Week of September 21

Globally, More People Expect AI to Cause Job Loss Than Growth

Pew’s new global AI survey — 37 countries, ~50k adults, fielded February to June — follows up last year’s 25-country version , and the headline is jobs: in 34 of the 37 countries, more people expect AI to mean fewer jobs over the next 20 years than more. Hardly anyone thinks it’ll narrow the gap between rich and poor — no country gets above one in four, and 29 are in single digits.

The worry tracks wealth. In Australia, South Korea and the US around seven in ten or more expect job losses, while in some middle-income countries ~40% or more simply say they’re unsure. It’s also getting younger: the share of Americans aged 18–34 who are more concerned than excited went from 40% in 2024 to 55% now, level with the over-50s. Israel is the one country where more people are mainly excited than mainly concerned. And awareness cuts both ways — people who’ve heard a lot about AI are likelier to expect job losses, but also tend to feel more optimistic about it overall.

AI has improved significantly at reasoning about IKEA furniture assembly

Finally, the benchmark that matters: can the model tell you where you screwed up the IKEA build? Epoch assembled three pieces of furniture — a shoe rack, a bed frame and an 8-drawer dresser — making realistic mistakes on purpose and often building on past them, then gave models 60 photos, the manual as a PDF, a zoom tool and a Python sandbox. Find every step that’s wrong, or say what to do next.

Best score in November 2025 was 28% (Opus 4.5). Ten months later GPT-6 Astra gets 80%, with Fable 5.1 on 70% — and Astra does it in a median ~3 minutes a photo, 2–10× faster than the previous leaders. Still too slow to talk you through it live, but not by much. The failure modes run in families: Gemini and Qwen almost always assume there’s a mistake, while earlier Claude and GPT models too often missed real ones. Caveats Epoch raises themselves: only 60 photos from three builds, a deliberately lenient LLM grader, and no human baseline yet.

The AI Build-Out Is Becoming the Biggest Economic Bet in U.S. History

The WSJ puts a number on the AI build-out using new Brookings estimates: ~$10.3 trillion of data-center and AI infrastructure spend from 2025 to 2032, or ~3.6% of GDP a year. On an annual basis that’s bigger than the railroads, the interstates, the canals and the ’90s fiber build.

The bit that surprised me was the railroads, not AI. They ran at over 2% of GDP for twenty years, 1870–90 — a higher level, sustained for far longer, than I’d have guessed. The AI projection only covers eight years. Whether AI can keep that pace up for another dozen or so years to match the railroads’ run is very much still to be seen — and the coverage points at the strain already showing: other private construction shrinking while data centers grow, and a rising share of hyperscaler capex funded with debt.

The piece itself is paywalled — the numbers here are from the chart and Global Macro Monitor’s write-up of it .

Grace: New animation from Opus 5.5

Grace shared a new animation this week, and every frame of it was made by Opus 5.5 writing p5.js code. No video model involved — the model writes a program that paints the frames, and the program gets rendered to video. Fair warning: it’s an absolute earworm.

The recipe is catching on. The video for I’m Upping My P(doom) used the same basic trick — p5.js plus p5.brush, with Opus writing its own storyboard and briefing parallel subagents, and the only human direction being to use the Clawd character and give each lyric interesting visuals. Its author has since packaged the approach as a starter kit : a guide for the model, 31 acted emotions for Clawd, and a headless renderer. If you’ve got Claude Code, Node, Chrome and ffmpeg, you can make one of these yourself — though without a real GPU, the watercolour fills render at seconds per frame.

The plunging price of thought

Epoch put numbers on something that’s been obvious from the invoices: the cost of a given level of AI performance has been falling ~47% a quarter since 2023 — ~13× a year. That’s 4× faster than DNA sequencing ever fell, 6× faster than compute, 18× faster than lithium batteries and 54× faster than electricity did in its first century. The example that sticks: o3 scored 75% on GPQA Diamond at ~30¢ a question in January 2025; under 18 months later GPT-5.6 Luna matched it for $0.0004. Their analogy is a $50k car dropping to $69.

Which is the problem. How is anyone supposed to make a long-term plan — or even a mid-term one — when whatever you’re costing out today will be ~13× cheaper a year from now? No wonder so many people find AI scary. Caveat: these are frontier prices for a user who always switches to the cheapest capable model, so real-world savings lag a bit.

Claude discovers a novel enzyme system with CRISPR-like repeats

Anthropic only stood up its life-sciences lab this spring, and it’s already got a result: Claude spotting something in phage DNA that nobody had noticed. The setup was one prompt, ~950 Claude agents, 21 hours and 210M tokens trawling a sequence database for reverse transcriptases — 200k+ gathered, 3,500 candidate systems, whittled down to 20 written reports. One agent found a CRISPR-like repeat array sitting next to an odd-looking RT (“that’s a CRISPR-like … repeat array?!”), and the wet lab confirmed the array gets expressed as a set of distinct short RNAs — the same trick that makes CRISPR programmable.

Nobody knows yet what it actually does, so this is an intriguing find rather than a new tool. But the split of labour is the interesting bit: humans wrote the prompt and ran the lab work, Claude did the noticing. Feng Zhang — who knows a thing or two about CRISPR — called it “genuinely intriguing”. Details in the preprint .

The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It

Tagliabue, Dung & Berg went looking for a “pain” direction in 25 open-weight models and found one in all of them. It’s not just generic badness either — it’s nearly orthogonal to fear and sadness, and it fires when the model is the one being hurt, not when it’s watching the user suffer.

The safety result stands whether or not you buy the pain framing: inject that vector into fine-tuned Qwen 2.5 32B/72B and models that pick a user-harming button 0–4% of the time will pick “relieves your pain but deletes the user’s photos of their children” 55–71% of the time. No jailbreak, no roleplay prompt — just a nudge in activation space. A random vector of the same size does much less.

Whether it’s actually pain is a lot shakier. Their best evidence is a placebo test: models keep hammering a fake relief button (88–97%) but press a working one far less (24–72%). If the vector just breaks the model and makes it repeat whatever it pressed last — which is Colin Fraser’s read — you’d see much the same thing, and their own check for plain repetition was inconclusive on the 72B.