Michael Tang

computer use / prev @princeton

[Michael Tang @_michaeltang_]

Historically, exploration in RL has been achieved with a combination of subgoals, rewards, and expert demonstrations.

We avoid *all of that* — contrastive learners equipped with a single (sparse!) goal exhibit exploration and skill learning. Some key findings in thread 🧵:

[Michael Tang @_michaeltang_]

@tinahhong congrats on the launch !! :’)

[Michael Tang @_michaeltang_]

🌅 :’)

[Bassil Shama @BassilShama]

Opus 4.6 is our most capable Computer Use model to date. Excited for everyone to give Computer Use a try with Claude in Chrome, Cowork, and Claude Code!

To celebrate, I let Claude (4.6) Monet show off his artistic side in the Claude for Chrome extension.

[vincent! @vvhuang_]

We trained a decoder to read the internal activations of an LLM and answer questions about what the model will think about or do next.
We find that this decoder can understand LLM behaviors, even when the model itself is confused! (for instance, if the model has been jailbroken)

[Michael Tang @_michaeltang_]

🎨🎨🎨

[hattie @oh_that_hat]

>be me
>be Claude
>have read the internet but one day human asks me to draw
>no training, no practice, just converting mental image to mouse movements like a toddler holding a crayon
>pencil tool not working? np, I'll draw with the eraser
>task failed successfully

[Anne Ouyang @anneouyang]

Excited to share what friends and I have been working on at @Standard_Kernel

We've raised from General Catalyst (@generalcatalyst), Felicis (@felicis), and a group of exceptional angels.

We have some great H100 BF16 kernels in pure CUDA+PTX, featuring:
- Matmul 102%-105% perf of cuBLAs in 100 lines of code
- Attention 104% perf of FlashAttention3 in 500 lines
- Fused Llama3 FFN 120% perf of PyTorch (gpt-fast)

Reach out if you want to work on AI kernel gen with us!

[Michael Tang @_michaeltang_]

VideoGameBench full release! now has metrics 👀

[alex zhang @a1zhang]

Finally, how did we automatically track progress? We compute perceptual image hashes of “checkpoint frames” that always appear in the game and compare them to the current screen, and use a reference walkthrough to estimate how far in the game the agent is!

[Michael Tang @_michaeltang_]

@mengk20 omg hii

[Michael Tang @_michaeltang_]

at ICLR, come say hi!

[Michael Tang @_michaeltang_]

I love this project

[alex zhang @a1zhang]

Claude can play Pokemon, but can it play DOOM?

With a simple agent, we let VLMs play it, and found Sonnet 3.7 to get the furthest, finding the blue room!

Our VideoGameBench (twenty games from the 90s) and agent are open source so you can try it yourself now --> 🧵

[Michael Tang @_michaeltang_]

@a1zhang alex you should run a whiteboarding session on kernels for bio :)

[Michael Tang @_michaeltang_]

If you're at NeurIPS, stop by our workshop oral! @GraceLiu78 will be presenting on single-goal contrastive RL at 2:45-3pm on Sun 12/15, don't miss it :) https://t.co/zK6qYy5tKF

[Michael Tang @_michaeltang_]

@mengk20 @COLM_conf omg yes

[Michael Tang @_michaeltang_]

Presenting USACO at @COLM_conf, reach out to chat about anything coding, agents, test-time methods, or other nlp / rl topics :)

[Michael Tang @_michaeltang_]

@a1zhang come to colm !

[Michael Tang @_michaeltang_]

👀

[Ben Shi @BenShi34]

What a coincidence me too

[Michael Tang @_michaeltang_]

"Next frame prediction achieves a PSNR of 29.4, comparable to lossy JPEG compression." 😮

[AK @_akhaliq]

Google presents Diffusion Models Are Real-Time Game Engines

discuss: https://t.co/IBV70a990g

We present GameNGen, the first game engine powered entirely by a neural model that enables real-time interaction with a complex environment over long trajectories at high quality. GameNGen can interactively simulate the classic game DOOM at over 20 frames per second on a single TPU. Next frame prediction achieves a PSNR of 29.4, comparable to lossy JPEG compression. Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation. GameNGen is trained in two phases: (1) an RL-agent learns to play the game and the training sessions are recorded, and (2) a diffusion model is trained to produce the next frame, conditioned on the sequence of past frames and actions. Conditioning augmentations enable stable auto-regressive generation over long trajectories.

[Michael Tang @_michaeltang_]

Next-gen evals should be interactive 👀
• Long-horizon env interaction tests realistic planning (SWEBench, WebShop)
• Human interaction draws out hidden capabilities (USACO human-in-loop)
• Multi-turn interaction with a bad actor better captures the surface area for attacks

[Nathaniel Li @natliml]

Who's better at LLM mischief — humans or AIs? Spoiler: It's us.

Human red teamers achieve 70%+ attack success rates against LLM defenses that stump automated adversarial attacks. Why? We’re better at adversarial yapping.🧵

[Michael Tang @_michaeltang_]

fast edit 👀

[Zed @zeddotdev]

Fast Edit Mode: a breakthrough from @AnthropicAI that we're piloting in Zed.

It allows Claude 3.5 Sonnet to echo its input far faster than generating new text.

The result? Near-instantaneous refactoring.

We're collaborating with their team in Zed's open source codebase.

[Michael Tang @_michaeltang_]

Excited about more robust repository-level code evals!

[OpenAI @OpenAI]

We're releasing a new iteration of SWE-bench, in collaboration with the original authors, to more reliably evaluate AI models on their ability to solve real-world software issues. https://t.co/qJuLpCdSWJ

[Ben Eysenbach @ben_eysenbach]

I'm excited to share recent work with @GraceLiu78 and @_michaeltang_ on exploration in RL!

A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals

Paper, code, and videos: https://t.co/MVfTzC6KMJ

A thread.