A thread about proofs described as "AI is intractable / AGI can't work". That 2024 paper and similar ones are worst-case analyses. That is, there are distributions that no machine learning technique can learn effectively. Yes, computers cannot learn everything. But what does that show?

Mike Guerzhoy@guerzhoy.bsky.social · 4mo ago

To appear in Computational Brain & Behavior soon: the claimed 2024 proof (also in CBB) that AGI via learning is intractable also "proves" that ImageNet is intractable. My reading of the hole: equivocation on the variable D. Preprint here: arxiv.org/abs/2411.06498

To appear in Computational Brain & Behavior soon: the claimed 2024 proof (also in CBB) that AGI via learning is intractable also "proves" that ImageNet is intractable. My reading of the hole: equivocation on the variable D. Preprint here: arxiv.org/abs/2411.06498

Barriers to Complexity-Theoretic Proofs that "AGI" Using Machine Learning is Impossible

A recent paper (van Rooij et al. 2024) claims to have proved that achieving human-like intelligence using learning from data is intractable in a complexity-theoretic sense. We point out that the proof...

arxiv.org

I entertain myself with Fable Max queries and hipster coffee shop coffee. The dollar per minute entertainment ratio favours Fable but not by a lot

Back from the dentist! (1) Insurance auto-approves. Bro do you even logistic regression? (2) How does he know what instrument to use? Serious world-modelling there (3) I waived the x-ray explanation; can't check (4) Some anaesthetic, more only if patient asks: that one I get

Area man assigns homework, surprised range-restricted homework doesn't predict exam. (I don't doubt there was ChatGPT'ing with the free ChatGPT, and the R^2 is way lower than I'd guess even after accounting for ChatGGPT (see plot), but the point stands)

Kieran Healy@kjhealy.co · last mo.

lol

I've been promoted, get ready for some hot takes

Two fundamental problems (1) humanities people thinking it's normal to give a talk from a script (2) the guy didn't come back with "you're absolutely right to push back on this"

Crystal Fleming クリスタル@alwaystheself.bsky.social · 2mo ago

I'm attending a digital humanities event in Montreal. There was a keynote on AI and something about the talk made me wonder if it had been written by Claude. I said as much in the Q&A. As I posed the question, the speaker shifted, looking slightly uncomfortable. What he said next shocked the room +

When I retire, I expect UofT Bookstore to retire the generic collegiate font UofT ENGINEERING t-shirt

To appear in Computational Brain & Behavior soon: the claimed 2024 proof (also in CBB) that AGI via learning is intractable also "proves" that ImageNet is intractable. My reading of the hole: equivocation on the variable D. Preprint here: arxiv.org/abs/2411.06498

Barriers to Complexity-Theoretic Proofs that "AGI" Using Machine Learning is Impossible

A recent paper (van Rooij et al. 2024) claims to have proved that achieving human-like intelligence using learning from data is intractable in a complexity-theoretic sense. We point out that the proof...

arxiv.org

"Please don't feed the utility monsters"

Some personal news: I solved the Wordle in 3 guesses before my Friday afternoon ESC180 class. While you're here, the final version of our Findings of the Association for Computational Linguistics: IJCNLP-AACL 2025 paper on biases in human Wordle gameplay is out

New from us at ICCV workshops this weekend: www.cs.toronto.edu/~guerzhoy/Op... www.cs.toronto.edu/~guerzhoy/EC... Stay tuned for NeurIPS stuff :-)

I was cited in Psychology Today! Specifically, implying the opposite of what I actually wrote, by a person arguing that that's the kind of thing LLMs would do. www.psychologytoday.com/ca/blog/how-... Preprint: arxiv.org/abs/2411.06498

At the Statistical Society of Canada Annual Meeting in Saskatoon (my first time in the Praries)

Happy belated Valentine's day! New one is in orange

Low Cost Airline (FLAIR) captain at the door, pointing at my NeurIPS/etc posters: what are those, your plans to take over the world? Me: Yes and the amazing thing is you didn't charge for those as extra carry-on (OK fine I was two minutes late with this and didn't come back up to the plane)

Geoff Hinton was awarded the Nobel Prize in Physics. I wasn't but I did score an 87 in his course in 2005, so I got to be on (Canadian) national TV talking about it. youtube.com/watch?v=rQxv... (CTV) youtube.com/watch?v=6yDz... (CityTV) globalnews.ca/video/109117... (Global)

This Sunday, I went down to Chestnut Residence to play Ping Pong against some 17 @uoft.bsky.social EngSci students who challenged me to a game! For every point difference against me in a game I lost I donated $1 to the Toronto Humane Society (rounded up a bit to $42 + credit card fees).

I miss the times when I could just post to Facebook and get engagement. New socials suck: LinkedIn: Easy engagement, but tacky Instagram: Easy engagement, but don't understand it Bsky: People too nice Twitter: People too mean Openreview/EasyChair: People *way* too mean

Chemin de l'Église on a warm autumn night — in Montbonnot-Saint-Martin, France., Nov 24, 2012