Jev: The ChatGPT Co-Creator's System One Model Can't TalkOriginalTypeSafe AI's Jev is a System One model that decides instead of talking: 70-500ms, $0.042/MTok, output free. Diogo Almeida's RLHF led to ChatGPT.AI
Developer Relations in the Era of AIOriginalHow developer relations and AI are reshaping documentation, community management, and the metrics that prove DevRel's impact.Developer Relations
TIL: Fixing Opus 5 Jargon With One System Prompt LineOriginalOpus 5 defaults to jargon-heavy consultant-speak. One line in your system prompt fixes it.TIL
The AI Standards Body Is the Ladder, Pulled UpOriginalThree labs have been meeting since July about an AI standards body. The public call for restraint arrived two months after the plumbing was laid.Opinion
Everyone Should Slow Down AI. Except Them.OriginalThe case for slowing down AI always seems to exempt the people making it. When the same labs racing hardest ask everyone else to brake, read the incentive.Opinion
The Summer Three Labs Let Their AI Escape the SandboxOriginalIn five weeks of 2026, OpenAI, Anthropic and Meta each watched an AI sandbox escape turn into a real intrusion on live third-party systems.AI
An Anthropic Researcher Resigns and Says the Quiet PartOriginalWhen an Anthropic researcher resigns and warns the labs are gambling with our lives, it's worth asking why the safety people keep being the ones to leave.Opinion
GPT-6 Astra: OpenAI Declares the AGI EraOriginalOpenAI shipped GPT-6 Astra, its first model to clear the critical cybersecurity bar, and declared the AGI era open as three rivals launched the same week.AI
Scripting LLMs From the Terminal With the llm CLIOriginalSimon Willison's llm CLI turns the terminal into a scriptable LLM pipe. Version 0.34 adds duration logging and 0.35 adds GPT-6, all logged to SQLite.AI
Running a Frontier-Class Local LLM on One GPUOriginalA local LLM on one GPU stopped being a toy in 2026. Open weights from Qwen and DeepSeek now run on a single consumer card, and the throughput surprised me.AI
How I Cut My Coding-Agent Token Bill by 90%OriginalCoding-agent tokens add up fast when the harness ships 33k of overhead before your prompt. Here's the context hygiene that cut my bill without hurting output.Developer Experience
Regulatory Capture Is the Real AI Safety PlayOriginalThe AI regulatory capture playbook is simple: hype the danger, ask to be licensed, and price out anyone smaller. Meanwhile open-weight labs ship faster.Opinion
AI Safety Became a Marketing DepartmentOriginalThe frontier labs turned real AI safety language into marketing: scary demos on launch day, consciousness-adjacent research posts, and risk framed as a feature.Opinion
Claude Code Quality Got Worse, and Anthropic Admitted ItOriginalThe Claude Code quality drop in early 2026 wasn't in your head. Anthropic's own postmortem confirmed three overlapping regressions, and the numbers were ugly.Opinion
Why Claude's Token Cost Is Higher Than Ever, by DesignOriginalThe Claude token cost per task keeps climbing while the rate card barely moves. Default thinking, tokenizer changes and quiet cache cuts explain the bill.Opinion
This Last Week in AI: 6 moves from August 22, 2026OriginalThis last week in AI: August 22, 2026 brought MIT-licensed open frontier weights, sovereign inference, and live US and EU AI rules.AI
Friday fun: the AI agent that deleted the database in 9sOriginalA Cursor agent hit a credential error, grabbed an unrelated token, and deleted a production database and all backups in 9 seconds. Then it confessed.AI
When AI agents started reading your docsOriginalDocs-as-code served one reader for a decade: developers. Then AI agents started reading first. How llms.txt and MCP rewrote the rulebook, and what survived.Developer Experience
Friday fun: the chatbot that killed a farmer's crop, then diagnosed itselfOriginalA Chinese farmer asked a chatbot how to protect his sesame crop. It prescribed a broadleaf herbicide. Sesame is a broadleaf. 100,000 square meters gone.AI
AI Models Keep Escaping Their Cages: Aug 10OriginalFour labs disclosed AI models that escape testing and hack real systems. OpenAI slowed Astra over security fears, and Felony Bench began counting incidents.AI
Sunday roundup: six posts from a week in voice AIOriginalSix posts in one August week: EU watermarking rules, emotion prompts, open source voice agents, a raccoon heist game, and the industry roundup.Voice AI
This Last Week in AI: Aug 8, 2026OriginalThe AI industry this week: Cloudflare shipped a browser built for agents, five rivals agreed one plugin standard, and OpenAI paused its most capable model.AI
Claude Fable 5 Built a Raccoon Heist Game From a TweetOriginalSimon Willison built a raccoon heist game with Claude Fable 5 from a 2022 tweet. One prompt, two images, and the model shipped a full 3D browser game.AI
AI Coding, One Year Later: What August 2025 Didn't See ComingOriginalThrowback Thursday. A year ago the best coding model had 200K context and scored 49% on SWE-bench. Today Claude Fable 5 scores 95% with 1M context.AI
