Anthony Rinaldi

AI Researcher | MSc @UofTCompSci

[Anthony Rinaldi @anthonyprinaldi]

Prediction error isn't enough. A model can be locally coherent and globally blind. On frame revision, metacognition, and what AI still can't do.

[https://t.co/ZJDUWoy9OP]

[Anthony Rinaldi @anthonyprinaldi]

[5/5]

Full writeup with all 4 experiments (latent structure, degradation, surprise under artifact injection, action counterfactual).

Post: https://t.co/rPtfHb1o3w

Code: https://t.co/a8uADnu5jr

Always open to collaborate on world models / embodied AI — reach out!

[Anthony Rinaldi @anthonyprinaldi]

[1/5]

I ran 4 diagnostic probes on LeWM - https://t.co/6XBCSUnbeL

Read the thread to find out how far you can trust this JEPA predictor for planning.

🧵...

[Anthony Rinaldi @anthonyprinaldi]

[4/5]

Ran an action counterfactual: same rollout, three conditions - real actions, shuffled actions, all zeros.

Real actions give ~30-40% lower error at steps 1–5. Gap closes by step 10.

Predictor conditions on actions, but only in a short window. Beyond that it ignores them.

[Anthony Rinaldi @anthonyprinaldi]

A more elegant, possibly hybrid-symbolic, approach would not need this scaling trickery to solve the problem. It makes you wonder if all the engineering tricks used in modern LLMs are really just forcing an unsuitable solution to work, instead of finding better approaches

[Anthony Rinaldi @anthonyprinaldi]

Reproducing Marcus' experiment from The Algebraic Mind on multi-nodes-per-variable NN inability to generalize UQOTOM functions

Without seeing training data with a 1 in the last digit of the array, the model quickly learns to force the last digit to be 0 ...

[Anthony Rinaldi @anthonyprinaldi]

It turns out we can! Without changing anything except the scaling of the input data, we can improve the validation results of the final digit. This solution is not elegant and does not solve the underlying problem Marcus is pointing at; backprop being local ...

[Anthony Rinaldi @anthonyprinaldi]

@AlexanderNaume2 Not symmetrical by a diagonal line!

[Anthony Rinaldi @anthonyprinaldi]

@GaryMarcus Which systems can we rely on for solving these abstract (and hierarchical) problems? Is semi-symbolic the way?

[Dan Roy @roydanroy]

Fascinating in the context of LLMs and the debate about the language of thought.

[Pedro Domingos @pmddomingos]

How many trillions of dollars do we need to spend to conclude that LLMs won't get us to AGI?

[Anthony Rinaldi @anthonyprinaldi]

@xlr8harder AI is smarter than us!

[Anthony Rinaldi @anthonyprinaldi]

Wonder what AI philosophers think of this

[Kart ographien @kartographien]

“When ChatGPT hallucinates, it tells us something more fundamental than the truth — namely, the distribution from which the truth itself is sampled.”

[Anthony Rinaldi @anthonyprinaldi]

Great example showing AGI isn’t that close. A simple finite-state machine could get this task right.

[Anthony Rinaldi @anthonyprinaldi]

Don't forget: narrow generalization

[Pedro Domingos @pmddomingos]

Transformers: deep learning, shallow understanding.

[Anthony Rinaldi @anthonyprinaldi]

@pmddomingos Wouldn't it be their whole "conscious" life, since we may not remember things from adolescence? Or do you believe those early histories still affect us subconsciously?

[Tal Linzen @tallinzen]

What's the point of following all of the twists and turns of the OpenAI drama in real time? Let's go back to tweeting about science and wait for the recap in three days after the dust settles