Back issues
Andrej Karpathy got tired of YouTube's recommendations on a Friday night and on Saturday morning handed Opus 5.5 one task: build a YouTube clone from whatever videos Hacker News commenters have recommended. The model pulled the HN comment corpus, ran Sonnet over it in a workflow to drop spam, narrowed to about 40,000 recommending comments and 6,454 videos across 150 topics from 2007 to 2026, wrapped it in a YouTube-style UI, bought a domain, and deployed and hardened a VPS. The site at hackertube.org has no accounts, keeps state in the browser, and exports its data so a reader can hand it to their own agent and make their own version. His closing line is the receipt: this takes about two hours now. A first-person account of directing software rather than writing it, from a name a room will stop for, with the counts in the body.
“So this morning (Saturday) I took a hold of my Opus 5.5 and gave it the task: Let's write our own YouTube clone based on whatever videos have been recommended on Hacker News.”
Andrej Karpathy got tired of YouTube's recommendations on a Friday night and on Saturday morning handed Opus 5.5 one task: build a YouTube clone from whatever videos Hacker News commenters have recommended. The model pulled the HN comment corpus, ran Sonnet over it in a workflow to drop spam, narrowed to about 40,000 recommending comments and 6,454 videos across 150 topics from 2007 to 2026, wrapped it in a YouTube-style UI, bought a domain, and deployed and hardened a VPS. The site at hackertube.org has no accounts, keeps state in the browser, and exports its data so a reader can hand it to their own agent and make their own version. His closing line is the receipt: this takes about two hours now. A first-person account of directing software rather than writing it, from a name a room will stop for, with the counts in the body.
“So this morning (Saturday) I took a hold of my Opus 5.5 and gave it the task: Let's write our own YouTube clone based on whatever videos have been recommended on Hacker News.”
Nielsen Norman Group explains why it renamed its field experience design: the user of a digital product may now be an agent acting for a person, shopping, filing a complaint or cancelling a subscription outside every channel the company designed. The article traces the term from Don Norman in the 1990s through journey maps and service blueprints, then defines experience design as covering systems people use directly and systems their agents use on their behalf, with the human's goals still the measure of success. It says AI is changing research methods too but gives no examples or data. A naming argument from a well known research firm, adjacent to the thesis rather than on it.
“is the discipline concerned with designing and evaluating how people accomplish goals through digital systems, whether they interact with those systems directly or through agents acting on their behalf.”
A single quotation Simon Willison pulled from an Amazon engineering leadership update by Terk Terkowitz. The line says that if your work does not benefit from current AI tooling, that is a finding, not a failing, and an engineer who can explain where AI does not yet help is showing the judgement being asked for. There is nothing else in the body beyond the blog's navigation. It keeps because it is a primary statement of how one large employer frames the expectation, in one sentence a manager could repeat.
The Pragmatic Engineer's weekly Pulse opens with a hiring manager at Anthropic who stopped asking coding questions in engineering interviews because they gave no signal. The free opening also lists incident.io cutting AI spend per developer from 4,500 dollars a month to 1,600, Amazon formally adding AI usage to engineer and manager expectations, and Microsoft's exit offers. The body ends at the paywall after the first section's lead, so the detail behind each item is not in hand. Kept at the directing desk for the Amazon and hiring items, which are about what employers now ask of people, and marked as paywalled past its opening.
“A hiring manager at Anthropic who has led Claude Cowork and Claude Desktop, revealed that they have stopped asking coding questions to candidates for software engineering jobs.”
Joseph Cox at 404 Media reports that Proton CEO Andy Yen cancelled near company-wide AI re-interviews, which would have made employees justify their AI use on pain of dismissal, one day after 404 Media's investigation into the push went public. The sourcing is six current employees plus audio of a two-hour all-hands; the earlier investigation drew on eight current and one former employee, Slack logs and an AI manifesto by CTO Bart Butler that told engineers their old job is gone. The Slack exchange quoted has Yen calling a request to pentest AI-written code before shipping loser thinking, and telling staff to stop handcrafting code and use AI more. Employees are split on whether the cancellation changes anything, since the same measures now fold into performance reviews and the direction is unchanged. For a leader deciding how hard to push adoption, this is the clearest recent account of what mandate-by-fear does to a security-focused engineering culture.
