Working at the intersection of AI and gaming. CS PhD '26 @Stanford, advised by Kayvon Fatahalian
Joined October 2021
- Tweets63
- Following128
- Followers158
- Likes156
Vishnu Sarukkai retweeted
AI is using more and more of the world’s energy. But how much economic value are we getting from AI?
We’re introducing Joules-to-GDP (J2G), a framework that measures how well AI turns energy into economic value and shows which tasks, models, and hardware give the best return.
Across 44 occupations in finance, engineering, law, administrative work, and more in @OpenAI's GDPval, we arrive at three findings:
1) The tasks with the highest economic value are not the tasks with the highest economic value per joule. For example, CAD design earns ~10% more per task than building websites, but the latter earns ~27% more value per joule.
2) The most powerful model isn’t always the best use of energy. Gemma 4 31B delivers 56% of GLM 5.2’s net value with 23% of the energy, for 2.5x the value per joule.
3) More chips don’t guarantee proportionate returns. Moving the same model from 1 to 8 NVIDIA H100s uses 3.2x the energy but cuts completion time by only 22%, with about the same value.
With @SnorkelAI, we’re also releasing GDPWorks, a benchmark for economically valuable tasks across 36 additional occupations.
Joint work co-led by @amirzeinaalii and @Avanika15, with @Mersad_Abbasi, @Jon_Hartley_, John Hennessy, @HazyResearch, and @Azaliamirh.
More details below 🧵
Anyone tested using a different timestep shift for text-to-image diffusion inference vs when conditioned on a LoRA (depth/pose/etc)? Intuitively feels like you might need less shift when the conditioning makes it easier to learn rough image structure.
Anyone have a favorite method of compressing diffusion models/reducing flops? Seen a lot of techniques out there (sub quadratic attention, fewer tokens, etc.)
It's pretty amazing that Opus can quickly vibe-code game concepts, but the hard work of making a game is still in crafting the mechanics and gameplay. A game isn't just a short video with cool visuals :)
Playing STS2, it's crazy how playing a run as Silent takes twice as long as Ironclad. But winning on Silent feels way more strategic/clever than just stacking vulnerability and strength with Ironclad.
What's the most compelling game demo that's been built on top of Generative Agents? What's the best way to use agents with short and long-term planning as a key part of gameplay?
I stopped playing Catan as soon as I discovered Civ 6. The brilliant thing about Civ is that it's "historical," without being too serious. Where else can the Mongols build the pyramids, recruit da Vinci, and conquer with tank armies?
It's sad that AAA studios and blockbuster movies have fallen into the same trap: spending more and more on visuals, relying on tried-and-true (tired-and-true?) franchises, forgetting that novel stories/gameplay are what make new movies/games special.
I love Breath of the Wild, but the game had no replay value for me. Once I've covered all corners of the map and found (most of) the game's secrets, the mechanics weren't enough by themselves to bring me back. While more sim-style games (ex. Civ) have infinite replay value.
I've been playing with a few different strategies to give NPCs the ability to learn from their past. Would this type of simple learning loop make video game combat more fun?
Vishnu Sarukkai retweeted
Most times I use an agent it spends minutes rereading all my stuff – spending dollars and energy just to gather the context it needs to understand my work.
But it won’t always be this way. We can train that understanding into the model – amortizing across tasks.
We started Engram to build the systems required for a future where there are billions of models, each learning from users and deeply understanding their work.
I’m so grateful and excited to be working on this problem with this incredible team
Vishnu Sarukkai retweeted
Fresh blog post!
@modal partnered with @ScalingIntelLab, @HazyResearch, and @chelseabfinn's IRIS Lab to speed up research on speeding up AI research.
Read how scientists at the cutting edge are building the machines that build the machines with Modal.
modal.com/blog/accelerating-…
One key differentiator between effective automation and AI slop is caring — how much developers care about output quality and how well they can identify deficiencies.
I love Catan. I’ve played in tournaments, won one, and competed in the Catan National Championships. That care motivated how I approached building a digital version of the game.
I set up an agentic loop where agents generate game logic, other agents play the game to surface errors, and every human intervention becomes part of a growing library of regression tests. The result: a reliable digital implementation of Catan, alongside LLM agents that can actually play the game.
Here's how I did it: vsanimator.substack.com/p/ne…
Presenting our work on self-improving LLM agents via in-context examples today at 4:30pm at #NeurIPS2025! Location #3400
Paper: arxiv.org/abs/2505.00234
Come by! 👋
We also explore distilling strong "teacher" models into weaker "student" models—all via in-context learning 🎓
arxiv.org/abs/2512.02543
For more informal explorations of agentic self-improvement, check out my Substack:
vsanimator.substack.com