DeepMind Blog

Deepmind (http://deepmind.com/) comes to mind. It was acquired by Google though. Their vision it is "We combine the best techniques from machine learning and systems neuroscience to build powerful general‑purpose learning algorithms." AFAIK they work mostly on reinforcement learning coupled with Neural Networks

IBM, Facebook, Nuance, and most importantly: Google Google Google. "Google Search Will Be Your Next Brain"[1] is a good start, plus it's follow-up[2]. Some lay-of-the-land, least-to-most cool IMO: * Facebook's interest is in AI for social purposes (face recognition, NLP), rather than AGI (reason DeepMind turned them down). But they're still solidly in the field / a worthy target. * IBM has been interested in AGI for a very long time, hence Watson. NLP, neural nets, machine learning, etc. Dig into Watson, lots of fun here. * Nuance. Smaller, Dragon Speak. NLP & Markov Models primarily, but their creator Ray Kurzweil (now at Google) is the biggest AGI champion you'll see. He started this Singularity religion based on what AGI will bring to bear, see "The Singularity is Near" and "How to Create a Mind". Super fun stuff; inspiring follow, if not exaggerated. Both Nuance, and a ~similar company Wolfram Alpha (which Siri uses) would be really cool targets. * Google. G's always been interested in AI for NLP for translating web queries & mobile speech queries; speech synthesis for play-back; computer vision, robotics, & machine-learning (neural nets) for self-driving cars; you name it. If it's AI, G's on it. Importantly, one Geoff Hinton recently tweaked an AI algorithm which unblocked a clog in the field, and now shit is _throwing down_, catalyzing Hinton & Jeff Dean's G-internal "Google Brain" project. Shortly after, G acquired DeepMind[3], and _this_ my friend... this is the…

There is a huge overhang. Same techniques and sometimes even the same models can drive a lot of robotics coordination and decision making. Here is probably one of the more clear examples. A model trained on video and also robotic simulation/recording (probably ensembled with control systems/mobility models) will likely be at the core of how robots make decisions and plan. https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intel… This is way outside my area of expertise though. I've only dabbled in more classic robotics and control systems, but these multi-modal sequence to sequence models are highly adaptive and can effectively transfer learning across very different domains.

no matter your luck with hardware or your sysadmin skills, doing local inference for just yourself and/or to emulate typical usage (e.g. your coding workflow and deep research, etc.) is just very inefficient in current model architecture. to me, this is a "truck" approach to city driving as a single person who does not do furniture hauling every weekend. the sense of privacy and freedom is nice but online inference is more "economical" as multi-user load is more effectively served than going solo. maybe new architectures would make it effective to do text inference locally [1], till then great on you if you can spend car money on your setup. hope it is a great learning experience as well. [1] https://deepmind.google/models/gemma/diffusiongemma/

> AlphaEvolve’s procedure found an algorithm to multiply 4x4 complex-valued matrices using 48 scalar multiplications, improving upon Strassen’s 1969 algorithm that was previously known as the best in this setting. This finding demonstrates a significant advance over our previous work, AlphaTensor, which specialized in matrix multiplication algorithms, and for 4x4 matrices, only found improvements for binary arithmetic. > And in 20% of cases, AlphaEvolve improved the previously best known solutions, making progress on the corresponding open problems. https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent…

I was working some time ago on LLM assisted optimizations and algorithm discovery and this does not look like a novel idea. AlphaEvolve from google is evolutionary algorithm which uses LLMs for Idea generation following very similar loop: - https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent… - Open source implementation of the algorithm: https://github.com/algorithmicsuperintelligence/openevolve

Google Deepmind Page: https://deepmind.google/models/genie/ Try it in Google Labs: https://labs.google/projectgenie (Project Genie is available to Google AI Ultra subscribers in the US 18+.)