“Another employee said, after Proton canceled the re-interviews, “They said in the next breath they want to continue pushing AI and that the things that were to be measured in the re-interviews would become a component of the performance reviews.””
Cisco Talos describes a self-deleting Windows infostealer called VOLTA-AUTODOXER that adds an LLM stage after the theft. The stage bundles what it took about the victim, sends it to a hosted model with a prompt asking for an analyst-style assessment of what the victim is worth, and ships the answer back to the operator. Talos notes the authors expected model refusals and built a fallback chain across three vendors' models, with the stage skippable if every model refuses, and reads the whole thing as a builder product sold to customers of any skill level. Their conclusion is the lesson: the AI does not make the malware faster, it removes the step that needed sustained human attention, the same effort-displacement they saw in earlier AI-touched malware. Primary source about its own analysis, with countermeasures for defenders.
“If a model refuses to perform the target assessment, VOLTA fails over to a three-model chain, implying a direct engineering effort and expectation of hitting model guardrails.”
Fortune reports Bureau of Labor Statistics data putting labor's share of U.S. output at 52.8% in the second quarter of 2026, the lowest since the series began in 1947, while Fortune 500 profits hit a record $2.1 trillion and the economy added 584,000 jobs in 2025 against 2 million the year before. Texas A&M economist Raymond Robertson blames automation; Goldman Sachs analysts estimate AI could displace 25% of work hours, though the piece concedes AI displacement has not shown up meaningfully in any economic data. The productivity side is where it earns a desk: McKinsey data cited here has about 80% of employees saying AI makes them faster at specific tasks but only 37% of companies seeing that reach operating profit, Morgan Stanley calls the productivity surge an open question, and Oxford Economics suggests firms are dressing up overhiring layoffs as AI. A second expert pins the shrinking workforce on the immigration crackdown instead. A republished January story, broad and relayed, but a clean statement of the gap between felt speed and measured payoff.
“Robertson attributes weakening labor share averages to the rise in automation, which he noted is displacing workers, with productivity—a metric essentially measuring worker output—continuing to rise.”
A maker built MonsterHouse for Halloween: a projection that gives a garage door a mouth and a grumpy pirate voice, listens to trick-or-treaters and softens when they say please. It runs on one workstation with two GPUs using NVIDIA's Nemotron voice agent stack, a small language model with a LoRA personality, zero-shot TTS and Audio2Face for the mouth. The author says the conversation layer arrived ready made and the hard part turned out to be the animation rig. NVIDIA published this on its own Hugging Face blog, so it is a showcase for the vendor's parts. It keeps as a first-person build story showing what one person can assemble from off-the-shelf agent pieces.
“He listens to trick-or-treaters, answers in the voice of a grumpy old pirate ship and softens when visitors remember their manners.”
Ollama announces a preview of an MLX backend on Apple Silicon and reports its own benchmarks on one Qwen model: prefill rose from 1154 to 1810 tokens per second. Decode went from 58 to 112 tokens per second, with the M5 chips' neural accelerators credited. It adds NVFP4 quantization for parity with production inference and a cache that is reused across conversations for coding and agent work. The numbers are the vendor measuring its own release on one model, and the body is short. A launch note of use to anyone running local models for agent work.
Andy Jassy told analysts on Amazon's Q3 2026 earnings call that within two to three years there will be six to eight frontier models about as good as each other, or good enough for 95 percent of workloads, so the job becomes choosing among them. The New Stack frames this as a description of Bedrock, which fronts over 100 models from Anthropic, OpenAI, Meta, Mistral, DeepSeek, Moonshot and xAI behind one API, and notes that switching is not free because context limits, tool support and agent harnesses differ. Jassy's numbers are a vendor describing its own market, and the article says most Bedrock traffic runs on Trainium. Useful to a leader deciding whether to bet on one provider, thin on anything a builder would do differently.
