Open source ML @ Google NYC
New York, NY
Joined April 2011
- Tweets1.8K
- Following702
- Followers30.9K
- Likes2.7K
New Kaggle competition. Add reasoning capabilities to a small Gemma model using Tunix (a new popular RL library in JAX). This is really great, thanks to Wei on my team + many others for driving it. Check it out!
📣 Hackathon Launch Alert! Google Tunix Hack - Train a Model to Show Its Work hosted by @googlecloud
🎯 Train Gemma models to show reasoning using Tunix
💰 $100,000 Prize Pool
⏰ Final Submission: Jan 12, 2026
kaggle.com/competitions/goog…
Josh Gordon retweeted
KerasHub now includes HGNetV2!
We’re excited to bring the high-efficiency, high-accuracy HGNetV2 image classification backbone into KerasHub’s model family.
Model details and quickstart notebook are available on Kaggle: kaggle.com/models/keras/hgne…
#keras #kerashub #HGNet
Josh Gordon retweeted
The JAX team is hosting a dinner / networking event during ICML on Thursday. Join us for an evening of food, drinks, and discussion of all things JAX.
@SingularMattrix
and other JAX team member will be attending. Please register early as capacity is limited.
RSVP: rsvp.withgoogle.com/events/j…
Josh Gordon retweeted
🎬 Generate videos with the Gemini CLI
Add:
🧑💻 GenMedia MCP servers for Imagen, Veo & Chirp
📝 A GEMINI܂md file explaining your ✨ creative process
And you too can take 🙀 Rusty the Cat on an adventure
⬇️ Full tutorial in the vid ⬇️
Josh Gordon retweeted
We’re building Keras Recommenders at Google, and would love to hear from people working in RecSys to understand what they need. What features matter most to you?
DMs are open, feel free to reach out!
keras.io/keras_rs/
People are getting monster offers to work on LLMs behind closed doors, and meanwhile this team from Stanford is doing everything in the open and making it available to everyone. Open science all the way!
It was *so good* seeing everyone at the JAX and OpenXLA DevLab this week! Best event in a long time. Let’s do it again!
Josh Gordon retweeted
I am PUMPED to finally share what we’ve been working on:
🖥️ Introducing the Gemini CLI!
It can code, sure, but with access to your system shell, files and MCP servers, it can also:
👩🔬 Do research
💽 Organise your MP3s
Resolve rebases
🔬 Even strace that weird hung process
Check it out: goo.gle/gemini-cli
And I REALLY want to see how it helps you, so tag or share your workflows here!
ALT Install the Gemini CLI with npm install -g @google/gemini-cli
ALT Gemini CLI screenshot
Here's more detail on how to load a Hugging Face checkpoint into a KerasHub model. Thanks for the walkthrough, @yufengg , @divyasheess, and @monicadsong !
developers.googleblog.com/en…
You can find performance & scale optimized JAX models in MaxText and MaxDiffusion:
* github.com/AI-Hypercomputer/…
* github.com/AI-Hypercomputer/…
You can also use Keras / JAX to tune many Hugging Face Transformers model checkpoints by loading them into a KerasHub model. It's pretty cool!
* huggingface.co/blog/keras-hu…
* keras.io/keras_hub/presets
Josh Gordon retweeted
KerasHub supports loading checkpoints of many model architectures from HuggingFace.
So if there is a model checkpoint on HF that is not in the Keras format, you can easily load it in Keras and use it as a regular Keras model in any backends (JAX, TF, or PyTorch).
You can find performance & scale optimized JAX models in MaxText and MaxDiffusion:
* github.com/AI-Hypercomputer/…
* github.com/AI-Hypercomputer/…
You can also use Keras / JAX to tune many Hugging Face Transformers model checkpoints by loading them into a KerasHub model. It's pretty cool!
* huggingface.co/blog/keras-hu…
* keras.io/keras_hub/presets
Josh Gordon retweeted
If you're running JAX and you need to grab a model checkpoint from HuggingFace, KerasHub has you covered. Load, fine-tune, quantize, export for inference.
You can find performance & scale optimized JAX models in MaxText and MaxDiffusion:
* github.com/AI-Hypercomputer/…
* github.com/AI-Hypercomputer/…
You can also use Keras / JAX to tune many Hugging Face Transformers model checkpoints by loading them into a KerasHub model. It's pretty cool!
* huggingface.co/blog/keras-hu…
* keras.io/keras_hub/presets
You can find performance & scale optimized JAX models in MaxText and MaxDiffusion:
* github.com/AI-Hypercomputer/…
* github.com/AI-Hypercomputer/…
You can also use Keras / JAX to tune many Hugging Face Transformers model checkpoints by loading them into a KerasHub model. It's pretty cool!
* huggingface.co/blog/keras-hu…
* keras.io/keras_hub/presets
Really excited for the upcoming JAX & OpenXLA DevLab this Monday! This is a small group deep dive on the latest techniques, with breakouts on special interests. We'll record the tutorials for everyone, too.
Opportunity: If you're interested in *healthcare research* with JAX / LLMs, we have a couple more seats available at that breakout. Send me a DM here if you'd like to attend.
openxla.org/events/summer_de…
Yes! And you can start developing with the Gemini Developer API in just a couple minutes.
If you're just getting started, here's a quick walkthrough that shows you how to get your API key and run a neat audio generation example in Colab.
1/ Visit aistudio.google.com to get your API key
1/ A quick recap from another week of rapid shipping:
Great response to our Gemini 2.5 Pro update now in preview. Find it in AI Studio, Vertex + @GeminiApp.
And good news for @GeminiApp Pro plan members: your 2.5 Pro limit doubled from 50 to 100 queries per day.
10/ You can play the audio right in Colab.
There's also a neat example that prompts the model to read a discussion between two speakers, like NotebookLM.
Add your own prompts and see what the model can do.
11/ You can find lots more examples in quickstarts folder in the cookbook.
Also check out the developer docs on ai.google.dev/gemini-api/doc… for lots more walkthroughs, and code examples for JS developers.
If you have questions or run into bugs, the best place to ask them is on discuss.ai.google.dev/c/gemi…
Have fun!
