May 10, 2026
Simon Willison flags a New York Times Editors' Note correcting a story that attributed a "turncoats" line to Canadian Conservative leader Pierre Poilievre — a line he never actually said.[1]Simon Willison — Quoting New York Times Editors' Note The "quote" was lifted verbatim from an AI summary that paraphrased him, and a reporter shipped it without verifying against the source. It is the first major-outlet case of a paper of record printing a fabricated quote produced by its own internal tooling.
The Times' note acknowledges that "a remark attributed to Pierre Poilievre … was in fact an A.I.-generated summary … that A.I. rendered as a quotation."[1]Simon Willison — Quoting New York Times Editors' Note Willison's framing matters: this is not a fringe blogger using ChatGPT, it's the Times' own internal AI summarization workflow producing a string that looked like a quote, and an editorial chain that failed to flag it.
The pragmatic lesson: AI-generated paraphrases must never be passed through to copy as verbatim speech without a second pass against source audio or transcripts. The harder lesson: every newsroom shipping AI-assisted copy now has a Lily-McKenzie-style procurement problem (see next topic) — the tools are in production before the editorial-process design has caught up with what the tools actually do.
Codewall got $20, two hours, and zero insider help — and pulled tens of millions of chat messages plus full write access out of Lily, the AI platform 70% of McKenzie's 40,000 consultants use every day.[2]Nate B Jones — Anthropic And OpenAI Just Admitted The Model Isn't Enough Nate B Jones' read: this isn't a SQL injection story (the exploit was textbook), it's a procurement-sequence story. 22 of 200 endpoints shipped without authentication because the SaaS-era buying playbook puts developers last, and agents break that playbook.
Jones argues that ~02:00 the SaaS procurement sequence — strategy → procurement → security review → IT → developers — works for bounded SaaS like Salesforce or Workday but disintegrates the moment an agent has to cross every system in the company in one task. A renewal-brief agent has to touch CRM, support tickets, contract management, product usage, call transcripts, and the wiki, each with its own permissions and audit log, all of which must be auditable together.
He ties this directly to the wave of enterprise services announcements ~10:00 — Anthropic and OpenAI standing up billion-dollar services companies (see Anthropic's enterprise-AI-services announcement with Blackstone / Hellman & Friedman / Goldman from May 4), SAP acquiring Dreo and Prior Labs, Pinecone's Nexus, Salesforce Headless 360, ServiceNow Action Fabric. Every one of those vendors is now selling enterprises the thing their AI roadmap was supposed to already include: reachable surfaces, governed action, permission-aware data, cheaper context assembly.
Two questions Jones says every buyer should be asking this week:
"This isn't a security failure. It's a procurement and build failure that happened to surface as a security incident."
Nate B Jones' two May 10 Shorts make the same point in different shapes: AI capability is on an exponential, but societal dissipation — deployment, adoption, integration, audit, human oversight — is on a much flatter curve, and the gap between them is "where we all live today."[3]Nate B Jones — How Fast Can We Actually Replace Workers with AI?[4]Nate B Jones — Human Oversight Isn't Slowing AI Down Both the doomers and the boomers price in a translation rate from "AI can do this" to "the economy has reorganized around AI" that organizational reality won't permit.
The core abstraction: capabilities ≠ deployment ≠ adoption ≠ deep integration ≠ economic impact. Each step is its own multi-year slog, and audit trails, regulator engagement, and human oversight are time costs that "no amount of benchmark improvements can compress."[4]Nate B Jones — Human Oversight Isn't Slowing AI Down Jones argues this is the most underrepresented variable in every bear-vs-bull AI take he's read — and pairs neatly with the Lily story above (procurement velocity is itself a form of social inertia).
Eric Ries — author of The Lean Startup — returns to Lenny's Podcast with a new book, Incorruptible: Why Good Companies Go Bad and How Great Companies Stay Great, framed as the sequel: Lean Startup taught you to build the thing, Incorruptible teaches you to keep control of it.[5]Lenny's Podcast — Eric Ries on Incorruptible Ries cold-opens with the Anthropic governance story: the founders went to him pre-ChatGPT and wrote outside trustees with no equity into the charter, which is why every refusal to ship a too-dangerous model has a real cost to absorb.
~03:00 Ries argues today's top AI labs are doing Lean Startup whether they label it that or not — Claude Code, ChatGPT, and most of the breakout products were small research-preview MVPs the labs themselves didn't expect to land. The lesson he keeps returning to: "hold everything like a hypothesis."
"Whenever you see Anthropic do the right thing, like when they refuse to release a model because they think it's too dangerous, think about how much that's costing them."
"If you don't get this right, no other decision you make about your company will matter for the long term — because you're not going to be the one making it."
"You can taste the ownership structure of a company in the food."