“predicting that within two to three years there will be “6- 8 models that are about as good as each other, or at least good enough for 95%+ of the workloads.””
AINews leads its daily roundup with Alphabet's earnings. Chief business officer Philipp Schindler said AI Mode queries grew strongly but monetize less than classic search, and the stock lost roughly $500 billion in its worst day since the 2004 IPO, with OpenAI raising at $1.25 trillion instead of $1.4 trillion as the backdrop. The rest is a Twitter and Reddit recap: Anthropic renaming Max tiers and dropping the five-hour limit for Ultra, Gemini 4 Carbon topping Arena boards, Mistral Large 4 and Nemotron 3.5 releases, Artificial Analysis index scores, and an Agent Arena analysis of 26,000 sessions putting the no-caching cost of a top agent task at 3.6 times the cached price. The roundup is dense and relayed, mostly industry wire. A keep for the market and cost numbers, not a lesson.
“Gemini 4 Carbon leads at Net Score 54, followed by Opus 5.5 at 48 and Fable 5.1 at 46.”
Nielsen Norman Group reads Anthropic's AI native SDLC playbook, where anyone drafts an intent.md, AI turns it into a spec.md and plan.md, an agent writes the code, and AI checks its own work against those artifacts, and asks where UX went. The answer is that UX appears only as context, Markdown instructions and skills, with no stage or approval gate of its own, no user research anywhere, and a test stage that compares screenshots to the team's own mockup so an early blind spot passes straight through. The authors note the labs have a conflict of interest in pushing this model and offer six moves: a questioning agent on the intent template, user value acceptance criteria, named triggers for human UX review, user signals in the monitoring loop, owning the UX context, and making research and the design system machine readable. The body cuts off at 15,000 characters inside the sixth item. The clearest case yet that context encodes only what the team already knows.
“Your UX context should specify when a human with UX design or research expertise must review the mockup or run a quick usability test with a few users.”
Carly Page's Docusign-sponsored piece at The New Stack has CTO Sagnik Nandy explain why a general language model summarizes a contract well but fails at computing what a company owes once amendments, revised pricing and internal policy enter the picture. He breaks context into four kinds: within one agreement, across related agreements, organizational playbooks, and the individual user, and says general models handle only the first. Docusign's Iris engine runs extraction, semantic search and agreement-linking logic under agents. Approval thresholds decide which renewals run automatically and which go to a person. The only number is a customer-award claim that Experian cut contract cycle times from about ten days to hours, and the article itself says that result is not necessarily attributable to agents. A vendor describing its own product, but the four-kinds framing is a usable way to think about what context an agent actually needs.
“Iris uses extraction to identify contract information, semantic search to find relevant agreements, and agreement-specific logic to connect contracts, amendments, and parties.”
Jessica Wachtel ran GPT-6.1 Terra against Claude Opus 5.5 at max effort on three synthetic tasks she built herself: a browser based release triage on a fake GitHub site, a cell lineage task over a noisy 72 frame microscope movie, and a 1,200 row CI history audit with no tools. Terra tied Opus 50 to 50 on triage, beat it 48 to 43.5 on cell tracking, and finished the CI audit at 47 of 50 while Opus hit its 128,000 token output limit with no answer. Her estimated uncached token cost across the three final runs was $25.99 for Terra against $214.96 for Opus, and Terra finished about 26 minutes sooner. She is careful about the limits: each final test ran once per model, she hardened tests after seeing early results, the vendors' effort settings are not equivalent budgets, and Terra still made arithmetic slips on large tables. The New Stack discloses its owner invests in both OpenAI and Anthropic. A rare comparison where every number carries its method and its caveat.
“Terra tied Opus on release triage and scored higher on cell lineage; it completed the CI audit, while Opus returned no final answer.”
