Maximilian Kasy

microgpt Musings of a Computer Scientist.

A complete GPT — dataset, tokenizer, autograd engine, transformer architecture, Adam optimizer, training loop, inference — in 200 lines of pure Python, zero dependencies:
🔗 karpathy.github.io/2026/02/12/m...

A work of art by Andrej Karpathy (@karpathy.bsky.social) | h/t F.Geiecke

He has co-written a new policy brief about Universal Basic Income and Job Guarantees. The brief discusses the key theoretical issues as well as the latest results of his empirical work on these topics in Austria and Germany. Read the brief on the EfIP website: econfip.org/policy-brief...

The means of prediction and the production function of AI - Economics for Inclusive Prosperity Executive Summary Introduction How should we regulate AI? This debate is often dominated by arguments between “AI boomers” and “AI doomers” (Hao, 2025). AI boomers claim that building “artificial gene...

New on the Economics for Inclusive Prosperity website: A fascinating research brief by University of Oxford Professor Maximilian Kasy on why the true battle for future of AI is between different groups of people with different priorities—not humans vs machines.
#econfip
econfip.org/policy-brief...

Not Inevitable: Democratizing power over AI for public good vs private gain Podcast Episode · Economics for Inclusive Prosperity · February 25 · 43m

New episode! University of Oxford economist Max Kasy demystifies AI and explores ways to ensure that it's not exclusively controlled by tech titans. podcasts.apple.com/us/podcast/e...

YouTube video by Oxford Martin School Book talk - 'The Means of Prediction: How AI Really Works (and Who Benefits)' with Maximilian Kasy

Next was an excellent talk by @maxkasy.bsky.social on how the very human forces shaping the development of generative AI and who benefits from its current trajectory at @oxmartinschool.bsky.social www.youtube.com/watch?v=Lm9w... (5/8)

Book talk - 'The Means of Prediction: How AI Really Works (and Who… As economist Maximilian Kasy shows in The Means of Prediction , artificial intelligence, far from being an unstoppable force, is irrevocably shaped by…

Tomorrow 5pm at the Oxford Martin School!

www.oxfordmartin.ox.ac.uk/events/the-m...

Can A.I. Be Pro-Worker? As fears of mass unemployment grow, three leading economists advocate some policies to shift the focus from job displacement to job enhancement.

Can A.I. Be Pro-Worker? www.newyorker.com/news/the-fin...

Not inevitable: How to democratize power over AI away from the tech titans | Economics for Inclusive Prosperity We need an alternative vision of economics to replace failed neoliberalism and rising economic nationalism. In a world facing record inequality, climate peril, and rising illiberalism, how do we…

Just released: Efip podcast episode on how to democratize power over AI
inclusive-prosperity.simplecast.com/episodes/not...

(Other episodes with @rodrikdani @AtifRMian @S_Stantcheva: inclusive-prosperity.simplecast.com)

cc @econfip.bsky.social
1/2

The means of prediction and the production function of AI - Economics for Inclusive Prosperity Executive Summary Introduction How should we regulate AI? This debate is often dominated by arguments between “AI boomers” and “AI doomers” (Hao, 2025). AI boomers claim that building “artificial…

Efip policy brief on the same topic: econfip.org/policy-brief...

Which in turn summarizes key ideas from press.uchicago.edu/ucp/books/bo...

2/2

Not inevitable: How to democratize power over AI away from the tech titans | Economics for Inclusive Prosperity We need an alternative vision of economics to replace failed neoliberalism and rising economic nationalism. In a world facing record inequality, climate peril, and rising illiberalism, how do we…

Just released: Efip podcast episode on how to democratize power over AI
inclusive-prosperity.simplecast.com/episodes/not...

(Other episodes with @rodrikdani @AtifRMian @S_Stantcheva: inclusive-prosperity.simplecast.com)

cc @econfip.bsky.social
1/2

The robots who predict the future Three books unpack our infatuation with prediction, and what we lose when we outsource this task to machines.

Another nice review of "The Means of Prediction" in @technologyreview.com / @techreview:

www.technologyreview.com/2026/02/18/1...

(Also discussing lovely new books by @beenwrekt.bsky.social / @beenwrekt and @carissaveliz.bsky.social / @CarissaVeliz )

🔥 Just accepted for publication at AEJ:Policy 🔥

"Employing the unemployed of Marienthal:
Evaluation of a guaranteed job program"

maxkasy.github.io/home/files/p...