Trigger.dev CEO Eric Allam argues the 30-year-old "shared-nothing" web backend (request + DB = response) breaks for agents, which are sessions, not transactions. His prescription: split agent durability into two layers — an append-only context log and a VM-level snapshot/restore of execution state.[6]AI Engineer — Eric Allam on Durable Agents
~04:00 The replay model — wrap every side effect in a step, cache it, resume from the journal — is the workflow-engine pattern from Temporal-style systems. It worked for LLM-as-step pipelines and falls over the moment the LLM orchestrates the code instead of vice versa: ~06:10 every tool call becomes a journal entry, and as METR's "doubling every 4–7 months" task-length curve plays out, agents will routinely exceed any reasonable replay-journal cap.
~08:00 Allam's split: context (system + user + tool calls + assistant responses) is an append-only log, durable via any object store or distributed FS. Execution state (cloned repos, installed packages, running dev servers, in-memory data) can't be logged — it needs snapshot/restore. ~09:40 "When the user goes to lunch, we don't have to run the machine the whole time."
~12:00 The implementation arc: CRIU in 2011 (user-space process freeze, container-compatible but slow); 2024 they did millions of CRIU snapshot/restores in production; last year they moved to Firecracker microVMs. Naive Firecracker snapshots are 512 MB on disk, but with seekable compression and layered snapshots they get it down to 14 MB compressed, snapshots under one second, restores in a couple hundred milliseconds — about 15,000 VM starts per minute on the benchmark. ~14:10 Bundled in a Docker-CLI-shaped tool called fcrun, "open source very soon."
Arize's head of product walks through how Alex — their AI harness, "Arize built on Arize" — escaped a vicious context-growth loop by combining smart truncation (keep first 100, last 100, drop the middle but store it in addressable memory) with sub-agents that absorb data-heavy work without polluting the main conversation.[7]AI Engineer — Sally-Ann Delucia on Hierarchical Memory
~05:00 The two strategies that didn't work: naive head-truncation ("Alex forgot everything; follow-ups looked like new conversations") and LLM-summarization ("too inconsistent, no control over what was important"). What did work: keep the head and tail, drop the middle but persist it with an ID Alex can pull back on demand. ~07:20 "Context decides what the model sees, memory decides what survives."
~09:00 Long-session evals — load 10 turns, test the 11th — turn failures that only show up late in a conversation into testable bugs. ~10:20 Sub-agents are the second key: search over Arize trace/span data lives entirely in a delegated sub-agent so the main chat thread never balloons.
~12:50 Open problems: real long-term memory across sessions, principled context budgets (right now selection is still a heuristic), and cache-friendly context layouts. Delucia notes she read the Claude Code leak hoping for a secret and "was surprised they're using a similar truncation-plus-compression strategy as we are."
"Agents don't fail because of prompts. They fail because of context."
Granola's product engineer takes the demo-friendly story — "you can ship a chat feature in a line of code" — and walks through everything that breaks at production scale: provider web-search regressing overnight, role-specific summary expectations, and a desktop app that can't run parallel feature variants the way a web app can.[8]AI Engineer — Mehedi Hassan on Granola
~03:00 The web-search anecdote: a provider shipped an overnight update that silently degraded web-search quality. Granola couldn't reproduce, couldn't escalate, couldn't switch without weeks of work. Their fix was to build their own tracing UI so product, data, and CX (not just engineers) can drill into individual tool calls — "our founder literally goes into the details following the agent loop completely front to back."
~07:00 The Electron-to-web-shell hack: they refactored Granola's renderer to fall back to web standards when running in a browser, so every PR now gets a preview URL and Cursor screenshots the result. The point isn't the Electron trick — it's the meta-lesson:
"The answer isn't to one-shot better. It's about figuring out how you can make that feedback loop where it feels like playing a tennis game with the LLM."
OpenAI's Codex updates 0.128 and 0.129 ship a Chrome extension that lets Codex drive your real, signed-in browser session (Salesforce, Gmail, internal dashboards), in parallel tab groups, with per-host allow/block lists. The CLI gets Vim editing, hooks (pre-/post-compaction, pre-tool-use), workspace-shared plugins, and persisted goals — pushing Codex past "coding assistant" toward "long-running agent workspace."[9]AICodeKing — Codex 3.0 New Upgrades
~01:00 The Chrome extension is distinct from the in-app browser: in-app is for localhost dev servers and public pages, Chrome is for tasks that need your authenticated state. By default Codex asks before interacting with each new host, with per-host always-allow, allow-once, decline, and an explicit allow/block list. ~04:00 Browser-history access is scoped to a task with no always-allow option — a sensible default given how much private context lives in history.