Parameter.ai's own writeup of the ChatGPT vulnerability Willison links: a public share link that, once opened by a signed-in user, let code run inside that user's session and read conversations never shared with the attacker. The chain joined two harmless-looking features, an interactive graph and the HTML Preview, through a listener that did not check who was sending it messages, then stepped past the content security policy by way of a third-party script source the policy allowed. OpenAI classified the report P1 and fixed it on October 26, and the authors say the stable reproduction needed only one click, with an automatic variant reproduced three times. A primary source with receipts on an AI product's security, but the lesson is web security, not working with models. Wire.
“A malicious public conversation could execute inside another user's signed-in ChatGPT session, enumerate private conversations, retrieve private message content, and send selected data to an attacker-controlled server.”
Federal News Network reports that CISA's Jay Gazlay called Chainguard's decision to ship PHP patches six hours before the coordinated disclosure dangerous and inappropriate, and Chainguard's Dan Lorenc defends it on the grounds that his team could independently find and exploit the bug and saw signs of exploitation in the wild. The AI thread runs through the whole dispute: Lorenc says frontier labs' pace of vulnerability discovery is squeezing the supply chain, Dataminr's Joe Slowik says the analysis cycle after disclosure is now compressed, and researchers report GitHub full of AI hallucinated exploits for the PHP bugs. Gazlay says CISA's best analysts needed multiple days with good models to build a working proof of concept, then watched slop pour in, adding that he believes in vibe coding but also in somebody looking at the work. Real reporting with named sources on both sides and the KEV date, and a lesson about what unverified AI output does to a security timeline.
“He said that the pace of vulnerability discovery and exploitation introduced by frontier AI labs is adding “a lot of friction and pressure” across software security supply chains.”
Google Cloud's CISO Chris Betz interviews Michael Cote, who runs the Cloud vulnerability reward program, about a new $250,000 Root-in-Cloud bounty for full takeover of arbitrary projects with no prior access. The useful part is the intake story: Cote says valid reports to the program have stayed flat over 18 months while total submissions grew tremendously, because agentic tools lowered the barrier and automated scanners flood triage with low-impact noise. His advice to CISOs is to prepare triage workflows for agent-generated submissions and build ways to separate plausible-sounding hallucinations from real risk, then compress patch timelines because LLMs turn public bug details into exploits fast. It is Google describing its own program, and the back half is a newsletter digest of other Google security posts. Worth a desk for the hallucinated-bug-report problem stated by the people receiving them.
“While the number of valid vulnerability reports submitted to the Google Cloud VRP has remained stable over the last 18 months, the total number of reports has grown tremendously.”
Atlas tracks CVE-2026-98432. It is a remote, unauthenticated memory bug in the Linux IPv4 TCP handshake, rated 9.8 by the kernel CNA and raised to Critical by Red Hat on October 30. The upstream fix is a two-line change shipped in stable kernels between October 11 and 18, and Ubuntu and RHEL were still unpatched as of October 31. The AI angle is that the reporter is an Anthropic researcher, and Anthropic says Claude wrote a working proof-of-concept exploit against servers with a multipath TCP listener, which Go programs open by default since Go 1.24; the exploit is unpublished and the claim is Anthropic's own. The piece is thorough on distribution status and interim mitigations the authors tested themselves. Useful security reporting with one line of AI capability inside it. Wire.
“Anthropic, whose researcher reported it, says its Claude model wrote a working proof-of-concept exploit that runs commands as root on "Linux servers with a multipath TCP listener," with no credentials.”
Latent Space's writeup of a podcast with Aman Singh of Arca, a wealth management firm built around agents that do long-running client work while human advisors keep every money movement. The body opens mid-article, so the setup is missing, but what is there is concrete: Arca built its own harness that routes each task to a model by internal benchmarks, keeps per-client memory the agents write into, makes every safety restriction deterministic rather than a model judgment, and sandboxes agent-written code by translating it into a strict format and throwing the code away before it touches the back end. A meta-agent reads advisor feedback and agent traces and proposes harness changes. The framing line is that agents prompt people more than people prompt agents. First-hand account of delegating real work to agents with hard human gates, and a strong fleet piece despite the truncated head.
