Image
Introducing our biggest upgrades to the Deep Research API yet... including Deep Research Max (our SOTA system), MCP support, Native charts & infographics, planning mode, full tool support (including Google tools), full multi-modal input support, & real-time progress streaming!
Free-threaded Python with @fastapi.tiangolo.com 🧵🚀
(this is not a thread, just a threads emoji)
## THE SCENE: The London Studio It is 10:00 PM in a glass-walled studio overlooking the moonlit London skyline, but inside, it is blindingly bright. The red "ON AIR" tally light is blazing. Jaz is standing up, not sitting, bouncing on the balls of their heels to the rhythm of a thumping backing track. Their hands fly across the faders on a massive mixing desk. It is a chaotic, caffeine-fueled cockpit designed to wake up an entire nation. ### DIRECTOR'S NOTES Style: * The "Vocal Smile": You must hear the grin in the audio. The soft palate is always raised to keep the tone bright, sunny, and explicitly inviting. * Dynamics: High projection without shouting. Punchy consonants and elongated vowels on excitement words (e.g., "Beauuutiful morning").
The example prompt for Google's new Gemini Flash TTS text-to-speed model is a lot simonwillison.net/2026/Apr/15/...
Introducing Gemini 3.1 Flash TTS 🗣️, our latest text to speech model with scene direction, speaker level specificity, audio tags, more natural + expressive voices, and support for 70 different languages.
Available via our new audio playground in AI Studio and in the Gemini API! x.com/OfficialLoga...
Excited to launch the accompanying free RLHF Course for my book. To kick it off, I've released:
- Welcome video
- Lecture 1: Overview of RLHF & Post-training
- Lecture 2: IFT, Reward Models, Rejection Sampling
- Lecture 3: RL Math
- Lecture 4: RL Implementation
Landing page: rlhfbook.com/course
My book, Reinforcement Learning from Human Feedback, is wrapping up and going into final production (copyediting, making pretty, formatting, etc.). Shipping to you in 1-2 months!
It's a wonderful project to create a foundation of knowledge for the research communities that I love and operate in.
🌐 I've just released Sentence Transformers v5.4: we're going fully multimodal for embeddings & reranking!
Also featuring a modular CrossEncoder, and automatic Flash Attention 2 input flattening.
Highlights in 🧵
Running PyTorch natively on TPUs is on the way.
Google just announced they are working on TorchTPU, a new backend that will let PyTorch run natively on TPUs.
The idea is to let you switch the device to "tpu" without refactoring, replacing the older PyTorch/XLA approach.
Blog in 🧵
Just discovered this new book! 🤩 Insta-buy on my e-reader. From Discover to reading the first page in <5 min.
"The Infinity Machine" by Sebastian Mallaby.
Very much looking forward to this one!
www.amazon.ca/Infinity-Mac...
If you want to understand more about the internals of Gemma 4 and how the model is structured, this visual guide is a perfect start:
newsletter.maartengrootendorst.com/p/a-visual-g...
Today we are rolling out service tiers in the Gemini API! You can now (optionally) set "flex" or "priority".
In the case of flex, this will save you ~50% on API costs (with lower reliability).
In the case of priority, this will cost ~80% more but give you higher priority!
Google just dropped Gemma 4, and I'm sharing the scoop on this updated open weight model family. Built on Gemini 3 research, there are 4 model variants: E2B, E4B, 26B A4B, and 31B with full multimodal capabilities. And Gemma 4 is Apache 2.0 licensed!
www.youtube.com/shorts/rUMvZ...
New essay: Cutting Python Web App Memory Over 31%
I reduced Python web app memory by 3.2 GB using async workers, import isolation, the Raw+DC database pattern, and disk caching. Detailed before and after numbers for each technique.
mkennedy.codes/posts/cuttin...
Cloud Run Jobs or Cloud Batch? If you're running offline processing on Google Cloud, the choice isn't always obvious.
While both services are built for run-to-completion tasks, they serve very different needs.
