Ahmed

ml scientist @Cision. ms @GeorgiaTech.

[Ahmed @KickItLikeShika]

We have released a technical report (+ a qwen checkpoint) for the reproduction of Self-Distillation Fine-Tuning (SDFT), check it out here https://t.co/o1MmWQ67t8

[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

[Ahmed @KickItLikeShika]

pretty sure tons of these vibe-checks were part of the the training set, wanna see how it does it in a very tense research environment

[Tib3rius @0xTib3rius]

My god, its incredible.

[Ahmed @KickItLikeShika]

the fuck are we about to witness

[Cursor @cursor_ai]

Claude Fable 5 is now available in Cursor.

It sets a new state of the art on CursorBench at 72.9%, 8 points above the previous best.

[Haitham Bou Ammar @hbouammar]

I have so much fun writing this position with some of the most amaaazing people in robotics!

Have a look at it here: https://t.co/zM3NBtobkx
#AI #MachineLearning #Robotics

[Ahmed @KickItLikeShika]

llms can break a bunch of things and try and fix them and break other things etc

[Ahmed @KickItLikeShika]

quantifying this with number of lines isn’t something i can trust tbh

[Alex Albert @alexalbert__]

We just published internal data on how much of Claude's development is already being done by Claude:
- Over 80% of all code merged into our codebase is now written by Claude
- It's been months since many researchers at Anthropic hand-wrote code
- The typical Anthropic engineer ships 8x as much code as they did in 2024
- On the most open-ended engineering tasks, Claude's success rate jumped from ~26% to 76% in 6 months
- When research sessions went off-track, Claude proposed a better next step than the human took 64% of the time

We're not at recursive self-improvement yet, but it could come sooner than most expect. I highly recommend reading the full blog post.

[Ahmed @KickItLikeShika]

llms scale fried my brain so bad that seeing JEPA's performance with only a few million parameters makes me doubt my life choices

[llm_enjoyer @LLMenjoyer]

u js trained on test bro, it's not that deep

[Dave Banerjee @DaveRBanerjee]

lol

[Ahmed @KickItLikeShika]

@hbouammar @MarcHoelle our experiments were based on the codebase released by the authors, which is fully built on TRL https://t.co/tcBrj8ZlUF
i believe it's been copied over to trl library later after release, so you should see similar results if you try trl

[Ahmed @KickItLikeShika]

@md_kasif_uddin qwen is clear

[Ahmed @KickItLikeShika]

@aamixsh this was one of the main motivations for publishing this report, you can find the detailed evaluation results in the shared HF model repo!

[Aayush Mishra @aamixsh]

Couldn’t reproduce the “continual learning” claim of the original paper. Their released checkpoint showed drastic performance drops in prior tasks. No updated releases for months now.

Great claims should come with great reproducibility.

Happy to see this effort; hope it works!

[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

[Ahmed @KickItLikeShika]

why would they call rlm dynamic workflows

[ClaudeDevs @ClaudeDevs]

New in Claude Code (research preview): dynamic workflows.

Claude writes an orchestration script on the fly, then spins up a large fleet of coordinated subagents in parallel to take on your most complex tasks.

Use the word "workflow" in a prompt to get started.

[Ahmed @KickItLikeShika]

placed my bet on $mu months ago, went in with 60% of my portfolio weeks ago

[The Kobeissi Letter @KobeissiLetter]

BREAKING: Micron stock, $MU, officially hits $1 trillion in market cap for the first time in history.

12 months ago, this stock was worth just $70 billion.

[Ahmed @KickItLikeShika]

burning some gpus

[HandofArsenal @HandofArsenal]

Greatest picture I have laid eyes on

[Ahmed @KickItLikeShika]

@20alaamahmoud دايمًا معترف

[Arsenal @Arsenal]

Good morning from your Premier League champions 👋