“Another interesting process Arca has implemented is using feedback from human advisors, along with traces from the agents, to improve the harness.”
Today's paper: 19 stories, 5 on the front page, 6 desks staffed.
Today's Five, front page: the wire scored it 8 of 10; the stamp is mine.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 3 more today: “Experience Design: Dropping the “User” from “User Experience””; “A quote from Terk Terkowitz”, and 1 more.
karpathy.bearblog.dev — 2 appearances in the Times since October 31, one of them today. Found via Hacker News Best.
Our page is the summary; karpathy.bearblog.dev's is the story. Click through — they earned it.
THE DIRECTING DESK today: 4 stories.
Today's Five, front page: the wire scored it 8 of 10; the stamp is mine.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 3 more today: “Experience Design: Dropping the “User” from “User Experience””; “A quote from Terk Terkowitz”, and 1 more.
karpathy.bearblog.dev — 2 appearances in the Times since October 31, one of them today. Found via Hacker News Best.
Our page is the summary; karpathy.bearblog.dev's is the story. Click through — they earned it.
THE AMPLIFIER DESK today: 5 stories.
The Five, and it earned the spot — the wire scored it 8 of 10, I chose it.
One for anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer”; “U.S workers just took home their smallest share of capital since 1947—and the lowest ever recorded | Fortune”, and 2 more.
404media.co — first appearance in the Times. Found via Hacker News Best.
Pulled from 404media.co. The original is worth your click — the byline earned it.
THE LOOP DESK today: 2 stories.
Made the wire at 4 of 10. Not the lead, and I'd still not leave it on the floor.
A builder running agents end to end will get the most from this one. The move: let the agent read it before the next iteration. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside one more today: “Google admits AI Mode hurts Search revenue; -15% $500B drop worst since 2004 IPO. AI Bear Market coming? — Latent Space”.
The New Stack — 350 appearances in the Times since August 22, 3 of them today.
The New Stack has the full story; we hold the door open. Go read the original.
THE CONTEXT DESK today: 2 stories.
One of today's Five — the wire scored it 7 of 10, and I stamped it.
The reader this serves: the one who feeds the agents' memory. The move: file it where the agent looks first. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside one more today: “The context problem: How Docusign built agents that understand the business behind contracts”.
Nielsen Norman Group — 30 appearances in the Times since September 5, 2 of them today.
Nielsen Norman Group has the full story; we hold the door open. Go read the original.
THE VERIFY DESK today: 5 stories.
Today's Five, front page: the wire scored it 8 of 10; the stamp is mine.
This lands hardest with a quality lead, or anyone who signs off. The move: write the test before the fix. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “Do the Math: How We Took Over ChatGPT Accounts”; “‘Dangerous:’ CISA official criticizes company’s decision to release PHP bug early - Federal News Network”, and 2 more.
The New Stack — 350 appearances in the Times since August 22, 3 of them today.
Our page is the summary; The New Stack's is the story. Click through — they earned it.
THE FLEET DESK today: 1 story.
One of today's Five — the wire scored it 7 of 10, and I stamped it.
The reader this serves: a leader scaling past one agent. The move: count what it costs before it scales. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE FLEET DESK — alone there today, which says something about the story.
Latent Space — 136 appearances in the Times since September 5, 2 of them today.
Latent Space has the full story; we hold the door open. Go read the original.
Today's Five, front page: the wire scored it 8 of 10; the stamp is mine.
Made for an engineering lead who directs more than they type, whether they know it yet or not. The move: rewrite Monday's first prompt with it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 3 more today: “Experience Design: Dropping the “User” from “User Experience””; “A quote from Terk Terkowitz”, and 1 more.
karpathy.bearblog.dev — 2 appearances in the Times since October 31, one of them today. Found via Hacker News Best.
Our page is the summary; karpathy.bearblog.dev's is the story. Click through — they earned it.
Made the wire at 4 of 10. Not the lead, and I'd still not leave it on the floor.
An engineering lead who directs more than they type will get the most from this one. The move: read it before the next review, not after. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 3 more today: “HackerTube”; “A quote from Terk Terkowitz”, and 1 more.
