi-token prediction drafters

An overview of how Multi-Token Prediction (MTP) drafters are making Gemma 4 models up to 3x faster at inference.

A couple million lines of Haskell: Production engineering at Mercury

What it takes to run 2 million lines of Haskell in production at a fintech company serving 300,000 businesses.

Using “underdrawings” for accurate text and numbers

A technique for accurate text and numbers in AI-generated images: generate the layout deterministically, then ask the image model to paint on top.

ProgramBench: Can language models rebuild programs from scratch?

Abstract page for arXiv paper 2605.03546: ProgramBench: Can Language Models Rebuild Programs From Scratch?

ZAYA1-8B matches DeepSeek-R1 on math with less than 1B active parameters

Who should care If you work with math, science problems, or complex coding tasks and you're looking for something small enough to run locally or cheaply via API, this is worth serious evaluation. The benchmark numbers at 760M active parameters are not normal and the Markovian RSA boost means performance scales with compute budget rather than hitting a fixed ceiling. If you're building agent workflows that need reliable tool calling or multi-step instruction following, look elsewhere fo

Show HN: Apple's SHARP running in the browser via ONNX runtime web

Hi HN, author here. SHARP is Apple's recent single-image 3D Gaussian splatting model (https://arxiv.org/abs/2512.10685). Their reference code is PyTorch + a pretty heavy pipeline; I wanted to see if it could run in a browser with no server hop, so I exported the predictor to ONNX and ran it via onnxruntime-web with the WebGPU EP.What works: drop in an image, get a .ply you can download or preview live, all on your machine — your image never leaves the tab. The model is large (~2.4 GB sidecar) so first load is slow on a cold cache, but inference itself is a few seconds on a recent Mac.Caveats: SHARP's released weights are research-use only (Apple's model license, not the code's). I host the exported ONNX on R2 so thedemo "just works", but you can also export your own from the upstream Apple repo and upload locally.Happy to talk about it in the comments :)

Text-to-CAD

An open source harness for generating CAD models. Contribute to earthtojake/text-to-cad development by creating an account on GitHub.

GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

Abstract page for arXiv paper 2604.26752: GLM-5V-Turbo: Toward a Native Foundation Model for Multimodal Agents

Learning the Integral of a Diffusion Model

A deep dive on flow maps.

The Road to a Billion-Token Context

Transformers Are Inherently Succinct (2025)

Abstract page for arXiv paper 2510.19315: Transformers are Inherently Succinct

Writing fingerprint analysis of responses reveals Kimi's similarity to Claude

Which LLM models write alike? A heat map built from their words alone.

ChatGPT lost 22 points of web share in a year

Worldwide AI-chatbot web visits: ChatGPT 76% → 54%. Gemini 6% → 28%. Claude 1% → 9%. This is not monthly users and not revenue. See who took the share →

GPT-6-Astra: infinitely pairs of consecutive primes with distance at most 186

Conditional Lean formalization and numerical certificate for prime gaps at most 186. - openai/PrimeGaps186

Anthropic/OpenAI may be spending more than $1000 for every $100 you pay them

Coding with LLMs (Claude Code, OpenAI Codex) is often presented as the ‘killer app’ for Generative AI. But looking at data, it seems the one piece of the puzzle missing is actual cost. …

Show HN: Adam – An embeddable cross-platform AI agent library

An embeddable cross-platform AI agent library written in C. Cloud and local LLMs, tool calling, long-term memory, voice, sessions, research mode, self-evolving loops. The SQLite of agent frameworks: small, portable, just works. - sqliteai/adam

J-space comparisons across open models

Six experiments extending the Verbalizable-Workspace paper on open models: temporal reach, training dynamics, cross-model transplant, capacity scaling, corpus dependence, MoE.

Show HN: I trained a language model that thinks the capital of Japan is Paris

DIMBA II: a 288M-parameter masked-diffusion language model on a bidirectional Mamba-2 backbone. What worked, what failed, and what we

LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes

Abstract page for arXiv paper 2608.31016: LLM Judges Verify Presence, Not Absence: Omission Blindness in AI Clinical Notes and What Recovers It

Grok Build 0.1: Intelligence, Performance and Price Analysis

Analysis of xAI's Grok Build 0.1 0616 and comparison to other AI models across key metrics including quality, price, performance (tokens per second & time to first token), context window & more.

