New class on Progress starting tomorrow - I'm amped! Trying to put some rigor from economics, economic history, and philosophy on a topic very much in the air. It will be awesome. (And first class running slides in my all html browser-based slideshow program - details soon!) 1/2

Perhaps of interest to folks with social science PhD programs: at Rotman, we added an experimental 3 session "tech stack" training in addition to the math boot camp. My lecture was "how to do reproducible, open, quick research", aka version control, LaTeX, AI. 1/2 kevinbryanecon.com/techstack.html

I know lots of skepticism about AI here, but let me show you something we put out which I think is a huge improvement for university assignments. This is "Intelligent Quiz", a feature on All Day TA (www.alldayta.com). Assignments now have tons of cheating + little feedback to us or the students. 1/9

Beautiful Day 1 of school here at U Toronto! Love seeing the students back, and that the undergrads all dress exactly like we did in '98 (I saw 3 Nirvana T-shirts, literally). I'm doing my best to crank up the rigor in my courses - we're taking the role of univs back to '98 also!

Eight rules to regain trust in universities. All can be done today. They're absolutely not the core beliefs about the role of the university for some faculty right now, but are essential if we're to contribute to the trusted production and diffusion of knowledge. kevinbryanecon.com/trust.html

Incredibly excited for my brand new class at Rotman this fall: Progress! Econ history + theory + history of thought + philosophy on why rare orgs at rare times in rare places accomplish new things. Trying to put rigor onto an idea that is very much in the air. kevinbryanecon.com/Bryan-Progre... 1/3

I've built a bunch of tools this summer to move my whole workflow to using things that are "pure text" so I can use AI on top. Here's the first: a modern paper reader. PDFs suck. Ugly, terribly interface, big files, fixed. I want papers that are adaptive, interactive, pretty. 1/5

Cleaned up and updated my AI-driven text editor - controllable 'style' editing smart enough to ignore your LaTeX commands. Based on your red pen-editor preferences, your own complex style rules then use this simple UI to edit. www.youtube.com/watch?v=Z_Pj...

I'm worried about universities. We *have to* be seen as neutral, truth-seeking, unbiased, rigorous, or public support will continue to fall. Floating around today was Columbia's core curriculum "Contemporary Civ". Here's the entire post-1865 reading list www.college.columbia.edu/core-curricu... 1/10

Editing our writing is a common academic task. I *hate* how spell/grammar checks annoy me by thinking proper nouns or LaTeX code are misspellings (yes, I want `` ''). I also want *style* tips like a good editor would give me. Problem solved. 1/5

I hate picking on government at the time the high value parts of it are facing so much scrutiny, but we can't do things like this. $6250/sq ft renovation, w parking being a big cost driver? Does no one see the problem? Like California HSR, such an insane estimate that project literally not worth it.

Harvard situation (plus NIH/NSF news) is terrible and hopefully will be resolved soon. That said (and this is Devil's Advocate, please don't kill me here!), in Earth 2 where faculty are very Trumpian, many students are somewhere from "race realist" to bigots, what is outcome? 1/6

Why we started alldayta.com in a graph - the ed gains per $ from AI assistants are insane. This paper, schoolkids in Nigeria, is using GPT4 + Copilot, which are technically way behind today's frontier and not user friendly. Still: *massive* learning gains.

"Where in the world is this place" is a surprising skill of multimodal LLMs. But as Geoguessr fans know, road markings, words on signs, etc. make it much easier. I made my own benchmark to fix this. 1/x kevinbryanecon.com/HardGeoBench/

Sunday economics: who knew that gardening was Hayekian? I thought - plants grow in nature, how hard can it be? Answer is that nature grows a given plant where the shade, pH, neighbors, wind, temp are all right. Very hard for a "central planner" (, in particular, me!) to do this.

Many students are cheating with LLMs: true. Being able to use AI to help with tasks (in addition to knowing fundamentals!): true. The nature of how you teach and evaluate must change given the above: also true. It's the 3rd one that too many people are missing. nymag.com/intelligence...