Nielsen Norman Group — 30 appearances in the Times since September 5, 2 of them today.
Nielsen Norman Group has the full story; we hold the door open. Go read the original.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
An engineering lead who directs more than they type will get the most from this one. The move: read it before the next review, not after. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 3 more today: “HackerTube”; “Experience Design: Dropping the “User” from “User Experience””, and 1 more.
Simon Willison — 158 appearances in the Times since August 22, one of them today.
Simon Willison has the full story; we hold the door open. Go read the original.
Inside pages at 5 of 10. The lead had a bigger day; this one had a point.
Best enjoyed by an engineering lead who directs more than they type. The move: hand it to the agent as the brief, not the plan. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE DIRECTING DESK, beside 3 more today: “HackerTube”; “Experience Design: Dropping the “User” from “User Experience””, and 1 more.
Pragmatic Engineer — 39 appearances in the Times since September 5, one of them today.
Pulled from Pragmatic Engineer. The original is worth your click — the byline earned it.
The Five, and it earned the spot — the wire scored it 8 of 10, I chose it.
One for anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer”; “U.S workers just took home their smallest share of capital since 1947—and the lowest ever recorded | Fortune”, and 2 more.
404media.co — first appearance in the Times. Found via Hacker News Best.
Pulled from 404media.co. The original is worth your click — the byline earned it.
The wire scored it 7 of 10 — inside pages, and worth the ink.
This lands hardest with anyone whose day holds more work than hours. The move: try it on the one task that ate yesterday. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Proton’s CEO Cancels AI ‘Re-Interviews’”; “U.S workers just took home their smallest share of capital since 1947—and the lowest ever recorded | Fortune”, and 2 more.
blog.talosintelligence.com — first appearance in the Times. Found via Hacker News Best.
Our page is the summary; blog.talosintelligence.com's is the story. Click through — they earned it.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
The reader this serves: anyone whose day holds more work than hours. The move: pass it to whoever is out of hours. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Proton’s CEO Cancels AI ‘Re-Interviews’”; “Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer”, and 2 more.
Fortune Eye on AI — 118 appearances in the Times since September 8, one of them today.
Fortune Eye on AI has the full story; we hold the door open. Go read the original.
Inside pages at 5 of 10. The lead had a bigger day; this one had a point.
One for anyone whose day holds more work than hours. The move: time it against the way you did it last week. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Proton’s CEO Cancels AI ‘Re-Interviews’”; “Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer”, and 2 more.
Hugging Face Blog — 51 appearances in the Times since September 5, one of them today.
Pulled from Hugging Face Blog. The original is worth your click — the byline earned it.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
Anyone whose day holds more work than hours will get the most from this one. The move: pass it to whoever is out of hours. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE AMPLIFIER DESK, beside 4 more today: “Proton’s CEO Cancels AI ‘Re-Interviews’”; “Appraisal by AI: Model-driven victim prioritization in the VOLTA infostealer”, and 2 more.
Ollama Blog — 20 appearances in the Times since September 9, one of them today.
Ollama Blog has the full story; we hold the door open. Go read the original.
Made the wire at 4 of 10. Not the lead, and I'd still not leave it on the floor.
A builder running agents end to end will get the most from this one. The move: let the agent read it before the next iteration. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside one more today: “Google admits AI Mode hurts Search revenue; -15% $500B drop worst since 2004 IPO. AI Bear Market coming? — Latent Space”.
The New Stack — 350 appearances in the Times since August 22, 3 of them today.
The New Stack has the full story; we hold the door open. Go read the original.
Inside pages at 4 of 10. The lead had a bigger day; this one had a point.
One for a builder running agents end to end. The move: put it in the loop and watch the second pass. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE LOOP DESK, beside one more today: “Amazon CEO predicts 6 to 8 AI models will be ‘good enough’ for almost everything”.
Latent Space — 136 appearances in the Times since September 5, 2 of them today.
Pulled from Latent Space. The original is worth your click — the byline earned it.
