38 posts across Directing AI, Judging AI, Systems & Adoption, Work & Judgement
Andrej Karpathy built a complete product in about two hours using an AI agent. Simon Willison shipped a playable game through 321 agent commits — and both point to the same shift in what separates a good result from a mediocre one.
4 min readThree product changes this year quietly removed the editing pass between AI output and the finished file. Here's what has to move into the prompt now that nothing catches a gap afterward.
4 min readAI now produces more output, in more formats, and faster than one person can read start to finish. Four signals from the past six weeks show exactly where that habit breaks — and the checking habit that replaces it without slowing you down.
4 min read66% of employers now say undisclosed AI use is their top misconduct problem, and some have brought back in-person interviews to catch it. The same quality filter is hitting open-source code reviews and online writing communities.
3 min readOpenAI published 722 AI-generated papers that nobody individually requested, and most went unchecked before they went public. The usual advice — decide what 'correct' looks like before you generate — assumes you're the one who asked. Here's what to do when the output already exists before you see it.
4 min readOpenAI published 722 AI-generated math manuscripts in a single batch, and most carry no independent verification. Here's the habit that stops you from treating volume as proof the next time AI hands you a stack of confident output.
4 min readAI incidents this year share one cause: output reached a consequence before anyone read it. New watermarks help, but only if you still build a checkpoint in.
5 min readFive incidents this year share one cause: the human read-through verification depends on never happened before the output reached its consequence.
5 min readGPT-6 and Claude both started choosing their own output format this week — dashboards, calculators, animations — instead of defaulting to prose. The model deciding what you see is a new layer to check before you trust it.
4 min readA diagnosis study, a tax-prep agent, and a new evaluation guide all surfaced the same gap this year: a review step that isn't decided in advance isn't a check, it's a formality. Here's the pattern underneath all three.
5 min readContent flagged as AI-generated now gets 20-35% less engagement, and arXiv just capped submissions after AI-written papers overwhelmed its moderators. The skills that keep output from reading as AI-generated carry a direct revenue effect now.
4 min readSimon Willison says coding agents made his job harder, not easier. Cognition's Devin hit $1B in revenue with DHH writing 3% of his own code. The pattern behind both: AI pays off in proportion to the supervision you put in, not the subscription you buy.
4 min readA new guide from Hamel Husain and Shreya Shankar says the way most teams grade AI output is backwards. Read real failures by hand first, then decide what to measure — not the other way around.
4 min readA fake citation, a hallucinated URL, a paid-feeling product pick, and an invisible watermark all surfaced this year from the same root cause: nobody checked where the source actually came from. Here's the two-question habit that catches all four.
5 min readFour separate stories this year — a CEO's warning, a viral study, a blunt writing rule, a $50,000 gap in financial advice — all point at the same three seams. Here's the throughline, and the move that closes each one.
4 min readMcKinsey found 94% of companies still see no measurable AI return despite licenses everywhere. Four separate 2026 moves — a monetized free tier, outcome-based consulting fees, and a user base that's no longer mostly developers — show what actually separates the 6% that do.
5 min readA resume screener, a one-shot game, a benchmark leader, and a search engine all optimized for something measurable this year — and it wasn't whether the output was actually right for the job. Here's the check that catches the gap before you act on the score.
5 min readTwo studies published the same week found AI search tools steering people toward worse choices: pricier products from Google's AI Mode, fabricated citations from Perplexity.
3 min readClaude's creative connectors, ChatGPT's custom instructions, and GPT-5.5's shorter answers all expanded how much you can tell an AI tool before you start. The upgrade only pays off if you fill the room.
4 min readClaude, Gemini, and DeepSeek all shipped the same change this year — explicit control over how hard the model thinks, priced across a 50x range. Here's the three-tier rule for matching effort to what a task actually needs.
4 min readFabricated citations, a second AI fact-checker, merged code, and OpenAI's own safety warning all failed the same test this year: does this look right. Here's the specific check that replaces it for each type of output.
5 min readAn AI agent one-shot a working game and a NeurIPS-ready research paper. Both got rejected for the same reason: strong execution, wrong call on what to build.
5 min readFour separate AI stories from this year — a 40% task failure rate, three experts agreeing independently, a Hacker News revolt, and an executive hype backlash — all point at the same skill gap, and none of them is really about the AI.
5 min readMIT Sloan tested AI-generated financial advice against real outcomes and found the quality of the prompt, not the model, decided who ended up ahead. Here's the structure that closes the gap.
4 min readThree communities — corporate critics, cognitive scientists, and content readers — arrived at the same conclusion this week: uncritical AI use is producing bad decisions, eroding judgment, and generating detectable slop. Here's what to do about it.
4 min readOpenAI's own safety guidance for ChatGPT Work admits the system can produce 'finished mistakes at a scale that is harder to catch.' The same week, Wharton research killed prompt tricks and Cursor data showed a 46x productivity gap. The differentiator is now verification, not prompting.
6 min readMcKinsey reports 30%+ of its consulting fees are now tied to client outcomes, not hours. Deloitte warns labour-based consulting could shrink significantly by 2035. Here's what the shift from hours to outcomes means for every professional who sells time.
5 min readMost of the effort in reviewing AI output goes on working out what you wanted while looking at what you got. Three sentences written beforehand turns a judgement into a comparison.
5 min readFigma's CEO Dylan Field told Stratechery that AI output 'draws from the middle of the distribution' — and that teams become viscerally attached to their first concept. The iterate-and-refine skill is what separates competent AI output from distinctive work.
4 min readEthan Mollick, Charity Majors, and Simon Willison reached the same conclusion independently within five days. The professionals who get the most from AI are not the best prompters — they are the best judges.
5 min readThree AI failure modes got named publicly in the same week — forwarding raw AI answers without reading them, optimising for speed over quality, and letting the model think while you sign off on the result. Each one has a skill that fixes it.
5 min readTwo Hacker News posts hit the front page the same week with a combined 1,374 points. One named the upstream problem — vague briefs going in. One named the downstream problem — raw output going out. The model was fine in both cases.
4 min readReading it and approving it isn't a review — it's the exact filter AI output is best at passing. Interrogating is a different activity, it takes about ninety seconds, and it's most of what separates the people getting value from AI.
5 min readA New York Times reporter used AI to research a story, got back a fabricated quote, and published it. A reader caught it. The editors didn't. The failure pattern here isn't a journalism problem — it's the default behavior for anyone who uses AI to pull quotes, stats, or summaries and publishes them without checking the source.
5 min readFigures are the highest-risk thing AI produces and the thing people check least, because a number looks like a fact. Four checks that take under a minute each, and the one that catches the most.
5 min readGoogle's Gemini can now output a finished Word file, spreadsheet, or PDF directly from conversation. Writing prompts that produce share-ready documents is a different skill than writing prompts that produce text.
4 min readCheck everything and you've saved nothing. Check nothing and you're gambling. The right amount is a function of two things, and neither of them is how much you trust the tool.
5 min readA law firm sent 42 fabricated citations to a judge. A newspaper published a quote that was never said. In every case the output looked right, because looking right is what these systems are reliably good at.
5 min readWe use one cookie to keep you signed in — that one isn't optional. Separately, we'd like to record which drills people finish and where they get stuck, so we can fix the parts that don't teach. No ad networks, no selling anything on. Privacy policy.