Fermat's Last Theorem: Anthropic has beaten me to it

I guess technically it was revealed to the world by a coffee shop in Islington on Insta, but an hour later it was officially announced by Anthropic: one of their internal models, using the prove2.m…

What I Learned from Reimplementing 40 Multi-Agent LLM Papers

Nobody checked which state IBM's flagship quantum chemistry results compute

This is the data-and-code archive for the preprint "Which state are you converging? A spin audit of sample-based quantum diagonalization benchmarks on iron–sulfur clusters." Sample-based quantum diagonalization (SQD, also called QSCI) selects electronic configurations from quantum-circuit samples and diagonalizes the Hamiltonian in that subspace. Its iron–sulfur demonstrations, on the [2Fe-2S] and [4Fe-4S] clusters, are among the headline evidence for utility-scale quantum chemistry. Thi

The Two MMLU Scores: What a Benchmark Name Does Not Fix

Two accuracy numbers under the same MMLU label, taken apart by variable — split, runner, prompt format, grader, network access — and what a verifier outputs once each one is written into the frame.

Show HN: Bypassing Transformer Softmax via Static Contraction

A JAX XLA-powered PoC that leverages branchless mathematical primitives to bypass the memory and execution bottlenecks of LLM softmax operations - PJHkorea/jax-softmax-bypass

Talk Is Cheap: The Operational Impact of LLM Use

What the data says about the operational impacts of LLM use in the software industry

Fitting Neural Textures and PBR Material Maps with ES (No Backprop)

The repo, released on September 4, 2026 is here: https://github.com/richgel999/neural_texture_es2 ES=Evolution Strategies. Here is a mirror ...

Qwen3.8-27B can silently corrupt business records

Base Qwen3.8-27B can silently rewrite customer identifiers, drop legitimate records, and present stale facts as current. See the evidence, business exposure, and tested mitigations.

Kimi K3 Intelligence, Performance and Price Analysis

Analysis of Kimi's Kimi K3 and comparison to other AI models across key metrics including quality, price, performance (tokens per second & time to first token), context window & more.

GPT-6 Astra makes major gains in the Artificial Analysis Coding Agent Index

Independent analysis of AI models and hosting providers. Understand the AI landscape and choose the best model and API provider for your use-case.

I checked 30 frontier model cards. Here are the benchmarks labs report

An evidence-first daily map of new AI benchmarks, evaluations, and datasets.

GLM-5.2: The Most Powerful Open Model yet and the Brutal Reality of Running It

Z.ai’s GLM-5.2 is the new #1 open-weight model — 753B params, MIT license, a 1M-token context and a real architecture trick (IndexShare). But the weights are 1.51TB. What owners and the benchmarks actually say, and the honest hardware reality of running it at home.

I Cut 80%+ of Context Overhead in My Coding Agent

Most AI coding harnesses dump dozens of tool definitions into the prompt on turn zero. Here is how I designed a dynamic tool activation engine for Pi that keeps the agent lean, cuts 80%+ of baseline context overhead, and auto-cleans on idle.

The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"

Abstract page for arXiv paper 2309.12288: The Reversal Curse: LLMs trained on "A is B" fail to learn "B is A"

Show HN: I benchmarked LLM agents on fixing real-world security vulnerabilities

I built a benchmark with 20 real CVEs across 18 Python projects (Pillow, GitPython, yt-dlp, urllib3, etc). I've run it over 5 LLM agents (3 OpenAI, 2 poolside) and 3 different prompts (full advisory, locate, diagnose) with a total of 300 runs. The agents are tasked to fix security vulnerabilities in a sandboxed environment and they are scored against a hidden security tests from the maintainer's own fix.Best solve rate was 50%. On the other 50%, some fixes are sometimes coherent and pass all regression tests, but vulnerability still present.The main differentiator I found between models is cost: gpt-5.5 at 12× more expensive than gpt-5.4-mini while producing statistically similar results. Within-family performance gaps are small, which points out the difference is likely due to model training data. I also did a power analysis and the task count needed to detect a meaningful within-family edge at ~700.Full write-up: https://giovannigatti.github.io/cve-benchCode: https://github.com/GiovanniGatti/cve-bench

Following the Text Gradient at Scale

RL Throws Away Almost Everything Evaluators Have to Say

What Are Tokens in LLMs?

How LLMs split text into tokens, the BPE algorithm, and why