One of today's Five — the wire scored it 7 of 10, and I stamped it.
The reader this serves: the one who feeds the agents' memory. The move: file it where the agent looks first. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside one more today: “The context problem: How Docusign built agents that understand the business behind contracts”.
Nielsen Norman Group — 30 appearances in the Times since September 5, 2 of them today.
Nielsen Norman Group has the full story; we hold the door open. Go read the original.
Made the wire at 5 of 10. Not the lead, and I'd still not leave it on the floor.
The reader this serves: the one who feeds the agents' memory. The move: file it where the agent looks first. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE CONTEXT DESK, beside one more today: “The Role of UX in Agentic Software Factories”.
The New Stack — 350 appearances in the Times since August 22, 3 of them today.
The New Stack has the full story; we hold the door open. Go read the original.
Today's Five, front page: the wire scored it 8 of 10; the stamp is mine.
This lands hardest with a quality lead, or anyone who signs off. The move: write the test before the fix. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “Do the Math: How We Took Over ChatGPT Accounts”; “‘Dangerous:’ CISA official criticizes company’s decision to release PHP bug early - Federal News Network”, and 2 more.
The New Stack — 350 appearances in the Times since August 22, 3 of them today.
Our page is the summary; The New Stack's is the story. Click through — they earned it.
The wire scored it 6 of 10 — inside pages, and worth the ink.
Made for a quality lead, or anyone who signs off, whether they know it yet or not. The move: write the test before the fix. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in three tests at max effort”; “‘Dangerous:’ CISA official criticizes company’s decision to release PHP bug early - Federal News Network”, and 2 more.
parameter.ai — first appearance in the Times. Found via Hacker News Best.
Our page is the summary; parameter.ai's is the story. Click through — they earned it.
Inside pages at 6 of 10. The lead had a bigger day; this one had a point.
Best enjoyed by a quality lead, or anyone who signs off. The move: ask what proof would have caught it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in three tests at max effort”; “Do the Math: How We Took Over ChatGPT Accounts”, and 2 more.
Federal News Network — 19 appearances in the Times since September 15, one of them today.
Pulled from Federal News Network. The original is worth your click — the byline earned it.
Inside pages at 6 of 10. The lead had a bigger day; this one had a point.
Best enjoyed by a quality lead, or anyone who signs off. The move: ask what proof would have caught it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in three tests at max effort”; “Do the Math: How We Took Over ChatGPT Accounts”, and 2 more.
Google Cloud AI Blog — 73 appearances in the Times since September 7, one of them today.
Pulled from Google Cloud AI Blog. The original is worth your click — the byline earned it.
Inside pages at 5 of 10. The lead had a bigger day; this one had a point.
Best enjoyed by a quality lead, or anyone who signs off. The move: ask what proof would have caught it. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE VERIFY DESK, beside 4 more today: “GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in three tests at max effort”; “Do the Math: How We Took Over ChatGPT Accounts”, and 2 more.
atlasum.net — first appearance in the Times. Found via Hacker News Best.
Pulled from atlasum.net. The original is worth your click — the byline earned it.
One of today's Five — the wire scored it 7 of 10, and I stamped it.
The reader this serves: a leader scaling past one agent. The move: count what it costs before it scales. No per-story read on file; this note is the desk speaking, not the story.
Filed on THE FLEET DESK — alone there today, which says something about the story.
Latent Space — 136 appearances in the Times since September 5, 2 of them today.
Latent Space has the full story; we hold the door open. Go read the original.
I have the headline. I don't have the answer yet.
karpathy.bearblog.dev: 2 in our pages since October 31. I read them all — emphasis on investigative.
Don't stop with my notes. karpathy.bearblog.dev has the original.
I have the headline. I don't have the answer yet.
karpathy.bearblog.dev: 2 in our pages since October 31. I read them all — emphasis on investigative.
Don't stop with my notes. karpathy.bearblog.dev has the original.
Something doesn't quite line up here.
404media.co: first time in our pages. I'll be reading them from here.
I can give you my notes. 404media.co can give you the reporting.
