Thomas Edlich

Science at Google DeepMind. AlphaQubit. Ex-BenevolentAI. Opinions are my own.

[Thomas Edlich @grumpybavarian]

Gemini 4 Argon made our AlphaQubit model twice as fast! ⚡️
Quantum error correction has strict realtime requirements: for superconducting qubits we receive new measurements every microsecond. This allows us to decode larger codes in realtime and use larger, more accurate models.

[Pushmeet Kohli @pushmeet]

In quantum computing, Argon is accelerating both error correction and algorithm design. It translated our AlphaQubit decoder into a single optimized kernel, nearly doubling its inference throughput so we can deploy larger, more capable models and scale to larger quantum devices, while continuing to meet strict real-time thresholds.

[Silvia Sapora @silviasapora]

Accepted at #ICLR2026!
1/🧵 Inverse Reinforcement Learning typically produces opaque black-box rewards that are impossible to debug.
But what if we could learn rewards as executable, human-readable Python code instead? 🐍
Introducing GRACE: Generating Rewards As CodE. 👇

[Silvia Sapora @silviasapora]

🧵 Check out our latest preprint: "Programming by Backprop". What if LLMs could internalize algorithms just by reading code, with no input-output examples? This could reshape how we train models to reason algorithmically. Let's dive into our findings 👇

[Google @Google]

Introducing Gemini 2.0, our most capable AI model yet designed for the agentic era. Gemini 2.0 brings enhanced performance, more multimodality, and new native tool use.

[Thomas Edlich @grumpybavarian]

Check out the paper and the blog post!

Paper: https://t.co/5jIpu9NTpr
Blog: https://t.co/SzDhSqgl61

Very fortunate to have been part of this amazing collaboration between @GoogleDeepMind and @GoogleQuantumAI.

[Thomas Edlich @grumpybavarian]

Beyond excited to introduce AlphaQubit, now published in @Nature!

AlphaQubit is a neural network for quantum error correction and achieves state-of-the-art accuracy on simulated and real-world data.

[Adam Roberts @ada_rob]

I’m so proud of the updated version of #MusicFXDJ we developed in collaboration with @jacobcollier, available today at https://t.co/zizZMdGvwJ.

Over the past year I’ve spent countless hours experimenting with our real-time music models, and it feels like I’ve learned to play a new type of instrument, in the spirit of what we started @GoogleMagenta to achieve.

It was validating to hear Jacob echo these sentiments the first time he played an early version. With his help, we’ve added new controls, raising the ceiling on what players are able to do with it.

Now anyone can experiment with this new instrument and share what they create with the world.

I'll start things off with my most recent jam session...

[Thomas Edlich @grumpybavarian]

Check out our paper on how an agent can learn to unknot through reinforcement learning!

https://t.co/v7hOKTX7Up

[Nenad Tomasev @weballergy]

Another example of how reinforcement learning can help with exploring the space of mathematical conjectures - both in the broader sense, but also in more narrowly defined subdomains, for more specific questions:

[Google DeepMind @GoogleDeepMind]

We’re presenting the first AI to solve International Mathematical Olympiad problems at a silver medalist level.🥈

It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. 🧵 https://t.co/SYaLPSbIyj

[Silvia Sapora @silviasapora]

1 / 🧵 Excited to introduce our #ICML2024 paper: 😈 EvIL (Evolution Strategies for Generalisable Imitation Learning) a new inverse RL (IRL) method for sample efficient transfer of expert behaviour across environments – it's so good, it's downright EvIL!

[Thomas Edlich @grumpybavarian]

@ethanCaballero This was the case at BenevolentAI (pre-IPO) where leavers lost all vested equity. At some point they introduced a good-leaver policy where you could keep it if you had been there for several years.

[Google DeepMind @GoogleDeepMind]

Welcome to the next chapter in the Gemini era. ✨

We’re bringing our models to more @Google products you know, use & love.

From today, Bard will be known as Gemini — and you can access our most capable model, Ultra 1.0, with Gemini Advanced. → https://t.co/EfD6YDn6oK #GeminiAI

[Silvia Sapora @silviasapora]

1/🧵Jax fans, buckle up! Prepare to ditch slowpoke IRL training with the ⚡️jaxirl⚡️ launch, a JAX library built for Inverse Reinforcement Learning algorithms that's 250x faster 🏎️ 💨 than PyTorch.
Read on to learn more ⬇️⬇️⬇️

[Thomas Edlich @grumpybavarian]

@acvaucher @IBMResearch Wow, huge congratulations!!

[Tim Green @tfgg2]

New Research Engineer opening in the DeepMind science team: https://t.co/ilbHFcrO4H - please share with your networks!

[AK @_akhaliq]

A Neural Network Solves and Generates Mathematics Problems by Program Synthesis: Calculus, Differential Equations, Linear Algebra, and More
abs: https://t.co/7RTy5wKxTh

first work to automatically solve, grade, and generate university-level Mathematics course questions at scale

[Google DeepMind @GoogleDeepMind]

Can machine learning explain the world at the quantum level? New in @ScienceMagazine, with the code also released, our team shows that neural networks can radically improve density functionals with #DM21.

Read more: https://t.co/0moH36j4wk
Code: https://t.co/0dsrsh2OQt 1/

[Google DeepMind @GoogleDeepMind]

A new paper in @Nature details how machine learning was used to make significant new discoveries in pure mathematics by guiding the intuition of some of the world’s top mathematicians from @Sydney_Uni and @OxUniMaths: https://t.co/yWca5JOrHs

Paper: https://t.co/AvX78jY9zC 1/

[Google DeepMind @GoogleDeepMind]

Many single protein chains can be accurately predicted using AlphaFold, but multi-chain protein complexes remain challenging. Introducing AlphaFold-Multimer, a model that can predict the structure of multi-chain protein complexes with increased accuracy https://t.co/KMyN87wgDE 1/

[Google DeepMind @GoogleDeepMind]

Published today in @naturemethods together with colleagues from @calico: Enformer - a transformer model that has led to greatly increased accuracy in predicting gene expression from DNA sequence.

Blog: https://t.co/21zIewaiKR
Paper: https://t.co/Bg1gPp4AWe 1/