Senior AI Dev Engineer @googlecloud | Research, Open Models, Ray & TPU | Instructor @DeepLearningAI | Startup Advisor @ycombinator x Google Cloud
San Francisco, CA
Joined February 2012
- Tweets1.9K
- Following1.4K
- Followers2.1K
- Likes1K
Pinned Tweet
Presenting at the Next '26 Developer Keynote is one of those moments I'll remember for a long time.
Thank you to everyone who played a part.
Till the next Next!
Can’t wait for this !
From watching worlds on screen to interacting with them in real time, join @willy_visko, @buzzhays, @ivnardini, and Dave Elliott this Wednesday to explore a new form of entertainment powered by our Live Models.
@buzzhays brings over 25 years of Hollywood experience, having produced feature films for Sony, Paramount, Disney, and Universal. He began his career at Lucasfilm, leading research and development at its THX division 🫡
Pull up a seat: luma.com/83ftfeg6
@GoogleStartups
To the space
We’re sending TPUs to space (yes, really).
After years of research, we’re launching a satellite to evaluate if and how Google Tensor Processing Units (TPUs) hold up in orbit.
The test mission, as part of our latest moonshot — Project Suncatcher — is designed to gather data exploring how we can one day host machine learning infrastructure in space.
ALT White text on a black background with photos of a TPU and a rocket orbiting Earth. The text reads: “Meet Project Suncatcher, our moonshot to test whether we can host AI infrastructure in space.”
ALT An image showing circuit boards, overlaid with text that reads: “We partnered with Planet to send a prototype satellite carrying four TPUs into orbit. We’ll be looking to see how well the TPUs withstand the harsh conditions of space to assess whether we can eventually host machine learning in orbit.”
Google DeepMind and Google Research recently announced WeatherNext 2, a global medium-range weather and cyclone forecasting model
I ran it on a TPU VM and wrote an onboarding guide. It walks through TPU VM provisioning and shows how to generate and collect ensemble forecasts
Links in the 🧵
Guide: github.com/google-deepmind/w…
Repo: github.com/google-deepmind/w…
TPU Developer Hub: cloud.google.com/products/tp…
The second and last blog in the Run Ray on TPU series is out and it walks through three libraries you can use to actually build: Ray Serve, Ray Data and Ray Train
Blog:
developers.googleblog.com/ru…
I’m writing a two-part series on running Ray on TPU. Part 1 is 👇
As of Ray 2.55, TPUs are first-class accelerators in Ray. Official pre-built images are supported, along with official APIs and AI libraries
In Part 1, I cover what you need before writing and running code with Ray on TPU, then walk through the stack.
Blog:
developers.googleblog.com/ru…
AlphaEvolve just went GA on Google Cloud
Google DeepMind's LLM-guided evolutionary coding agent. You give it a program to improve, a goal, and a scoring function and it uses Gemini to rewrite the code, score each candidate, and evolve the best over generations
I tested it to tune LoRA recipe for Gemma 4 on a function-calling task using Ray. But you can use it for algorithm discovery, mathematical search, combinatorial optimization and more.
Just published a write-up on elastic training on TPU, in collaboration with Abhinav (MaxText) and Luke (Pathways)
In the blog, we discuss why it’s needed, what elastic training is, and how it works with MaxText, Pathways, and Orbax on TPUs with a simple example
Full blog 👇
developers.googleblog.com/we…
3/ W&B quietly added coverage for JAX/TPU
While logging a fine-tuning job, I noticed new first-class JAX metrics like jax/core/compile/* for XLA compile time and jax/orbax/write/* for sharded-checkpoint I/O throughput
Those are metrics to track in training and now they’re just there
I'm still experimenting here but both of these unblocked me. If you're running Ray/JAX on TPU, they're worth knowing
vLLM torchax: docs.vllm.ai/projects/tpu/en…
wandb PR: github.com/wandb/wandb/pull/…
2/ If a model isn't supported on vLLM-TPU, you're not necessary stuck. vLLM-TPU has two model ways to serve a model:
- flax_nnx: native JAX impls but only for archs in its registry
- vllm: runs the PyTorch definition through torchax instead
1/ Been testing with running Ray on TPU this week for serving and fine-tuning end to end. Two things surprised me 🧵
Introducing Claude Apps Gateway on Google Cloud
You can use Claude Code with Google Cloud for a while now. You set a few env vars, grant some roles, and tokens flow from your project. But this setup can be hard to scale because per-dev attribution is weak, and there are no hard spend caps
With Anthropic, we introduced Claude Apps Gateway. It is a single stateless container on Cloud Run (backed by Cloud SQL) that sits between Claude Code and Google Cloud. It holds a credential and calls the model on each developer's behalf
Developers `/login` . Admins get a centralized control plane
Roy and I wrote a dev guide
cloud.google.com/blog/topics…
We cooked! Stay tuned 🙃
Teamed up with @ivnardini again to cook our new course about JAX on GPU. I’m glad we got to do it in person this time in this super cool studio in Bay Area 😎
My first contribution to MaxText just merged and it is a get started guide to elastic training on TPUs with Pathways
Guide
maxtext.readthedocs.io/en/la…
Anthropic just announced a new batch of self-hosted sandbox providers, and Google Cloud is on the list with the GKE Agent Sandbox.
With this integration, Claude and the agent loop stay on the Anthropic platform, while every tool call runs in a gVisor pod on your cluster inside your environment, close to your data, in a pool that scales with demand.
Sample repo in 🧵
Claude Managed Agents can operate in a sandbox you control, on your own infrastructure or with any provider you choose.
Today we added new guides for @blaxelAI, @e2b, @googlecloud, @namespacelabs, and @superserve_ai, so you can choose the best fit for your use case.