Is the AI “Revolution” Overblown? | Robert Wright & Timothy B. Lee

[@Nonzero]

Click here to see an exclusive Overtime segment of this episode: nonzero.substack.com/p/is-the-ai-revolution-overblown-robert

[@therainman7777]

I believe Tim misinterpreted the result he referred to, where doubling the size of the ARC images caused performance to drop. This was not an indicator that the models have trouble with abstraction; rather it was evidence in quite the opposite direction: that the only reason current frontier models struggle on ARC at all is because their vision abilities lag behind their text abilities. We’ve long known that the models tend to be much smarter when dealing with text than with images, as vision has not yet developed as far as text. The point is to show increasing the size of the images simply taxes the models’ visual processing more, such that they will perform worse on the exact same reasoning task, if you give their vision system more that it must process. However. The underlying _reasoning task_ is the same regardless of the image size. IMO this shows that as soon as the frontier AI models’ vision catches up with their text processing abilities, ARC will likely no longer be a challenge for them.

[@rasen84]

No not vision. They are fed text formatted representations of the puzzles. These models have 128k token context window. They are supposed to see that far.

[@geaca3222]

@@rasen84 He's talking about vision as image processing compared to language processing, not images as symbols

[@rasen84]

@@geaca3222 right and I said arc is given as a text format they have 128k token context window.

[@geaca3222]

​@@rasen84 yeah when it's about tokens processing, then multimodal frontier models will get much better at imagery processing, so that it will be on par with language tokens processing, and the underlying reasoning capability will be even.

[@rasen84]

@@geaca3222 o3 is literally no more multimodal than 4o. It’s literally the same base model. The only change is extended rl based cot training. Multimodality isn’t why o3 solves arc.

[@cwcarson]

Hi Bob

[@tylermoore4429]

Google's new Titans paper already addresses some of the shortcomings mentioned by Tim.

[@cwcarson]

Tom Hanks' mother is Janet Marylyn Frager. Janet Marylyn Frager's son is John Travolta - chatgpt 15/01/2025

[@ryanmillerioux176]

I didn't realize Tom and John were brothers 🤔

[@donaldrobertson1808]

Perhaps print journalists should be known as text journalists going forward

[@jttj742]

I have a face made for radio, a voice made for the written word. And a brain for not writing anything ever.

[@therainman7777]

@@jttj742 😂

[@therainman7777]

@@jttj742 😂 Better Bachelor?

[@johnstewart7025]

Frankenstein: It seems as though a dumb universe made us and now we are making something like us. And, that something involves a trick where you juggle six balls at the same time.

[@jd35711]

must be rough working in the tech industry while named tim b lee

[@cwcarson]

He makes sure to print the B

[@jd35711]

@ so does tim berners-lee!

[@jangelbrich7056]

AI does neither _know_ nor _understand_ anything. It makes only statistically best guesses. Hence many results are wrong.

[@geaca3222]

I think it would be wise for humanity to just heed the warning of Geoffrey Hinton and proceed carefully with this immensely powerful technology. Enough said!

[@aleksandero7982]

When i hear dudes telling what's not going to happens and what's going to happen instead... i'm out!!!

[@vaughanmerrick]

I agree! Can’t wait for AI sex. It’s going to be so awesome! Maybe AI can make babies too so I can order one off Amazon and then I can have AI take care of the kid and educate it too! I don’t want to hear Debbie downers telling me that won’t happen. Don’t be a buzzkill ai skeptics!