aidaritoo

Born
Live...

[aidaritoo @aidaritoo]

@MotoniqLabs Wow, very interesting boost in performance!

[Motoniq @MotoniqLabs]

Not one π0 weight changed.

Precision fitting on a real robot: 0/10 → 9/10.

Early results.

Part 1 of 3: more capability from the models you already have. 🧵

[aidaritoo @aidaritoo]

@eppy_ephard @ggerganov @Alibaba_Qwen Wow! Llm on phones are awesome 👍

[Eppy @eppy_ephard]

Qwen3-4B on a 2019 phone, fully offline: 1.34× faster replies (2.2× on long outputs) from llama.cpp settings alone, with identical outputs.

@ggerganov @Alibaba_Qwen

[aidaritoo @aidaritoo]

@dankrad I think mathematicians need to find a criteria (smth similar to NP hard problems in CS) for math problems which unlikely can be solved by LLMs. Then, they will have a choice - to work on them and not being scooped with AI or take a risk to be scooped

[aidaritoo @aidaritoo]

@karthik_r_n Karthik, let me ask you, how did you get Princeton faculty position? Because you loved CS more than others OR because you had a great results in CS field? In CVs and resume I haev never seen "I love the game/subject/etc" But I saw results, published papers etc.

[aidaritoo @aidaritoo]

@DimitrisPapail @dejavucoder you dont get scinetific awards and univeristy positions for asking questions, don't you? You get them by showing results. Well, AI models show the results in mathematics.

[aidaritoo @aidaritoo]

@tonylfeng Well, sometimes very "primitive and naive" characteristics could actually work. I think mathematicians like yourself should keep in mind that the next token predictors solved 100+ open math problems already. So, primitive and naive not necessarily means bad.

[Haitham Bou Ammar @hbouammar]

Continual learning is a dear problem to me that i worked on as early as 2013. ❤️

In our new work, we provide a nice formalisation of the stability vs plasticity trade-off in terms of solutions path in the probability space between student and teacher distribution.

This work is led by the amazing Ahmed K Khamis @KickItLikeShika a must-follow researcher!

Check it out 👇 ❤️

Paper, Code and Checkpoints can be found here: https://t.co/t7kjw4fr2N

#AI

[aidaritoo @aidaritoo]

@CodexZeroTR Wtf is that with lightings????

[Haitham Bou Ammar @hbouammar]

Self-distillation has been an awesome idea! However, since the amazing paper was out: (https://t.co/nB4IX9duXZ), there have been many difficulties regenerating the results and many issues put out!

I am happy to tell you that the problem is not the idea, we have been able to reproduce the results on Tool Usage!

To help the community push this forward, we wrote a small technical report which provides the way to reproduce those results.

Not only that, we have also open-sourced HF checkpoints for the fine-tuned models for you to use.

We still think self-distillation is amazing ♥️♥️

Report: https://t.co/jrdynIs1E7
Checkpoints: https://t.co/PjKXpCaBda

Happy weekend! You can take it from there!!

#AI #MachineLearning

[Haitham Bou Ammar @hbouammar]

Hey Beautiful Amigos,

I have some ideas to explore for EAI - especially around world models and some cool extensions to VLAs :D

Come join us to work together! Ping me a message if you are versed with ML and Python and EAI !!

#AI #MachineLearningTrivia

[Haitham Bou Ammar @hbouammar]

I just wrote a Medium blog about our work: λ-RLM. 😄

The core idea:

LLMs don’t need infinite context.
They need better recursion.

Instead of asking the model to invent messy recursive code loops, λ-RLM gives it a typed runtime:

SPLIT → MAP → FILTER → REDUCE

The model reasons.
The runtime recurses.

In the blog, I explain why long-context reasoning is not only a memory problem! It is a control-flow problem.

This is why we call λ-RLM:

The Y-Combinator for LLMs.

Blog: https://t.co/0WwzF5jXXx

[aidaritoo @aidaritoo]

@hbouammar Great tool, btw!

[Haitham Bou Ammar @hbouammar]

super happy that lambda rlms are slowly growing!! 😆😆

give you llm some structure and monitor control!!

check it out the code here: https://t.co/sBm9DlsJbd

#ai #machinelearning

[Haitham Bou Ammar @hbouammar]

I made a small website for λ-RLM.

The idea: Recursive LLMs may not need to invent their own control flow.

λ-RLM gives them a typed functional runtime instead!

😱The LLM solves bounded leaf problems.
😱The runtime handles recursion.
😱Less chaotic agents.
😱More inspectable reasoning.

🌐 https://t.co/43jBoRFRh2
⭐ https://t.co/y8BL0nrxv5

#AI #MachineLearning #ICLR2026

[Han Xiao @hxiao]

bro goes casual mode at ICLR

[Haitham Bou Ammar @hbouammar]

I just finished reading the Harness blog of @TheodoreGalanos on Lambda-RLMs, and it is simply amazing and worth everyone's read!

1⃣ Walks us through RLMs from the amazing @lateinteraction's team
2⃣ Walks us through Lambda-RLMs from mine
3⃣ Independently runs fantastic experiments and reports clear results!

I am happy Lambda-RLMs did well!!

Blog: https://t.co/lieB2wQKYT
Code for Lambda-RLMs: https://t.co/y8BL0nrxv5
Paper for Lambda-RLMs: https://t.co/h1ha6HYBOx

Simply amazing! 😍

#AI #MachineLearning #RLMs #ICLR2026

[aidaritoo @aidaritoo]

@pmddomingos "out of idealism, not self-interest" - the level of dellusion is unprecedented.

[aidaritoo @aidaritoo]

@yoavgo Interesting way of thinking. Lets play with it a bit: stealing from Apple store definitely has small impact on Apple tiny if it exists at all, as you said). Does it make it a good/acceptable practice?