The robots who predict the future Three books unpack our infatuation with prediction, and what we lose when we outsource this task to machines.

Another nice review of "The Means of Prediction" in @technologyreview.com / @techreview:

www.technologyreview.com/2026/02/18/1...

(Also discussing lovely new books by @beenwrekt.bsky.social / @beenwrekt and @carissaveliz.bsky.social / @CarissaVeliz )

🔥 Just accepted for publication at AEJ:Policy 🔥

"Employing the unemployed of Marienthal:
Evaluation of a guaranteed job program"

maxkasy.github.io/home/files/p...

The robots who predict the future Three books unpack our infatuation with prediction, and what we lose when we outsource this task to machines.

Three books unpack our infatuation with prediction, and what we lose when we outsource this task to machines.

Worker activism in tech is back. It's starting with opposition to ICE, but I think it will spill over into other areas.

This thread will share what's happened recently and what I think might be next. As well as ways for folks to get involved if they want to.

“LLMs are not just labor-replacing and productivity-expanding, but can, in some circumstances at least, enable the production of totally new things.” @mattbruenig.bsky.social mattbruenig.com/2026/02/10/t... ht @mitchsaid.bsky.social

My slides for today's ML & Econ group meeting:

"Double descent,” and what linear regression can teach us about the coming collapse (?) of the AI bubble.

maxkasy.github.io/home/files/o...

(Based mainly on "Learning Theory from First Principles," chapter 12, www.di.ens.fr/%7Efbach/ltf... )

The Means of Prediction Check out this great listen on Audible.com. This is an audiobook version of this book. An eye-opening examination of how power—not technology—will define life with AI. AI is inescapable, from its...

📕 🔊 "The means of prediction" is now available as audiobook! 🔊 📕

www.audible.co.uk/pd/B0GL99H5J...

I wanted David Attenborough. didn’t work out, but still got a great voice actor who beats my midnight-jazz-radio-dj voice :)

Title: Machine learning, causal inference, and economics Speaker: Professor Maximilian Kasy (University of Oxford) Date: 27th Jan 2026 - 9:30 to 10:30 🗓️ Event: (CIFW05) Causal Machine Learning for… Prof. Maximilian Kasy | Machine learning, causal inference, and economics

Recording of my talk on
"Machine learning, causal inference, and economics"

www.youtube.com/watch?v=Qufn...

Very nice review of "The means of prediction" by @B_Eichengreen:

smartthinkingbooks.com/smart-thinki...

The means of prediction and the production function of AI
maxkasy.github.io/home/files/p...

Thoughts and comments welcome!

2/2

Now with correct links - forthcoming policy-papers and essays:

Individual property rights or collective democratic governance? Privacy in the age of AI
maxkasy.github.io/home/files/p...

Welfare for the 21st century: Basic income and job guarantee policies
maxkasy.github.io/home/files/p...

1/2

The means of prediction and the production function of AI
maxkasy.github.io/home/files/p...

Thoughts and comments welcome!

2/2

Now with correct links - forthcoming policy-papers and essays:

Individual property rights or collective democratic governance? Privacy in the age of AI
maxkasy.github.io/home/files/p...

Welfare for the 21st century: Basic income and job guarantee policies
maxkasy.github.io/home/files/p...

1/2

oops -
posting correct version momentarily, thanks!

(“one brain region for one character trait”). Hard to see a priori why that should be true for trained neural nets. Moreover, not clear why mechanisms should generalize between different nets.

I remain skeptical.
MI seems predicated on finding underlying computations that are neatly separable, modular or sparse, similar to the (failed) Mendelian paradigm in genetics (“one gene producing one observable trait”) or 19th century craniology

Mechanistic Interpretability for AI Safety -- A Review Understanding AI systems' inner workings is critical for ensuring value alignment and safety. This review explores mechanistic interpretability: reverse engineering the computational mechanisms and…

What is “mechanistic interpretability” (MI) of neural networks?
This paper gives a great introduction:
arxiv.org/abs/2404.14082

MI empirically explores “how neural networks work,” in a spirit similar to neuroscience - without the ethical constraints on manipulating living brains.