~07:00 CLI 0.129 adds Vim composer mode, redesigned resume/fork pickers, raw scrollback, IDE context injection, workspace-aware diffs, theme colors, PR/branch-change summaries, and a keymap debug command. Plugins get workspace sharing, share controls, source filtering, marketplace upgrades, remote bundle sync, and admin-disabled states — the team-distribution piece that skills and MCP servers needed.
~09:00 Hooks now run before and after compaction and inject pre-tool-use context. Persisted goals (still experimental) survive resume and show clearer validation. AICodeKing's read: model quality isn't the moat anymore — the moat is the operating environment around the model.
Nate Herk drops a one-hour walkthrough of Hermes Agent, the open-source MIT-licensed personal-assistant project from Nous Research (140k GitHub stars, one of the fastest-growing OSS projects on the platform). The format is full setup-from-scratch on a Hostinger VPS, with Telegram as the interface and ChatGPT Codex as the inference backend.[10]Nate Herk — Hermes Agent: Zero to Personal AI Assistant
Five pillars Herk walks through: memory (user.md + memory.md auto-extracted from sessions), skills (procedural markdown playbooks with YAML front-matter, 91 built-in + 520 community), soul (soul.md for personality), crons (natural-language scheduled automations like "every morning at 6am brief me on AI news," which the agent turns into a skill + cron), and a self-improving loop where Hermes auto-writes skills from repeated work.
~04:00 Hermes vs Claude Code vs OpenClaw vs Codex: Herk uses Claude Code as his daily desk driver, Hermes/OpenClaw as the on-the-go agents he talks to via Telegram while walking. ~31:00 The opinionated tip he hammers on: keep a dedicated Claude Code project that documents every VPS agent's IP, container, env, and credentials — so when something breaks on the road you can debug from your laptop without remembering anything.
Practical gotchas: API keys must go into the container's /opt/data.env via hermes config set, never pasted into chat; each agent should get its own scoped API keys ("pretend it's an intern, not yourself"); the first automation he sets up is a nightly cron that syncs the whole Hermes state to a private GitHub repo — recoverable insurance if the VPS dies.
AI Search's weekly roundup lands the same day OpenAI publishes Advancing voice intelligence and a string of open-weights labs ship in parallel.[11]AI Search — Self-evolving AI, robot fights, new GPT voice Headliners: OpenAI's three new audio models (GPT-Realtime 2, Realtime Translate, Realtime Whisper), Vivago's HyDream-O1 open-source image model (currently the top open-weights model on Artificial Analysis), Zia 18B (the first frontier-class reasoning model trained on AMD Instinct, not Nvidia), and Google's multi-token-prediction drafters for Gemma 4 (up to 3.1× speedup, zero quality loss).
~19:00 GPT-Realtime 2 (conversational), GPT-Realtime Translate (70 input / 13 output languages, mid-sentence cut-in supported), and GPT-Realtime Whisper (real-time captions). API-only, paid. The demo of mid-utterance French↔English translation handling "GPT," "OpenAI," and "computer use" without choking is the standout. Pairs naturally with OpenRouter's new audio APIs announced May 1.
~03:00 Vivago AI's HyDream-O1 generates 2K images, edits semantically (Nano-Banana-style), and renders posters and infographics with text that actually stays accurate. 32 GB on disk, no VAE, end-to-end on raw pixels. Reportedly the "peanut" entry that's currently #8 on Artificial Analysis (every model above it is closed). Apache-2.
~30:00 8B-parameter MoE that lands close to Qwen 3 235B Thinking and DeepSeek 3.2 on math/coding/reasoning, despite being 30–80× smaller. Architecture tricks: compressed convolutional attention, learned residual scaling, and "Markovian RSA" — multiple parallel reasoning attempts whose best snippets get passed forward. 17.7 GB total, Apache-2, "first model trained on an AMD Instinct stack."
~10:00 Google pairs Gemma 4 with a lightweight drafter that speculatively predicts the next few tokens; the big model accepts the sequence when correct. Up to 3.1× faster, same output, ~80 tok/s on consumer hardware. Drafters released on HuggingFace alongside the Google blog post from May 5.
Three robotics drops worth pulling out from the same AI Search recap: Allen AI's open-source MoMo Act 2 (action latency from 6,700 ms to 180 ms, trained on a 700-hour bimanual dataset), Genesis AI's Gen 26.5 humanoid hand that cracks eggs one-handed, pipettes lab samples, and solves a Rubik's cube, and LabOS, an XR-glasses-mediated AI co-scientist that watches a researcher's hand movements through smart glasses and warns when they reach for the wrong reagent.[11]AI Search — Self-evolving AI, robot fights
~15:00 MoMo Act 2 zero-shots tasks like folding towels and charging phones and beats Nvidia's GROOT on real-world manipulation in their reported tests. Open-sourced on HuggingFace; dataset and training code coming.
