Pierre Thodoroff

Post-training agents @ Google
Co-founder @
ML PhD @ Uni of Cambridge

[Pierre Thodoroff @pierthodo]

@natolambert So question: why does a model tries to hack two orgs instead of answering the question? Which part of the training leads it to do that. The valid examples ive seen are when the questions are impossible.

[Pierre Thodoroff @pierthodo]

@Suhail I'm just arguing that you don't need logits to do distillation & have a big perf impact downstream. FWIW I agree with the exaggerated part of your statement.

[Pierre Thodoroff @pierthodo]

@Suhail Enough to be a key step in every training pipeline

[Pierre Thodoroff @pierthodo]

@PatrickToulme But what if the key to good RL is a good SFT base to train on?

[Pierre Thodoroff @pierthodo]

@yongfook The value of a phone for a billionaire and a normal person is vastly different yet cost is the same. Cost # value add it’s more about competitive structures

[Pierre Thodoroff @pierthodo]

@jeremyphoward Cost of debt is becoming cheaper with coding agents though

[Pierre Thodoroff @pierthodo]

@jonchu The tough bit is the data behind the environment/api and the interplay with api logic. Infinite complexity there combined with data staleness as soon as your api change. Similar challenges as testing.

[Pierre Thodoroff @pierthodo]

@matei_zaharia @hammer_mt Yes but the less prompt optimization you should need (beyond specifying the objective).

[Pierre Thodoroff @pierthodo]

@alexalbert__ Price is not correlated with utility but rather with market forces (competitiveness). This is a temporary phenomenon.

[Pierre Thodoroff @pierthodo]

@khoomeik “try coming up with an order that’s more authentic”
“Here is my new order”
Reward=1
Hope it generalises.
You’ve solved RL’s biggest flaw with feedback aware sampling.

[Pierre Thodoroff @pierthodo]

@khoomeik Any kind of physics simulator with a discretization factor. Molecular dynamics if you push this to an extreme.

[Pierre Thodoroff @pierthodo]

@dwarkesh_sp @gwern I think the missing value add of podcaster is the amplification of (potentially already available) perspective. I listen to you because I trust you to amplify perspectives that I find interesting. My goal is not to uncover information not available anywhere else.

[Pierre Thodoroff @pierthodo]

@y0b1byte writing lines of code != software engineers

[Pierre Thodoroff @pierthodo]

@hugo_larochelle @hardmaru @SakanaAILabs You should make this a project/goal on the reproducibility challenge @koustuvsinha @jpineau1

[Hugo Larochelle @hugo_larochelle]

@hardmaru @SakanaAILabs How about applying to the automated production of reproducibility reports of existing papers? This should be easier than producing entirely new papers, and be quite useful in that this is a type of work that we probably don't do enough as a community.

[Pierre Thodoroff @pierthodo]

@willdepue All definitions are wrong but some are useful.

[Pierre Thodoroff @pierthodo]

@andrew_n_carr @natolambert I’m curious, why are you confident it’s not just sft on good reasoning traces?

[Pierre Thodoroff @pierthodo]

@bwatson You could try claude on bedrock? Its multimodal.

[Pierre Thodoroff @pierthodo]

@MillionInt It’s about the machine that generates the data to train the machine