Sam Tukra

Chief AI Officer & Co-Founder | Obsessed with research and AI

[Sam Tukra @SamTukra]

After a lot of coding, experimentation, and engineering. Claude … didn’t make me a million dollars 😭

[Sam Tukra @SamTukra]

@kcarriedo That's awesome! requested an invite.
I've done a workaround by solidifying my workflow (the steps I take to do my ML Research) and then allow claude to bypass permissions. Since its operating in a sandbox, failure is limited. Though, still want to harden.
https://t.co/gCd0O45flh

[Sam Tukra @SamTukra]

Is anyone else tired of approving/denying Claude Code CLI requests? 😂
I get why it exists. Safety matters. Only way out I see: define workflow, allowed actions, hard gates, rollback points, and success criteria upfront.
Then let agents run inside that box.
Any other ideas?

[Sam Tukra @SamTukra]

@RyanOlunix Hey
@RyanOlunix
working on an open-source autonomous AI research team.
The idea is: user writes a plan + success metric, and a swarm of agents run the ML lifecycle around it, to deliver a working model for any data / problem use case
https://t.co/gCd0O45flh

[Sam Tukra @SamTukra]

@bambusdani Super useful, thanks so much for sharing. Hadn’t seen ECC before. The task-shaping angle is exactly what I was trying to get at.

Permissions still feel like the messy bit to standardise: easy to make them safe, harder to make them non-annoying. Will dig into this.

[Sam Tukra @SamTukra]

@bambusdani Yep, that sounds right. Smaller permission boxes + trustworthy logs feels like the right sandbox of autonomy.

Do you have any examples of setups/repos/blogs where people are doing this well? Especially around task shaping + permissions.

[Sam Tukra @SamTukra]

@NikhilKagita_ Hey @NikhilKagita_ working on an open-source autonomous AI research team.

The idea is: user writes a plan + success metric, and a swarm of agents run the ML lifecycle around it, to deliver a working model for any data / problem use case

[Sam Tukra @SamTukra]

@JohnnyNel_ Exactly. I feel the trick is not removing approvals entirely it’s moving them up a level. So only important stuff comes to user ... i.e. Approve the operating envelope, then only interrupt on boundary violations or failed gates.

[Sam Tukra @SamTukra]

@gregisenberg This is where agent systems are moving: not just more agents, but stronger workflow contracts. I’ve seen the same in ML/research workflows. Orchestration only works when success criteria are explicit. Otherwise 100 agents = 100 confident ways to drift 😅

[Sam Tukra @SamTukra]

@Jason I’m in

[Sam Tukra @SamTukra]

@beffjezos It’s crazy how real this looks!

[Sam Tukra @SamTukra]

@DudespostingWs What brand safe is this? If it can survive that, I sure as hell want one!

[Sam Tukra @SamTukra]

@TimesAlgebraIND Secularity …. Absolute nonsense!

[Sam Tukra @SamTukra]

@realGeorgeHotz The job market is gonna change in the next 2-3 years drastically. I feel for all younger grads. Tough times.

[Sam Tukra @SamTukra]

@HoodedClaw1974 Yehp, totally not a terror related incident, just happens to have happened on Christmas Day, and no reveal of race, age or any other details of the driver.

[Sam Tukra @SamTukra]

@worldranking_ 💯 why is this even a question?

[Sam Tukra @SamTukra]

@realMaalouf How is it, that these people aren’t getting arrested? But people commenting on Facebook are?

Man, UK has changed. I didn’t even recognise this was London from the video

[Sam Tukra @SamTukra]

@Salwan_Momika1 wtf!!! What more evidence is required to stop these people?!?!

[Sam Tukra @SamTukra]

@clownworld YES

[Sam Tukra @SamTukra]

Keras 3.0 runs on top of Jax, @TensorFlow and @PyTorch 🤯. Is this the first unification of DL frameworks? Can’t wait to try it out!

#deeplearning #MachineLearning

https://t.co/60OfVqNTee