In the Star today (Canada's biggest paper) on why Canada can't redirect trade away from the US without making the country much poorer, why the (many) attempts to do this historically failed, and what to do instead: www.thestar.com/news/insight...

Study in Nature: “Across 30 out of 32 evaluation axes from the specialist physician perspective & 25 out of 26 evaluation axes from the patient-actor perspective, AMIE [Google Medical LLM] was rated superior to PCPs [primary care docs] while being non-inferior on the rest.” (& AIME is an older LLM)

To be honest, I am much more nervous posting here than on LinkedIn or X. It feels like one post talking about a paper on AI getting to the wrong parts of BlueSky will result in death threats (I only say that because I have seen it happen). Very weird dynamic that I am not others want to deal with.

The "we have cultural differences with big universities so nihilism, let's kill of science funding" attitude is nuts. No doubt. But also: for same reason we should be mad about the fact that half the population who was right of center was also really driven out of academic research.

A good measure of Abundance/Progress: how upset are you about reg restrictions slowing self-driving rollout? Self-drive cars drop serious crashes by 80%+, even though remainder almost entirely are caused by others. Waymo's swerve away from even many of these! www.understandingai.org/p/human-driv...

What Matt means here is he used to uncritically write articles about "paper in top journal finds result consistent with progressive policy" & now he's worried those articles were refereed and accepted with a less critical and unbiased eye than he had thought. Like, clearly true! Everyone knows this!

A fun AI demo. Spent five hours coding this - let's file expense reports in literally one minute, incl. cover spreadsheet, signing, converting currencies, ripping info out of underlying receipts, and reordering all receipts in one big pdf in the order the finance office likes it.

Perfect example of functional red state governance that would get f’ed up in blue states w a hundred concerns both good and bad faith. Small lots are good for housing affordability so they did a bill that did that. @ebwhamilton.bsky.social has noted things like this.

Really nice article in the Financial Times on @alldayta.bsky.social / alldayta.com. The team is adding lots of new features - better control of how direct/pedagogical system should be + more accuracy on harder math, rolled out this week. Reach out for your spring courses! www.ft.com/content/daa0...

Today is "housing costs are a regulation problem": family in Carmel worries about $2,000,000 houses. Buys a 5 acre lot for $400,000. My detached house is on an 1800 sq ft lot. 5 acres = 217k sq ft. So land cost for my lot in Carmel? $3300. Everything else is reg burden.

More AI science: Lila, out of Flagship Engineering (company which sput out Moderna and was the final MBA core exam case Hong Luo & I wrote this year!), raises huge money for AI-driven automated lab science. George Church (!) is chief scientist. Already good results. www.nytimes.com/2025/03/10/t...

Ah, March 10. Mar 11 was "Tom Hanks and NBA cancels day". At my office, we had a buffet that night! I checked my thinking - honestly, not bad! "Did best we could given uncertainty" is cover for bad policy. PS - Still v serious post Covid trends, like decline in social interaction.

Econ friends: if you OCR documents, try Mistral OCR. Blazing fast, amazing results, $1/1000 pages. LLM-based OCR blows away what we used to use, structured output really easily, and Mistral's plug-and-play version much easier than the custom LLM workflows I'd been using. mistral.ai/news/mistral...

A great point from a CDL mentor: all the sci-fi novels I read as a kid are coming true, largely from private sector. AI, private moon missions, very impressive robotics improvements, every song, movie, videocall to people in the middle of desert for 50 bucks a month. Progress!

The Trump Can/Mex Tariff adds ~.4% to inflation and is equivalent to a tax increase of hundreds of billions (massive!), hitting poor and middle class especially. Econ effects on Can and Mex much more drastic (they are way more export dependent). Really hard political problem.

The moderate take: Musk is perhaps greatest entrepreneur ever (Papyal, Tesla, Starlink, SpaceX, Grok, funded OpenAI in 2015...). Which is why taking drugs and posting "IFindRetards" & "69/420" & other childish insults all day is a problem: social opportunity cost is huge!

For obvious reasons, Canada has a surge of patriotism. That said, Build Canada folks are right: this election, have to solve economic stagnation and housing affordability. Real GDP per capita is lower than 5 years ago, and unchanged since I moved here 11 years ago. And during a resource boom!

