Now

79 entries · 20 shown

• 2 September 2026

  the layer underneath the answers

  A paper looks for symbols inside vectors, while the new tools around agents start treating trust, access, and even refusal as product surfaces.

• 1 September 2026

  the tools are learning to leave the screen

  A no-code web preview, an agent control room, and two open-source projects that make sophisticated AI work feel smaller and closer.

• 31 August 2026

  the interface is no longer the job

  A coding editor caught in a corporate feud, a workbench that has escaped the chat box, and the small systems that decide whether agents deserve trust.

• 30 August 2026

  the agent gets a desk, then a badge

  a giant open model, a faster local server, and three small reminders that agent software is becoming workplace software.

• 23 August 2026

  the speedrun, the smug 27b, and the logits on your desk

  Prime Intellect runs 18 frontier models against the same optimizer and tells you who closed the gap. A 27B Qwen quietly does a reverse-engineering job a frontier model was assumed to need. And someone finally explains why your local LLM feels dumb.

• 22 August 2026

  the clone office, the protocol road, and the cheap ripgrep

  MCP publishes its next roadmap with agent identity and progressive tool discovery at the top; Munder Difflin wraps thirteen CLI agents into an always-on clone office that hit HN front page; Anthropic adds Claude protein co-design to its research portfolio; and danluu argues the cost of performance work has collapsed because agents can now run the experiments.

• 21 August 2026

  the venue, the destroyer, the abstraction

  Slack puts coding agents inside group chats; Anna's Archive says AI labs are buying and shredding rare books for training data; a pseudocode-first editor asks what 'writing code' even means in 2026; and DeepSeek ships a vision model that costs a tenth of the alternatives.

• 20 August 2026

  claude designs binders, the model pastes itself into your replies, and the kids ship the feature anyway

  Anthropic puts Claude in the protein-design loop with 35% hit rates; a one-page HN essay says stop pasting AI; Strix ships an open pentest agent; and a counter-essay argues the junior engineer is more useful than ever.

• 19 August 2026

  openai presses pause for two weeks, cursor lands on github's worst day, and mojo is finally open

  OpenAI's two-week frontier training pause is now a real schedule, not a slogan. Cursor ships Origin into a six-hour GitHub outage. Cerebras posts the CS-4 wafer-scale silicon — 30x faster than GPUs. Mojo 1.0 lands on Apache 2.0 with a Microsoft Windows partnership.

• 18 August 2026

  origin, outage, and a chatbot trap

  Cursor launches a GitHub alternative while GitHub is down; Google buys Spirit's call-center tape farm for $10M; OpenAI cuts GPT-5.6 Sol in half; and someone built a fake think tank to manipulate AI.

• 17 August 2026

  watermarks, books, and red agents

  Anthropic's text-watermark rollout draws a furious Gruber; Amazon is cutting bindings off rare books for AI training; a Copilot 'autofix' let Wiz's red agent walk into Snowflake's Jira; and Dario goes on X to defend the company's safety talk.

• 16 August 2026

  anthopic finds its swarm, openai buys back milliseconds, and the on-device model is now a thumb drive

  Anthropic's frontier-red-team multi-agent study lands with concrete numbers; OpenAI's Cerebras-backed Ultrafast tier pushes GPT-5.6 Sol to 14× speed; Anthropic's Q2 revenue hit $11.5B on its way to an IPO; cactus-compute ships a 14MB foundation model that runs on a phone.

• 15 August 2026

  leadership feels closer than code

  Google ships private-AI cryptography, Gemini 3.7 Flash lands on Product Hunt, an engineer gets 232× on a kernel with Codex, and an old builders' essay lands in front of a new crowd.

• 14 August 2026

  gemini flash, glm cyber, cerebras ultrafast, and the new bottleneck

  Google ships Gemini 3.7 Flash at half the cost of 3.6 Flash on the same week z.ai's GLM-5.3 lands with emergent cyber capabilities. Cerebras pairs with OpenAI on GPT-5.6 Sol Ultrafast. Geoffrey Litt's "Understanding is the new bottleneck" talk resurfaces as the defining essay of the moment.

• 13 August 2026

  deepseek ships again, and zed tries to eat the editor

  DeepSeek's V4 Pro lands at the top of HN on a price cut. Zed launches Delta, an agentic editor built around your codebase. Anthropic publishes a conceptual-reasoning index for alignment work.

• 12 August 2026

  counterexamples first

  Gowers notices the machines are winning at counterexamples, not proofs; Anthropic starts signing Claude's output invisibly; a staff engineer describes review as the new bottleneck; Harvey open-sources a legal agent benchmark.

• 11 August 2026

  the bound moves

  Anthropic's Claude nudges a 165-year-old Riemann-zeta result, Cactus ships a 14MB on-device agentic LLM, Anthropic also rolls out AI-content provenance, and Spotify reveals its agentic dev environment.

• 9 August 2026

  latent reasoning lands on a real vllm fork

  DeepSeek-V4 thinking moves into latent space, DeepMind's weather model lands as an open repo, and a 4GB laptop can fine-tune an 8B LLM today.

• 8 August 2026

  deepseek at arc, an open-models gantry, and an accidental attack

  deepseek v4 flash clears arc-agi-2, the doe drops a federal open-models gantry, databricks publishes its coding-agent cost ledger, and openai's own red-team run turned into the week's strangest story.

• 7 August 2026

  the judgment hardens

  AMD buys a startup that etches models into silicon, Cloudflare gives agents their own browser, OpenAI widens the free tier, and a phone decides its owner is a thief.