Jonathan Ouyang

intern @GoogleDeepMind @GoogleAIStudio

[Jonathan Ouyang @jonsouyang]

This is my vision for the future of Gemini and AI
Assistants.

Computer Use + Browser Use + Command Line Agent + On Screen Annotation.

I finally got around to building a quick experimental product and demo, using the Gemini models

If you also love this space, let's chat!

[Jonathan Ouyang @jonsouyang]

@ayushunleashed Ah well…

[Jonathan Ouyang @jonsouyang]

I’m scared of trying out fable, what if it burns through all my credits and still produces slop?

[Jonathan Ouyang @jonsouyang]

@rickymaggio02 @yacineMTB Washing and putting the rice into the cooker is too much work, make the robot do it

[Jonathan Ouyang @jonsouyang]

@tylerxdev Probably wouldn’t work very well lmao

[Jonathan Ouyang @jonsouyang]

Well I’ll be damned. Maybe local models actually CAN run on my MacBook now

[Philipp Schmid @_philschmid]

More Gemma 4! New QAT Gemma 4 checkpoints with similar performance while using ~4x less memory!

It comes with a new mobile quantization format that reduces memory footprint of Gemma 4 E2B to just 1GB.

Quantization-Aware Training (QAT) simulates low-precision operations during training to allow loss-less quantization afterwards for smaller, faster models while maintaining accuracy.

Available on @huggingface and directly runnable.

[Jonathan Ouyang @jonsouyang]

Would argue that agentic coding IS gambling. Either the project gets done in 2 hours or 2 weeks based on how smart Claude wants to be on that day

[Zack Voell @zackvoell]

Agentic coding is more addictive than gaming

[Jonathan Ouyang @jonsouyang]

@math_on_tap Fortune favors those with lots of GPUs

[Jonathan Ouyang @jonsouyang]

I interviewed for an ML sim role at nvidia and they laughed at me when I told them our ucla lab only had 8 GPUs. Apparently that was a bad answer

[steve @gpusteve]

you're interviewing for an ml performance role at nvidia and they ask:

"why can't we just replace hbm with huge on-chip sram?"

you say: "because all chips need hbm." wrong.

here's how you answer:

[Jonathan Ouyang @jonsouyang]

@JakeFleshner Quantifying Attention

[Jonathan Ouyang @jonsouyang]

I used to dream of swimming at the Olympics before pivoting to building a B2B vertical SAAS AI startup

[𝑹𝒊𝒐 ᯓᡣ𐭩 @rioserenity_]

Drop a random fact about yourself

[Jonathan Ouyang @jonsouyang]

@notcee_fan @danielwu28 dude this is literally your apartment living room

[Jonathan Ouyang @jonsouyang]

@0xkozue Did bro just describe a robotics egocentric data collection company

[Jonathan Ouyang @jonsouyang]

@lyc_aon Alg 😂

[Jonathan Ouyang @jonsouyang]

@lyc_aon Salt in the wound man

[Jonathan Ouyang @jonsouyang]

@0xkozue You liked the shorter hackathon styles more?

[Jonathan Ouyang @jonsouyang]

@lloydhightower I’ll discuss it with the team 😉

[Jonathan Ouyang @jonsouyang]

What if we built harness evals and leaderboards…. That might be cool

[Philipp Schmid @_philschmid]

My personal research question for today: Should we optimize the model for a harness or should the harness be optimized for the model?

[Jonathan Ouyang @jonsouyang]

@ash_twtz My MacBook might genuinely melt through my table if they put a nvidia GPU in it

[Jonathan Ouyang @jonsouyang]

@neerajjj6785 You can always magic wormhole it

[Jonathan Ouyang @jonsouyang]

@raayandhar Or hundreds Idr this was a while ago