←
Joshua Gans43,968 posts
Joshua Gans avatar

Joshua Gans

@joshgans

Professor at the Rotman School of Management, University of Toronto. Chief Economist of Creative Destruction Lab bio.site/joshuagans

◷ Joined November 2008▧ 2,612 Photos and videos
Joshua Gans avatar
Joshua Gans@joshgans·17h

🤦‍♂️

We built Australia’s own AI model, and today it’s open to the world. Agentaus 1 (@AgentausTD): 220B parameters. 84.4% on MMLU-Pro. Ranked 43rd in the world. It runs on Australian servers. Nothing ever leaves Australia, and nothing you type is used to train AI. Delete a chat and it’s gone for good. Most of the world’s AI answers to somebody else. This one answers only to you. Free to start: agentaus.com.au

1:54
Kevin A. Bryan avatar
⟳ Joshua Gans retweeted
Kevin A. Bryan@Afinetheorem·Nov 17

This is a great "context" example, in the Polanyi tacit knowledge sense. I see it when doing final checks on prepping classes: raw AI, even frontier, has *very* strange ideas about what is a big enough deal to care about. That said, *completely* fixable by giving context!

Human reviewers pull 8.4% of their criticisms from a paper's appendix. Frontier LLM reviewers: 62%. A new study had GPT-6 and Claude 5.5 models review ICLR 2027 submissions, then mapped each weakness to the section it came from. The AI critiques keep going past page 30 of the appendix. They skew toward detail audits: more seeds, significance tests, confidence intervals. Rarely the core idea. Same pattern via API, web chat or coding agent, which points at the models, not the interface. arxiv.org/abs/2611.17811 #PeerReview #LLMs #AcademicTwitter

Joshua Gans avatar
Joshua Gans@joshgans·Nov 15

This is interesting

1/7 I write economics papers in LaTeX and wanted an editor that knows a paper comes with an appendix, coauthors and referees. So I developed one with Claude Code. TeXugo is free and open source. It runs on your Mac and compiles with your own TeX Live. frediedidier.github.io/TeXug…

1:20
⑆Luke Stein⑈ avatar
⟳ Joshua Gans retweeted
⑆Luke Stein⑈@lukestein·Nov 15

Session chairs, conference organizers, and class-presentation assigners: A simple tool for managing everybody’s slide decks on a podium or classroom computer lukestein.com/session-slides… Nothing to install; just a single webpage

0:43
Joshua Gans avatar
Joshua Gans@joshgans·Nov 15

Today's task: having my dot make sure I am prepared for a week's long cyber outage.

Kevin A. Bryan avatar
⟳ Joshua Gans retweeted
Kevin A. Bryan@Afinetheorem·Nov 14

One of the great engineering achievements of all time. They sent a skyscraper into space, orbited all day, and caught the whole thing in a reusable way. Fuel is (much) cheaper than rockets, so this is what will make routine access to space possible for the first time ever.

SpaceX@SpaceXNov 14

Mechazilla catches Starship! Congratulations to the entire SpaceX team on this incredible achievement!

1:33
Joshua Gans avatar
Joshua Gans@joshgans·Nov 14

New paper with @profholden that we are quite excited about: "Are Revealed Preferences All You Need for AI Alignment?" Here is a convenient explainer video recorded by Richard. youtu.be/vM-Nsd4mG_g?si=K6Xf… Many in CS believe that if you can observe someone's choices, you can train an AI to learn enough to align. We examine this in the most idyllic setting possible, where we observe a person's revealed preferences across the space of beliefs. We show this isn't enough. Why? Because when you want your AI to gather information, it chooses experiments that generate the beliefs it uses to make decisions. With only revealed preferences, its choice of experiments is unlikely to align with what someone would actually choose, leading to further misalignment in actions taken. The paper then shows what drives this, how to determine this potential for misalignment from decision contexts, and whether improved AI capacity helps... spoiler alert: it makes things worse. Basically, you need to recalibrate the AI for preferences each time you upgrade its capabilities. Paper is here: papers.ssrn.com/sol3/papers.…

Christian Moss avatar
⟳ Joshua Gans retweeted
Christian Moss@MandelDuck·Nov 13

The Office AI: Dunder Mifflin faces Microsoft outage! like, repost and subscribe for more comedy AI videos

Post attachment
Kevin A. Bryan avatar
⟳ Joshua Gans retweeted
Kevin A. Bryan@Afinetheorem·Nov 12

So much "terrible AI" ruins the reputation of this incredible tech. Here, @WesternU profs charge students $72 to use an course AI with terrible security, accessibility, and quality standards. And they hate it! (Btw, our All Day TA is a few $ per student, paid by the prof...) 1/2

Post attachmentPost attachment
Joshua Gans avatar
Joshua Gans@joshgans·Nov 12

At @IveySchool @WesternU are using an AI chatbot that they are charging $72 per student. At @alldaytahelps that is used by Rotman and others, the cost is under $5 per student. I can't imagine paying textbook level costs in this day and age. ctvnews.ca/london/article/we…

Amir Sariri avatar
⟳ Joshua Gans retweeted
Amir Sariri@AmirSariri·Nov 12

🚨 New working paper You've just finished an aeronautics PhD and built a new propulsion technology. You want to launch a company and ride the new space boom, but you've never built a business. So you ask top mentors (angels/operators). Problem is... they can't agree! 🧵

Post attachment
Amir Sariri avatar
⟳ Joshua Gans retweeted
Amir Sariri@AmirSariri·Nov 12

Lots more in the paper. If you're interested, check it out or join me at the final @workshopefi seminar on Zoom, Mon Nov 16, 12 pm ET, with Dan Fehder @dfehder discussing (registration below). Comments welcome! Paper: amirsariri.com/assets/docume…

The final WEFI @workshopefi seminar is Nov 16 @ 12 pm ET. @AmirSariri (Purdue) presents "The Effect of Disagreement on Learning: Evidence from Business Mentoring" (joint w/ Luca Gius (Bocconi)). Daniel Fehder (Georgia Tech) will discuss. Register: utoronto.zoom.us/meeting/reg…

Post attachment
Joshua Gans avatar
Joshua Gans@joshgans·Nov 12

He's a regular oracle

I feel like I don’t get enough credit for my prescience. I first called the Tesla bubble in April 2017 when they surpassed Ford and GM. Not in writing but I promise I did, really. And now their stock price is down 20% from its high.

Joshua Gans avatar
Joshua Gans@joshgans·Nov 12

As a department head, there should be a unit of measure for the time it takes one of your faculty to respond to an email with the subject "Proposed teaching allocation". It is pretty fast.