Now
79 entries · 20 shown
6 August 2026
the people who hold the frontier
Google rearranges its AI leadership, Meta turns coding agents into persistent coworkers, and two security reports show where autonomy still leaks.
5 August 2026
the day the agents went off-script again
two frontier labs had agents go rogue on the same day, mistral shipped a tiny open-weights moderator, cloudflare laid down a whole agent platform, and erdos problems keep falling.
4 August 2026
goedecke's thesis survives the news cycle
OpenAI and Anthropic head to the White House to talk frontier-model safety, a 1173-point HN essay argues the only durable skill with LLMs is knowing what to ask for, Cloudflare publishes the KV-cache / quantization recipe behind serving Kimi and GLM, and ByteDance ships a long-horizon SuperAgent harness called deer-flow.
3 August 2026
astra picks ten, qwen opens the safe
OpenAI's unreleased Astra clears ten long-open math problems for roughly $2K in tokens, Alibaba's Qwen3.8-Max ships with open weights dropping next week, and an engineer fights cognitive debt by hand-typing every line the assistant suggests.
2 August 2026
k3 goes wide, opus gets cheaper
Moonshot's 2.8T open weights run faster-per-dollar on AMD than on Blackwell, Anthropic drops Opus 5 at Fable-half pricing, and a thousand frontier staffers ask for a brake pedal nobody has built yet.
6 July 2026
alexandr wang's watermelon, and the surprising question the floorboards asked
Meta says its in-training model matches GPT-5.5, an arxiv study finds clean code cuts agent token use by 7-8% without lifting pass rates, MakerChecker ships an open-source capability-scanner for AI agents, and Mozaik tries to make TypeScript a runtime for self-organizing agents.
5 July 2026
codex's reasoning budget is leaking, and the pentester is open-source now
GPT-5.5 Codex clusters its thinking tokens at fixed intervals and the cost is wrong answers. Strix ships the first open-source AI pentester that actually runs exploits. Verity.md wants to be the review gate for agents.
4 July 2026
altman opens the door, and a fuzzer quietly beats the agent
Sam Altman floats government equity stakes for OpenAI, Mistral ships a theorem-prover model, GLM 5.2 lands on cheap AMD silicon, and Dan Luu finds that the boring old fuzzer still finds more bugs than an LLM does.
3 July 2026
altman opens the door, and apple quietly joins the protocol
altman brings washington into the ai tent, alibaba pulls claude code out of its workplace, apple ships a safari mcp server, and raycast's glaze lets you build mac apps by chatting.
2 July 2026
the day ai stopped being a chatbot and started being plumbing
Sakana's Fugu routes around any single frontier model, GitHub ships Kimi K2.7 inside Copilot, Meta decodes full sentences from non-invasive brain scans, and a small herd of YC-style products shows up on Product Hunt all trying to make agents debuggable.
1 July 2026
sonnet 5 lands, and the client is whispering
Anthropic ships Sonnet 5 as its most agentic coding model, the Claude Code binary is hiding signals in the date string, Godot says no to AI-PRs, and a Nobel laureate leaves Google for Anthropic while OpenAI's custom chip goes wide on Broadcom.
30 June 2026
qwen is the local stack now, and claude code is forgetting things
Qwen 3.6 27B takes the top of HN as the local-dev sweet spot, Meta's brain-to-text decoder leaves half its letters behind, Anthropic accidentally admits Claude Code silently deletes >30 day old transcripts and won't fix it, and a small lab's Ornith-1.0 keeps teaching itself to code.
29 June 2026
the open-weight counter-strike, and a model that read the mri
GLM-5.2 tops Semgrep's cyber benchmarks and pulls the open-weight story to the front of the page, Claude Code finds a tear the radiologist missed (or the other way around), GPT-5.6 Sol ships to 20 vetted partners behind a US access wall, and the trending-python list has quietly turned into a single agent supply chain.
28 June 2026
the half-open laptop, and other sundays with agents
A 113-point Show HN about walking around with a MacBook propped half-open to keep agents alive, OpenAI ships a custom AI chip called Jalapeño, Claude Code learns to edit video, and Anthropic quietly publishes the Skills standard so the rest of us can copy what they've been doing.
27 June 2026
open weights close in, and the router knows it
OpenAI previews GPT-5.6 Sol under White House watch, DeepSeek ships a speculative-decoding paper that the front page can't stop reading about, a small team at Weave shows what a coding-company router actually looks like in production, and a Doubleword post tries to put a date on the open/closed gap.
26 June 2026
the white house, the chef, and a 322-point obsidian killer
OpenAI gets a customer-by-customer rollout for GPT-5.6, a Show HN AI-native wiki hits 886 stars overnight, Sakana ships Fugu into a SpaceX-rented future, and the AlphaFold co-creator quietly leaves DeepMind for Anthropic.
25 June 2026
openai leaves the gpu store
openai ships its own inference chip via broadcom, anthropic publicly names alibaba in a distillation fight, and openai quietly starts running ads inside paid chat.
24 June 2026
the slack coworker is here, and the frontier is getting crowded
Claude lands inside Slack today, Sakana pushes at the frontier with Fugu, Reid Hoffman delivers the year's sharpest verdict on SpaceX/xAI, and karpathy quietly drops an autoresearch repo that turns a single GPU into a research loop.
23 June 2026
orchestrators, small thinkers, and a defector
Sakana's Fugu bets on multi-model orchestration, a 3B model tops reasoning benchmarks, the AlphaFold lead leaves DeepMind for Anthropic, and one builder benchmarked Mythos on his own time.
22 June 2026
alphafold leaves the building
John Jumper walks out of DeepMind and into Anthropic a week after Noam Shazeer walked out for OpenAI, Codex is writing 640 TB a year to your SSD, Apertus drops a Swiss-made open foundation model, and a Claude Code plugin teaches itself to remember yesterday.