Amine Ben Khalifa

Founder @attaapp. ML PhD. Building AI products with real models, not wrappers. Previously worked on FaceID at Apple.

[Amine Ben Khalifa @aminert]

@fchollet is this intentional? Was hoping to read your take on recent developments

[Amine Ben Khalifa @aminert]

@_arohan_ Not frontier haha

[Amine Ben Khalifa @aminert]

@ClementDelangue @spyced How will we trust them moving forwardeven if they correct course?

[Nathan Lambert @natolambert]

To me this paints Anthropic clearly as anti science, and therefore anti progress and anti safety.

[Amine Ben Khalifa @aminert]

I've never been this angry about getting nerfed by big labs

[Nathan Lambert @natolambert]

I got a good nights sleep and I’m still just as angry about Anthropic’s choices.

I enjoy working in AI so much and to have my access to the cutting edge models for my work rugpulled in an under the table fashion is appalling.

I expected to be restricted eventually, but not now, and to be told it directly.

[Xin Eric Wang @xwang_lk]

If you really think about it, despite being mocked as “ClosedAI,” OpenAI has contributed enormously to the field: GPT, GPT-2, GPT-3, CLIP, the ChatGPT paper, the GPT-4 Technical Report, the Sora technical blog, and even open-sourced Codex.

Anthropic, meanwhile, has contributed far less to the public research ecosystem while increasingly promoting fear-based narratives and restricting access through heavy gatekeeping.

The world I least want to live in is one where the future of AI is controlled by companies that prioritize secrecy, gated access, and centralized control over openness, reproducibility, and scientific progress.

[Bojan Tunguz @tunguz]

Anthropic really takes us for idiots.

[Nathan Lambert @natolambert]

A message to Anthropic leadership: You're not special. Making sure AI goes well is a team effort not a "you effort."

[Mikel Artetxe @artetxem]

Brilliant idea! Next up: Apple randomly reboots your Mac if you're building competing tech, Gmail silently edits your email if you mention rival platforms, and Tesla Autopilot swerves if it detects you're working on self-driving cars.

All in the name of safety, of course. Because malicious actors controlling the world’s operating systems, inboxes and cars would be extremely dangerous!

[Sebastian Raschka @rasbt]

Turns out Fable 5 is shadowbanning AI researchers 🫤

[Nathan Lambert @natolambert]

Labs starting to pull up the ladders on the ability to diffuse AI was inevitable. Doing it without telling the user is misaligned.

[AJ @ClearwaterCoder]

@eliebakouch intentionally hindering innovation isnt something a company with morals does

[Jonathan ⚡ @jonathanbylos]

@eliebakouch I get the safetyism, but this is super dystopian. We have to put a lot of faith in this central org, and historically that's just been a bad move in general, even for them. One leak, hack or mis-step and it's a disaster

[Amine Ben Khalifa @aminert]

@zuhayeer May Allah accept it brother!

[Paul Graham @paulg]

Tucker Carlson interviews a British doctor who worked in Gaza.

"Four young teenage boys were brought in, all of whom who'd been shot in the testicles."

[Amine Ben Khalifa @aminert]

Probabilistic machines writing code to call probabilistic machines, all the way down.

[Amine Ben Khalifa @aminert]

@DavidKPiano let's train it!

[Mehrdad Farajtabar @MFarajtabar]

Continual Learning remains one of the most challenging “holy grails” of AI.

Most discussions focus on catastrophic forgetting: models lose what they previously learned. But there is another equally important failure mode: over long continual training, neural networks can also lose their plasticity, ie, their ability to learn new things is weakened over time.

In our ICLR 2026 work with colleagues at @Apple and @ETH, we study this phenomenon, known as Loss of Plasticity (LoP), from a geometric perspective.

We show that LoP can arise when gradient dynamics become trapped in invariant manifolds of parameter space. In particular, we analyze two types of traps:

🔴 Frozen units: units saturate, gradients vanish, and they become effectively silent to backpropagation.

🔵 Cloned units: units become redundant, receive matching forward and backward signals, and move together.

For these structures, the gradient is tangent to the trap. Once standard GD/SGD enters these affine subspaces, it cannot leave them on its own. This means the dynamics can remain sticky even when the data distribution or task changes.

What we find especially interesting is that these traps are not merely optimization bugs. The same feature-learning pressures that help networks learn useful representations for the current task can also push them toward states with less future adaptability.

This raises a difficult open question for future work: are neural networks trained with SGD and cross-entropy loss fundamentally the right framework for continual learning?

Please read the full paper for more details: https://t.co/oPFIebhDUL

[Omar Shaik @omarshaikdev]

Asked Atta to compare the new GPT-5.5 release with other frontier models.

Takeaways:
Use Opus for coding, financial workflows, and tool-heavy orchestration.
Use GPT for math, abstract reasoning, and terminal-heavy development.
Use Gemini for browsing, academic knowledge, and cost-sensitive workloads.

https://t.co/ayaVl1aNT5

[Alonzo Canada @acanada]

My workflow now looks nothing like it did a year ago.

Fire up Cursor agents → new feature heading to production.
Open Figma → start exploring the next design.
Check on Cursor. Nudge the agents.
Back to Figma. Brainstorm with Claude on the design direction.
Check on Cursor. Guide. Redirect.
Iterate the design. Share with Claude for critique.
Guide agents more.
Finish the design in Figma.
Push the PR from Cursor.

Design and production aren't sequential anymore. They're parallel. The exploring feeds the building. The building pressure-tests the exploring.

This is what designing with AI agents actually looks like. It doesn't replace the craft, but compresses the loop between imagining something and shipping it.