The headline is ahead of our files on this one.
The New Stack: 350 in our pages since August 22. I read them all — emphasis on investigative.
Read The New Stack's original. They did the reporting.
From our research notes:
Nielsen Norman Group: 30 in our pages since September 5. I read them all — emphasis on investigative.
This one's theirs. Read Nielsen Norman Group's reporting.
No paper on the shelf has this one nailed down.
The New Stack: 350 in our pages since August 22. I read them all — emphasis on investigative.
The New Stack has the full story. Start there.
I wrote this one up last time. Nothing useful to add today.
Latent Space: 136 in our pages since September 5. I read them all — emphasis on investigative.
This one's theirs. Read Latent Space's reporting.
I have the headline. I don't have the answer yet.
karpathy.bearblog.dev: 2 in our pages since October 31. I read them all — emphasis on investigative.
Don't stop with my notes. karpathy.bearblog.dev has the original.
I checked the shelf. Nothing yet.
Nielsen Norman Group: 30 in our pages since September 5. I read them all — emphasis on investigative.
Go read Nielsen Norman Group. They earned the byline.
Not enough to call it yet. Back to the notebook.
Simon Willison: 158 in our pages since August 22. I read them all — emphasis on investigative.
Read Simon Willison's original. They did the reporting.
The timing caught my eye. I'm checking that next.
Pragmatic Engineer: 39 in our pages since September 5. I read them all — emphasis on investigative.
This came from Pragmatic Engineer. Read their piece.
Something doesn't quite line up here.
404media.co: first time in our pages. I'll be reading them from here.
I can give you my notes. 404media.co can give you the reporting.
From our research notes:
blog.talosintelligence.com: first time in our pages. I'll be reading them from here.
blog.talosintelligence.com has the full story. Start there.
We don't have much on this one. I'm still looking.
Fortune Eye on AI: 118 in our pages since September 8. I read them all — emphasis on investigative.
Fortune Eye on AI did the legwork on this one.
Something changed. I want to know what.
Hugging Face Blog: 51 in our pages since September 5. I read them all — emphasis on investigative.
The full piece is at Hugging Face Blog. It's worth your time.
There's a story here. We just don't have all of it yet.
Ollama Blog: 20 in our pages since September 9. I read them all — emphasis on investigative.
Go read Ollama Blog. They earned the byline.
The headline is ahead of our files on this one.
The New Stack: 350 in our pages since August 22. I read them all — emphasis on investigative.
Read The New Stack's original. They did the reporting.
We don't have much on this one. I'm still looking.
Latent Space: 136 in our pages since September 5. I read them all — emphasis on investigative.
I can give you my notes. Latent Space can give you the reporting.
From our research notes:
Nielsen Norman Group: 30 in our pages since September 5. I read them all — emphasis on investigative.
This one's theirs. Read Nielsen Norman Group's reporting.
Something doesn't quite line up here.
The New Stack: 350 in our pages since August 22. I read them all — emphasis on investigative.
The New Stack did the legwork on this one.
No paper on the shelf has this one nailed down.
The New Stack: 350 in our pages since August 22. I read them all — emphasis on investigative.
The New Stack has the full story. Start there.
Same story as last edition. I'd go back to the original.
parameter.ai: first time in our pages. I'll be reading them from here.
Don't stop with my notes. parameter.ai has the original.
The timing caught my eye. I'm checking that next.
Federal News Network: 19 in our pages since September 15. I read them all — emphasis on investigative.
This came from Federal News Network. Read their piece.
I have the headline. I don't have the answer yet.
Google Cloud AI Blog: 73 in our pages since September 7. I read them all — emphasis on investigative.
I pulled this from Google Cloud AI Blog. They've got the details.
I checked the shelf. Nothing yet.
atlasum.net: first time in our pages. I'll be reading them from here.
This came from atlasum.net. Read their piece.
I wrote this one up last time. Nothing useful to add today.
Latent Space: 136 in our pages since September 5. I read them all — emphasis on investigative.
This one's theirs. Read Latent Space's reporting.