weakly typed

learning {ML, PL, maths} // CS pre-grad // DMs open :)

[weakly typed @weakly_typed]

@Adellama_ @trekkinglemon i perhaps should unpin this tweet
https://t.co/6tBGo8QGA3

[weakly typed @weakly_typed]

@trekkinglemon while this is an impressive demonstration of the capabilities of large language models to synthesise natural-language problem statements into formal / executable versions, we're still a long way off from 'true' system 2 mathematical reasoning

(1/3)

[Leo Gao @nabla_theta]

"it's crazy how there can exist something so incredibly intelligent and yet hobbled by maladaptive post-training", remarked the burnt-out bay area ai researcher,

[Mishka @mishapathy]

There's a really great joke about nonconstructive proofs.

[weakly typed @weakly_typed]

@LydNot *looks at huggingface*

[Kiran @kirancodes]

literally have none of these people ever read a CS textbook or what? like the lambda calculus is right there???

[Transluce @TransluceAI]

How could we train an AI model that was very good at overseeing another model: catching reward hacking and sandbagging, predicting unwanted behaviors or fine-tuning effects, etc.?

We propose a new approach for doing this at scale: oversight foundation models.

[weakly typed @weakly_typed]

@celestepoasts meeeeeeetamodels

[Talia Ringer 🕊🪬 @TaliaRinger]

I guess past-me is included in "only a tiny fraction of exceptionally brilliant PhD students," which is flattering and all, but I think one thing to mention here is that this you can hedge exploratory/visionary work by also having an immediately applicable side project

[gavin leech (Non-Reasoning) @gleech]

excellent framing from Dan Frank on themes that @patio11 @jachiam0 etc have been covering:

[weakly typed @weakly_typed]

@celestepoasts @viemccoy should i like just turn the lightning talk into a blogpost

[Celeste @celestepoasts]

it feels correct to say: I'm scared of repeating the mistakes of SAEs.

META MODELS SHOULD BE ONE BET AMONG MANY!!!

I got very excited about them because it felt like a good enough bet to have O(100) people on it, but was only O(10)

I think it should be more. But we are really talent constrained, lets not privilege this hypothesis too much!!!

ONE BET AMONG MANY. THANK YOU FOR YOUR ATTENTION TO THIS MATTER!

please dm if u wanna 1-1 on my last day in Seoul!

[weakly typed @weakly_typed]

@justinrgraham of quotient, cartesian product and mappend, one has the correct semantics

[Celeste @celestepoasts]

For all those excited about

Activation Oracles, LoRAcles, Natural Language Autoencoders, Introspection Adapters...

Come meet the authors at the scalable interp and meta models mixer!

[Eric Neyman @ericneyman]

Her copper coat now weathered by the rains,
By dreams deferred and promises half-kept,
Above the isle where weary millions stepped,
The mighty woman with the torch remains.
She cannot sway those weighty sunset gates
That barred the way and forced St. Louis on,
Yet parted for the ships that fled Saigon,
And now draw shut again. Undimmed, she waits.
“I want them all,” with silent lips cries she:
“The stranded allies, scarred and battle-sore;
The children borne through deserts; those who flee
From hunger, hatred, tyranny and war.
And, ready when once more the gates swing free,
My lamp still burns beside the golden door.”

[Asher @asher5772]

I built an open-source repo for efficient distributed training of NLAs. On a 4xH100, it reaches 70% FVE in ~3 hours on Qwen-3-8B. Feel free to reach out about this repo or anything else NLA-related—including project ideas, there's much to be done!
https://t.co/mVOyBBERxd

[weakly typed @weakly_typed]

@celestepoasts @juli_li_ :uwuracles:

[Vincent @vvvincent_c]

what does @TransluceAI know....

[Liv @livgorton]

I joined anthropic a ~month ago and have written ~no code myself. I went from getting quite frustrated with coding agents even 6 months ago and giving up and writing some of the code myself to a big part of my role now being agent management.

[weakly typed @weakly_typed]

@celestepoasts and then she failed to release loracles on arxiv