Saad

Machine Learning Engineer

[Saad @saad_im5]

@raila_san You might find this useful while traveling in Seoul.
https://t.co/Nbqikk5buA

[Saad @saad_im5]

If you're in Seoul for ICML, check out https://t.co/uVACyKzpVE
Ask it anything about Korea, from restaurant menus and directions to nearby places. It searches Korean sources in real time and gives you answers in seconds.
Free to use.

[Saad @saad_im5]

@icmlconf

[Saad @saad_im5]

If you're in Seoul for ICML, check out https://t.co/uVACyKzpVE
Ask it anything about Korea, from restaurant menus and directions to nearby places. It searches Korean sources in real time and gives you answers in seconds.
Free to use.

[Saad @saad_im5]

@jeongminby98858 Try this too

https://t.co/Nbqikk5buA

[Saad @saad_im5]

If you're in Seoul for ICML, check out https://t.co/uVACyKzpVE
Ask it anything about Korea, from restaurant menus and directions to nearby places. It searches Korean sources in real time and gives you answers in seconds.
Free to use.

[Saad @saad_im5]

@ohmyksh Try this too

https://t.co/Nbqikk5buA

[Saad @saad_im5]

If you're in Seoul for ICML, check out https://t.co/uVACyKzpVE
Ask it anything about Korea, from restaurant menus and directions to nearby places. It searches Korean sources in real time and gives you answers in seconds.
Free to use.

[Saad @saad_im5]

Claude Fable 5 got it right.

[Saad @saad_im5]

@kevinkoosk Please check now

[Saad @saad_im5]

Just open-sourced Clawd Phone, Android app for agentic search on mobile.

Uses Claude API and runs tools locally on-device, can search files, read PDFs, fetch from web, and create or edit files.

https://t.co/heDCfJnh40

[Saad @saad_im5]

@OpenAI Still couldn’t get this right

[Saad @saad_im5]

@tom_doerr I'm trying something similar but as a native android app
https://t.co/HpH1eJh198

[Saad @saad_im5]

Just open-sourced Clawd Phone, Android app for agentic search on mobile.

Uses Claude API and runs tools locally on-device, can search files, read PDFs, fetch from web, and create or edit files.

https://t.co/heDCfJnh40

[Saad @saad_im5]

@ai_for_success Same goes for left hand

https://t.co/pSjQeWW5Gu

[Saad @saad_im5]

Just as image generation models struggle to depict a person drawing or writing with their left hand, LLMs also struggle to accurately identify the writing hand in images. They mostly default to 'right hand.'

Both Grok-3 and o3-mini-high got it wrong

[Saad @saad_im5]

@pk_iv How fast is it?

[Saad @saad_im5]

@jezell Btw Manus is built on top of browser-use

[Saad @saad_im5]

@kalomaze Still no luck with left hands

[Saad @saad_im5]

@jianxliao @maximelabonne @huggingface @ben_burtenshaw For multi-turn tool use, you can check verifiers https://t.co/dJsPt8yylF from @willccbb

[Saad @saad_im5]

Just as image generation models struggle to depict a person drawing or writing with their left hand, LLMs also struggle to accurately identify the writing hand in images. They mostly default to 'right hand.'

Both Grok-3 and o3-mini-high got it wrong

[Saad @saad_im5]

@willccbb How can GRPO be applied to multi-turn conversations instead of tool calling? Would it require simulating user responses during training?

[Saad @saad_im5]

@jdchawla29 @willccbb https://t.co/FxBxHbEKXk

[Tiancheng Zhao (Tony) @tianchezhao]

Introducing VLM-R1!

GRPO has helped DeepSeek R1 to learn reasoning. Can it also help VLMs perform stronger for general computer vision tasks?

Our preliminary answer is YES and it generalizes better than SFT.
https://t.co/iffweRXcpO

[anton @abacaj]

You can have a strong base LLM, do a bad job post-training (aligning, sft etc) and it will still be a generally strong LLM. But if you have a weak base LLM you can't save it in post-training... why do you think no one uses llama 2 anymore