931
We bought the whole GPU, so we're damn well going to use the whole GPU
Benjamin Spector* , Jordan Juravsky* , Stuart Sul* , Dylan Lim , Owen Dugan , Simran Arora , Chris Ré
New
The latest blog posts to take off on Hacker News. 3,238 posts.
932
We reverse-engineered Flash Attention 4
Asynchrony, fast approximate exponents, and 10x more efficient softmax.
933
Code Mode: the better way to use MCP
It turns out we've all been using MCP wrong. Most agents today use MCP by exposing the "tools" directly to the LLM. We tried something different: Convert the MCP tools into a TypeScript API, and then ask an LLM to write code that calls that API. The results are striking.
934
LLM Observability in the Wild – Why OpenTelemetry Should Be the Standard
LLM Observability breaks in wild today as new libraries like OpenInference don't adhere to OpenTelemetry standards leading to silos and suboptimal performance for users who want to tie LLM observability with rest of their observability stack
935
936
Moondream 3 Preview: Frontier-level reasoning at a blazing speed
Think deeper, detect smarter, run just as fast.
937
Modular Manifolds
A geometric framework for co-designing neural net optimizers with manifold constraints.
938
Context is the bottleneck for coding agents now
Ship production-quality code at superhuman speed.
939
Windows ML is generally available
The future of AI is hybrid, utilizing the respective strengths of cloud and client while harnessing every Windows device to achieve more. At Microsoft, we are reimagining what’s possible by bringing powerful AI compute directly to Windows devices,
940
Ollama Web Search
A new web search API is now available in Ollama. Ollama provides a generous free tier of web searches for individuals to use, and higher rate limits are available via Ollama’s cloud.
941
Improved Gemini 2.5 Flash and Flash-Lite
Google is releasing updated Gemini 2.5 Flash and Flash-Lite preview models with improved quality, speed, and efficiency.
942
ChatGPT Pulse
Today we’re releasing a preview of ChatGPT Pulse to Pro users on mobile. Pulse is a new experience where ChatGPT proactively does research to deliver personalized updates based on your chats, feedback, and connected apps like your calendar.
943
Context Engineering for AI Agents: Lessons
This post shares the local optima Manus arrived at through our own "SGD". If you're building your own AI agent, we hope these principles help you converge faster.
944
Qwen3-VL
Qwen offers comprehensive functionality spanning chatbot, image and video understanding, image generation, document processing, web search integration, tool utilization, and artifacts.
945
Android users can now use conversational editing in Google Photos
We’re expanding the ability to edit your images by simply asking in Google Photos to Android devices.
946
OpenAI and Nvidia announce partnership to deploy 10GW of Nvidia systems
OpenAI and NVIDIA announce a strategic partnership to deploy 10 gigawatts of AI datacenters powered by NVIDIA systems, with the first phase launching in 2026.
947
CompileBench: Can AI Compile 22-year-old Code?
We tested 19 LLMs on their ability to handle real-world software engineering tasks like compiling old code and cross-compiling. See how Anthropic, OpenAI, and Google models stack up in our new benchmark – CompileBench.
948
Spectral Labs releases SGS-1: the first generative model for structured CAD
Spectral Labs releases SGS-1: the first generative model for structured CAD.
949
Hidden risk in Notion 3.0 AI agents: Web search tool abuse for data exfiltration
CodeIntegrity research on Notion AI security, Notion 3.0 AI Agents, lethal trifecta risk, indirect prompt injection, web search tool abuse, and data exfiltration.
950
AI tools are making the world look weird
LLMs think WEIRD - Western, Educated, Industrialised, Rich, Democratic. STRAT7 explores what AI cultural bias means for researchers working across global markets.
951
952
Chrome's New AI Features
Google Chrome is getting upgraded with the latest AI to make it safer, smarter and more useful
953
The quality of AI-assisted software depends on unit of work management
The craft of AI-assisted software creation is substantially about correctly managing units of work.
954
Meta Ray-Ban Display
The wait is over. Meta Ray-Ban Display hits shelves in the US today!
955
A postmortem of three recent issues
A technical report on three bugs that intermittently degraded Claude's responses: what happened, why fixes took time, and what we're changing.
956
Bringing fully autonomous rides to Nashville, in partnership with Lyft
Waymo is bringing the magic of our fully autonomous ride-hailing service to riders in Nashville, in partnership with Lyft in 2026. Riders will be able to hail via the Waymo app, and as our service grows, riders will also be able to use the Lyft app to match with a Waymo vehicle.
957
Tau² benchmark: How a prompt rewrite boosted GPT-5-mini by 22%
We expected small models to be fast, but our benchmarks revealed a common reliability trap. Here’s our deep dive on finding and fixing it.
958
Claude can sometimes prove it
We take pride in personally connecting with all interested partners, collaborators and potential clients. Please email us with a brief description of how you would like to be connected with Galois and we will do our best to respond within one business day.
959
Waymo has received our pilot permit allowing for commercial operations at SFO
The latest news, updates, and stories from Waymo.
960
Addendum to GPT-5 system card: GPT-5-Codex
This addendum to the GPT-5 system card shares a new model: GPT-5-Codex, a version of GPT-5 further optimized for agentic coding in Codex. GPT-5-Codex adjusts its thinking effort more dynamically based on task complexity, responding quickly to simple conversational queries or small tasks, while independently working…