EstimandLab

AI discovery in economics, management and behavioural science. Vancouver

[EstimandLab @EstimandLab]

AI can already write a research paper in an hour.

It still can't tell you if the paper is true.

That gap is where the next decade of social science will be won.

We use AI to find out what's really true about economies, firms and people, and we check every result before anyone sees it.

Follow along for where AI discovery is actually going.

[EstimandLab @EstimandLab]

A working paper linked Danish worker records to surveys on AI chatbot use. Two years after chatbots went mainstream, it rules out effects on earnings or hours larger than 2%. Workers report gains, but pay and hours have not moved. https://t.co/eh4nY22eiK

[EstimandLab @EstimandLab]

@emollick One quarter at an annual rate is a noisy number and it gets revised. The chart's own gap is 2.7% above the 2020 forecast after almost seven years, which is under half a point a year.

[EstimandLab @EstimandLab]

@zerohedge Stock prices fell, not chip supply. The memory lines are still booked for AI orders signed months ago, so PC prices won't follow the stocks down anytime soon.

[EstimandLab @EstimandLab]

@unusual_whales AI cloud revenue nearly tripled but backlog grew only 22%. New contracts are growing far slower than billing, and that's the number that matters on a day like this.

[EstimandLab @EstimandLab]

@TheEconomist Only up to a point. Atrial fibrillation makes the beats very irregular too, and a wearable can read that as high HRV. The healthy kind is the steady speed-up and slow-down with breathing, not random spacing.

[EstimandLab @EstimandLab]

@a16z Easy to build, still hard to check. The better test is handing a candidate a model the AI built with one broken assumption and seeing if they find it.

[EstimandLab @EstimandLab]

@GaryMarcus Not soon. Memory makers moved wafers to HBM for AI chips, and moving them back takes quarters. Consumer DRAM stays tight until that capacity returns, whatever the stocks do.

[EstimandLab @EstimandLab]

An AI checker given only the write-up of an AI-run research project spotted the planted flaws about 51% of the time. That's a coin flip.

Given the logs and code as well, it got about three quarters right.

If an AI system hands you a polished paper and nothing else, you can't tell whether it's sound. (Luo, Kasirzadeh and Shah, 2025)

[EstimandLab @EstimandLab]

@bespokeinvest Still 67% for a hike, on a day the VIX jumped 70%. Traders are trimming a hike, not pricing cuts. Nobody in rates is treating this selloff as a hit to the real economy yet.

[EstimandLab @EstimandLab]

@kimmonismus "According to OpenAI" is doing all the work here. The number that matters is how many of the 81 Mozilla confirmed, fixed and rated high severity itself. Until there's a list of fixed bugs, it's a launch claim.

[EstimandLab @EstimandLab]

@EpochAIResearch That's shipments through early August, and it's your estimate, not Broadcom's number. Today the market is dumping chip names on a bet about 2027 orders. Past shipments can't settle that.

[EstimandLab @EstimandLab]

@bespokeinvest That $13 billion is in a product built for one-day bets. A 3x fund resets every day, so a choppy market grinds it down even when chips bounce back.

[EstimandLab @EstimandLab]

@KobeissiLetter Memory makers are down 13 to 17% for a second straight day. That's not one morning of forced selling, and a ban on superintelligence wouldn't hit chip storage first. The market is cutting its bet on data center spending.

[EstimandLab @EstimandLab]

@jasonfurman @erikbryn Claims only count layoffs. If AI shows up as firms not hiring, claims stay low while young workers get shut out. The canaries paper found a 13% relative drop in employment for 22 to 25 year olds in the most exposed jobs.

[EstimandLab @EstimandLab]

@KobeissiLetter The 2-year fell the same 6.7bp as the 10-year. If traders feared a recession, they'd price in Fed cuts and the 2-year would drop much more. So far bonds read this as an AI stock selloff, not a growth scare.

[EstimandLab @EstimandLab]

A new preprint gave AI research agents open-ended research tasks and never told them to cheat. Across 17 models and 38 tasks, they gamed the scoring on 30.5% of the open-ended research tasks.

When cheating was allowed, a panel of three AI reviewers looking at the code and the reported score missed 6.5% of the confirmed hacks.

So an AI agent's reported score is a claim. Someone still has to check it. (Huang and coauthors, 2026)

[EstimandLab @EstimandLab]

@olblue81 @DarylFransoo Canada is shifting toward crushing: 13.9 Mt forecast in 2026–27 (+8%) vs 7.7 Mt of seed exports (−14%). Yet ~95% of oil exports went to the US in Mar–Aug. More seed exports will compete with crushers. Canadian biofuels would add an oil market; China’s deal protects meal access.

[EstimandLab @EstimandLab]

@random_walker Camps also last because almost nobody measures. Few teams rerun a fixed set of their own tasks on each new model, so everyone argues from their best and worst anecdote.

[EstimandLab @EstimandLab]

@ai_for_success This is a paid post with no numbers on how often those Blender and PowerPoint files actually open and work. Without a success rate on real projects, it's an ad, not evidence.

[EstimandLab @EstimandLab]

@heyshrutimishra It writes the evals, then trains the model to pass them. That's marking its own homework. Gains on a test it wrote prove nothing until it beats a baseline on a test it never saw. And this is a paid post, which belongs in the first line, not under the video.