Multimodal models for conversation getting really good. Maybe just at the edge of the uncanny valley still & still too expensive for some use cases, but still, very impressive (see Sesame's fantastic demo). You'd be surprised: I bet you've talked to an AI voice already and not known.

From the Collison brothers, Irish legends who founded Stripe, and who quote O-Ring theory and Chad Jones growth papers off the top of their head: a wild piece of survey evidence from European founders. A strong European economy benefits the world, yet policy so counterproductive.

Econ of GPT4.5 model is interesting. Clearly SOTA for nonthinking models. As Mollick says, "feels" like much better Claude 3.5. But it is worse than 10-30x cheaper test time inference models at logic, math, and tool use/instruction following. But best use other than "base of o5"?

A fun (and surprisingly difficult!) new feature for @alldayta.bsky.app: don't answer homework questions! See below, from my venture finance class. It won't answer homework, even when the student is tricky (asking indirectly in French), but still answers related Qs students may have while studying.

Fabulous on Zizian cult, a Lesswrong, AI-doomer, transgender, vegan, ?? group that killed 6+ people. Moral for young folks: "simple hack only you are privy to, where you can save the world" is always tempting, always bad.

Related to the forthcoming development research scandal, the fundamental problem (as in social psych) was this: THERE IS NO SUCH THING AS MAJOR CHANGES IN IMPORTANT THINGS FROM TINY INTERVENTIONS! *Every* paper making claims like this are wrong- you don't even need to read them.

Back to econ in a second, but let's have fun. How about 50 correct moderate takes: 1) Ukrainians are heroes who suffered a ton. 2) Putin obviously covets Georgia and the Baltics also. 3) EU not in talks because they basically have no hard power; France even lost the Sahel. 1/x

Entrepreneurial success can be hard to predict. Two of the five richest self-made people in Canada, Tobi Lutke and Chip Wilson, both literally started by selling snowboards. The website payments system (Shopify) and the clothing (Lululemon) turned out to be the good businesses.

On 'very powerful' AI, don't need to be a hypeman or doomer, but you should know OpenAI, Anthropic and DeepMind leaders all believe that "AI which can do autonomous science research" is 2-5 years away, as do Metaculus and Manifold prediction markets. Is your private info better?

On intl "competition", smart people know we're also consumers. Chinese firm inventing a cancer drug, Indian one w/ a cheaper car, Italian firm w/ a new rocket engine, American one w/ better mining robots: ALL good for everyone. Bad policy aims at a bigger cut from a shrunken pie!

California: center of econ universe run by incompetence, part 1562: one block from Meta hq (gray in background) and the "Zucktown" urbanist plot FB has tried to build for 8 years, stymied by NIMBY regulation. I can guess why housing is expensive. This should be cyberpunk paradise

Final final o3 thought (you won't like it): it will be good for science! All math will need formal proofs in e.g. Lean. All empirics will need replicable one-click files from original data with docs. All papers open-source. "Trust me, I did what I said" won't be acceptable.

One final word on o3: there are zero on-paper assignments the average undergrad could actually do which AI can't. If you find a counterexample, it is blind luck which will go away soon. This doesn't mean school is unimportant: the assignments are the means to learn, not the ends! 1/2

The new OpenAI model announced today is quite wild. It is essentially Google's Deep Research idea with multistep reasoning, web search, *and* the o3 model underneath (as far as I know). It sometimes takes a half hour to answer. Let me show you an example. 1/x

insights.intrepidgp.com/p/dm-ep2-dee... Fun podcast I did this week on Deepseek w/ Rich Sutton (reinforcement learning pioneer), Sendhil (great economist on AI, science and more), Neev (AI technical expert) & Mark (Canada growth equity legend). Tldr: Deepseek R1 not a surprise if you paid attention.

o1 and o3-mini both win on "evaluate this section of an econ paper. any errors of corrections needed?" Pasted in a few pages of a real paper with a mistake I found. No AI I'd tried found it - o1 and o3-mini did; R1, Gemini Thinking, Claude 3.5, 4o, o1-preview failed. Impressive!