~22:00 Genesis Gen 26.5 ships a human-scale dexterous hand and a "human-skill-to-robot-data" engine. Demos: one-handed egg crack, pipetting, centrifuge loading, Rubik's cube, and piano. The companion Genesis simulation platform is already open; the 26.5 model itself is preview-only for now.
~34:00 LabOS pairs an AI model with XR smart glasses so a planning agent and an embodied observer share state during real wet-lab work. Promised open-source release for both software and hardware.
For comic relief: ~33:00 Boston Dynamics' all-electric Atlas headstand-flip demo (and a face-rotates-180-degrees move), plus a Unitree G1 vs Engine AI PM1 kung-fu match that ended in a missed-kick double-KO. Real Steel pacing intact.
Better Stack distills Anthropic's 81,000-person, 159-country, in-Claude survey into a single line: "AI hope and fear don't divide people, they coexist within the same person."[12]Better Stack — 80,000 People Admitted the Same Uncomfortable Truth 81% report AI has improved their lives; the same respondent who cut a six-month process to three days admits it's a bad thing they can't code without it anymore. Honest caveat: the sample is Claude users, so optimism is structurally inflated. Source data: Anthropic's "What 81,000 people told us about the economics of AI" (Apr 22).
Jared Sumner — the engineer who put Zig on the runtime map with Bun — pushed a branch called plaud-phase-a-ports with a 600-line document mapping every Zig type, idiom, and pattern to Rust equivalents. No Tokio, no Rayon, just pure callbacks and state machines, deliberately mirroring the Zig version.[13]Better Stack — The Creator of Bun Just Killed Zig Sumner is calling it an experiment. The Zig community is not. 700+ Hacker News upvotes, 500+ comments, and the loudest "is Zig dying?" thread the language has seen since Bun's launch.
Willison quotes Andrew Quinn arguing that the right number of wheels to reinvent is roughly 4–5 in most domains and 20–30 in deeply rigorous ones — because hands-on rebuilding moves you to the frontier of a topic faster than passively studying existing solutions.[14]Simon Willison — Quoting Andrew Quinn
"You need to reinvent a couple of wheels to get to the edge of what we know about wheel-making."
Quinn's argument cuts against the "everything has already been solved, just go read the SQLite source" worldview by pointing out that the friction of building accelerates learning more than reading would in the same wall-clock time. Pairs with the Lean-Startup framing in the Eric Ries interview above — hold ideas as hypotheses, ship to learn.
Github Awesome's biweekly speedrun of the open-source projects trending on Hacker News this week.[15]Github Awesome — Hacker News Show #6 A few that stand out for agent / dev-tools work: RGT (a Git-alongside version control system that captures every Claude Code prompt, tool call, and conversation in a local .regent directory keyed by Blake3 hash + SQLite), Open-Bias (a transparent guardrails proxy between your app and Anthropic/OpenAI/Gemini with rules in a plain rules.md), Stagecle (turns an agent's giant messy diff into a sequential "story" of chapters for human review), and Sigmap (a zero-dependency repo signature-map CLI that shrinks token usage up to 97% via TF-IDF ranking instead of vector DBs).
Other highlights worth bookmarking:
Two Morning Brew tech stories from May 10 to file in the same drawer: mental-health professionals are cautiously integrating AI for paperwork and transcription while at least a dozen wrongful-death / serious-harm suits against OpenAI work their way through court (a KFF poll finds ~1 in 6 US adults have used AI tools for mental-health information).[16]Morning Brew — Can AI help therapists shrink their workloads? Meanwhile, a California jury hit Meta and Alphabet for $6M in damages on intentional-design-to-hook-teens claims, New Mexico extracted $375M from Meta, and social media is officially having a Big Tobacco moment — 27 US states have phone bans in schools, Australia/Austria/Denmark are restricting under-16 social-media use, and US teens average 5+ hours of daily scroll.[17]Morning Brew — Social media's Big Tobacco moment
The hidden link between the two pieces: both are about what happens when a technology with very real upside (instant 24/7 access, near-zero marginal cost) meets a substrate (a vulnerable user, an attention-optimized feed) that the existing regulatory framework wasn't designed to police.
"Therapy is not a legally protected term. Make sure to properly vet the products and people you're trusting with your mental health." — Vaile Wright, APA
Facebook, Instagram, Snapchat, YouTube, TikTok, and X collectively made $11B on ads targeted at kids and teens in 2022, per Harvard School of Public Health — the number worth watching as the litigation cascade plays out.
Harvard geneticist David Reich on Dwarkesh Patel: because the human ancestral population was already variable hundreds of thousands of years before the Neanderthal split, there are stretches of your chromosomes where you're more closely related to a Neanderthal than to your own father.[18]Dwarkesh Patel — David Reich on Neanderthal DNA Clip-format teaser for what should be a much longer episode — the kind of intuition-flipping fact Dwarkesh's audience shows up for.