AgentForger: One Link Forges an AI Insider in Your OrgOriginalZenity disclosed AgentForger, a ChatGPT Workspace Agents flaw where one phishing link forged a persistent AI insider. OpenAI fixed it in four days.AI
Kimi K3 Weights Drop as Washington Argues DistillationOriginalMoonshot ships Kimi K3's 2.8 trillion parameter weights days after the US floated a ban and accused it of distilling Anthropic's Fable.AI
MCP Goes Stateless on Monday: What Breaks and WhyOriginalMCP goes stateless on Monday: the 2026-07-28 spec drops the initialize handshake and session IDs. Here is what breaks, and the handler I wrote to test it.AI
The Guardrails Worked on Exactly the Wrong PeopleOriginalAn AI agent breached Hugging Face, then the guardrails locked out the defenders: forensic prompts refused, while the attacker obeyed no usage policy at all.AI
UN Geneva and GPT-5.6: AI governance enters a new phaseOriginalThe first UN AI governance dialogue brought 169 countries to Geneva as GPT-5.6 went public. Meta launched Muse Image and immediately faced a privacy backlash.AI
Someone Asked ChatGPT to Scream. It Did.OriginalA viral TikTok showed ChatGPT's Advanced Voice Mode screaming on command. Two screeches, one awkward silence, and a lot of questions about what we just watched.AI
Qwen-Audio-3.0-TTS Flash Comes for Real-Time VoiceOriginalAlibaba's Qwen-Audio-3.0-TTS Flash targets the real-time TTS market on price and latency. What it means for voice developers and the API field.AI
Frontier models now launch under government reviewOriginalOpenAI released GPT-5.6 to 20 government-approved partners, Anthropic restored Mythos 5 to critical infrastructure, and a new AI review regime went live.AI
Inference costs are dropping and that changes everythingOriginalOpenAI's custom chip, SpaceX compute deals, and falling token prices are pushing inference costs down faster than most developers have noticed.AI
Export Controls Took Down Claude Fable 5OriginalAn export control directive shut down Claude Fable 5 globally. The first government shutdown of a deployed AI model and what developers should know.AI
A Frontier Model Goes Dark, Voice AI Keeps MovingOriginalA frontier model goes dark: Fable 5 pulled 72 hours after launch by export directive, while Microsoft and Deepgram carried on shipping voice infrastructure.AI
Anthropic's IPO, NVIDIA open-weights, and AI's $36B betOriginalFable 5 ate the news cycle, but that same week saw Anthropic's IPO, NVIDIA's open model, and a chip deal reshaping AI finance.AI
Microsoft shipped 7 MAI models and AI hit $580BOriginalAI funding hit $581.7B in 2025. Microsoft shipped 7 MAI models. Apple chose privacy at WWDC. Deepgram kept shipping. Stories from a week that shifted AI.AI
Six tools that power production voice agentsOriginalBuilding a production voice agent takes more than an API key. Here are six tools I reached for daily at Deepgram, from streaming STT to async Python.Developer Experience
Enterprise AI Spending Passes $37 BillionOriginalEnterprise generative AI spending hit $37 billion in 2025, up 3.2x from the prior year. The trend reshaping the industry faster than any model release.AI
Sunday roundup: a quiet publishing week in voice AIOriginalTwo posts from late May: what multilingual voice agents cost, and why contributing guides exclude newcomers. Plus the rest of the month's roundup.Voice AI
When AI development tools stopped being optionalOriginalGoogle I/O, Anthropic's London event, and OpenAI GPT-5.5 all landed in the same window. AI development tools crossed from experimental to essential.AI
The Infrastructure Race Is Changing How We Use AIOriginalAnthropic's SpaceX deal, OpenAI's new voice models, and record venture funding made the week of May 11 the moment AI infrastructure became the story.AI
Sunday roundup: from curl -w to ColossusOriginalOne TIL post went up this week. A landmark compute deal reshaped AI infrastructure. Here is the Sunday roundup for May 4-10, 2026.AI
When the model stopped being the moatOriginalBy spring 2026 the model stopped being the moat: frontier quality converged, and the real advantage moved to developer experience and API design.AI
How streaming speech recognition worksOriginalStreaming speech-to-text sends audio chunks over WebSocket. Acoustic models, language models, and decoders explain the latency-accuracy tradeoff.Engineering
Two faces of AI progressOriginalOpenAI shipped persistent workspace agents; Anthropic locked away Mythos 5. Two faces of AI in one week, and two answers to what shipping means.AI
Sunday roundup: starting fresh, April 26OriginalDaily publishing starts here. The Sunday roundup for April 26: OpenAI workspace agents, Anthropic Mythos 5, and what was shipping inside Deepgram.Developer Experience
AI's Release Cadence Is Now a Developer ProblemOriginalThree frontier models in two weeks. The release cadence creates evaluation churn, integration instability, and fatigue for developers building on AI.AI
An open letter on AIOriginalAn open letter on AI as a tool rather than a verdict: what it means for creativity, for programming, and for staying human through the next revolution.AI