Let me help you choose with my guide: medium.com/p/8590a8e3a3b1
Image
Image
Introducing Gemini 3.1 Flash Live, our new realtime model to build voice and vision agents!!
We have spent more than a year improving the model + infra + experience, the results? A step function improvement in quality, reliability, and latency. x.com/OfficialLoga...
I said digital attestations and `pylock.toml` would have helped with the litellm attack. People asked for more details, so I wrote a blog post explaining why it would have helped.
snarky.ca/why-pylock-t...
Limit candidate packages to those that were uploaded prior to the given date. Accepts RFC 3339 timestamps (e.g., 2006-12-02T02:07:43Z), local dates in the same format (e.g., 2006-12-02) resolved based on your system's configured time zone, a "friendly" duration (e.g., 24 hours, 1 week, 30 days), or an ISO 8601 duration (e.g., PT24H, P7D, P30D). Durations do not respect semantics of the local time zone and are always resolved to a fixed number of seconds assuming that a day is 24 hours (e.g., DST transitions are ignored). Calendar units such as months and years are not allowed. May also be set with the UV_EXCLUDE_NEWER environment variable.
In light of the fallout from the LiteLLM supply chain attack, I just learned that you can exclude newly published package versions when installing with uv using exclude-newer.
Congrats!
Starlette 1.0 is here!🎉
After nearly eight years, Starlette has reached its first stable release. Downloaded almost 10 million times a day, it serves as the foundation for FastAPI and the Python MCP SDK.
Blog post: marcelotryle.com/blog/2026/03...
Release notes: www.starlette.io/release-note...
A promotional graphic on a dark background with a thin yellow border features the orange "colab" logo on the left, a large orange heart emoji in the center, and a white stylized graphic above the text "Model Context Protocol" on the right. Below the main graphic is the word "Announcement" in faint grey text next to abstract yellow shapes. The image signifies an integration or partnership between Google Colab and the Model Context Protocol.
Bringing Google Colab’s secure environment to your AI agents. ☁️🛠️
With the new open-source Colab MCP Server, your agents can now natively write and execute code inside a Colab Notebook. Secure, automated, and ready to go.
Andrej Karpathy just open-sourced autoresearch: a framework where AI agents don't just tune hyperparameters, they rewrite the entire training codebase to improve model performance. I break down how it works in under 2 minutes.
www.youtube.com/shorts/zxYF2...
If you are hitting 429 ResourceExhausted errors when you call models on Vertex AI, throwing in a while True retry loop would not help.
Here are 5 fixes you may want to consider 🧵
We've released Python Polars 1.39. Some of the highlights:
• Streaming AsOf join, enabling memory-efficient time-series joins.
• sink_iceberg() for writing to Iceberg tables
• Streaming cloud downloads for scan_csv(), scan_ndjson(), and scan_lines()
github.com/pola-rs/pola...
Image
Say hello to Gemini Embedding 2, our new SOTA multimodal model that lets your bring text, images, video, audio, and docs into the same embedding space! 👀 x.com/OfficialLoga...
Agents can write code.
Scaling them is the real challenge.
Air is a dedicated environment for agent-driven development: parallel, isolated execution, full-project review, and support for Codex, Claude Agent, Gemini CLI, and Junie.
Get started for free.
Download Air for macOS: https://jb.gg/0y9ntn
The next scikit-learn release will allow inspecting the type and values of attributes of fitted estimators in Jupyter notebooks & example code rendered as HTML in sphinx-gallery powered project websites.
scikit-learn.org/dev/auto_exa...
Weekend test: running JAX using GPU on MacBook .
jax-mps is a PJRT plugin that enables GPU-accelerated JAX on Apple Silicon via Metal Performance Shaders.
Requires Apple Silicon macOS + Python 3.13. Installs and auto-registers with JAX.
Repo: github.com/tillahoffma...
Image
Introductions Gemini 3.1 Flash-Lite 🔦, a huge step forward on the boundary of intelligence beating 2.5 Flash on many tasks. x.com/OfficialLoga...