Literally the only thing people cared about was Audio Overviews. NotebookLM had launched months earlier without that feature and was immediately ignored. The reason people liked audio overviews was because the voice model was amazing. This guy and the labs team had nothing to do with that. It was developed by a research team at Google Research/GDM who got maybe 1% of the credit: https://deepmind.google/discover/blog/pushing-the-frontiers-... It’s amazing how these people still insist on denying the massive contribution the audio model made, which is apparent by the lack of credits or acknowledgements to the team anywhere public despite these self aggrandizing blogs and postcast appearances.

Lemme start by saying this is objectively amazing. But I just really wouldn't call it a breakthrough. We had one breakthrough a couple of years ago with GPT-3, where we found that neural networks / transformers + scale does wonders. Everything else has been a smooth continuous improvement. Compare today's announcement to Genie-2[1] release less than 1 year ago. The speed is insane, but not surprising if you put in context on how fast AI is advancing. Again, nothing _new_. Just absurdly fast continuous progress. [1] - https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation…

Odyssey Systems is six months behind way more impressive demos. They're following in the footsteps of this work: - Open Source Diamond WM that you can run on consumer hardware [1] - Google's Genie 2 (way better than this) [2] - Oasis [3] [1] https://diamond-wm.github.io/ [2] https://deepmind.google/discover/blog/genie-2-a-large-scale-foundation… [3] https://oasis.decart.ai/welcome There are a lot of papers and demos in this space. They have the same artifacts.

"Gemini Flash fails even for simple tasks." On the Gemini Flash page (https://deepmind.google/technologies/gemini/flash/), it claims to be 'best for fast performance on complex tasks.'. I always use Gemini Flash in my project for demos and testing, and it performs very well, if a project requires a large, expensive model to handle simple tasks, that could be an issue to users.

This is the go community saying a computer will never best human go players. We already have examples of a model finding more performant sorts [0], given the right incentives and time, and the right system for optimizing (LLMs trained on “average code” probably aren’t it) the computer will best us at creating things for the computer. Is “vibe coding” real today? Not in my experience, with even Claude code. My hand has to be firmly on the tiller, using my experience and skill to correct its mistakes and guide it. But I can see the current trajectory of improvement, and I’m sure it’ll get there. [0] https://deepmind.google/discover/blog/alphadev-discovers-faster-sortin…

As someone who has no skin in the game and is only loosely following this, there is a tool: https://github.com/google-research/circuit_training, the detractors claim to not be able to reproduce Google's results (what Dean is commenting on) with it, Google and 1-2 other companies claim to be using it internally to success (e.g. see the end of this article: https://deepmind.google/discover/blog/how-alphachip-transformed-comput…

It is officially "Google Deepmind" now, according to their website - https://deepmind.google

Meanwhile Google has already used AI for this exact stage of chip design with AlphaChip: https://deepmind.google/discover/blog/how-alphachip-transformed-comput… And they've responded to the absurd degree of skepticism that followed: "That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design" - https://arxiv.org/abs/2411.10053

In academy for example, including the publicly funded institute mentioned in the article (European Molecular Biology Laboratory’s European Bioinformatics Institute). Also, there is a blog page on the deep mind website where they cite "millions of researcher" around the world using alphafold [1]. Some of those institutions might be happy to develop the new alphafold. On the other hand, even in academy, scientific articles are not always the best examples of open and reproducible research [2,3]. And in computer science, it is common for articles to be more focused on algorithms than on implementations. [1]: https://deepmind.google/impact/meet-the-scientists-using-alphafold/ [2]: https://en.wikipedia.org/wiki/Reproducibility [3]: https://en.wikipedia.org/wiki/Replication_crisis

Probably worth mentioning Deepmind's SIMA if you haven't seen it yet [0]. It definitely pushes the state of the art for agent behavior in a typical game environment. Perhaps in the future we won't code NPCs any differently from player characters- they'll have access to the full action space, with long-term memory and natural language prompts guiding their behavior. [0]: https://deepmind.google/discover/blog/sima-generalist-ai-agent-for-3d-…

SIMA author here - SIMA is betting on simulations and games. We use real-world open-ended language, but games/sim obviously have simplified physics/graphics. In terms of alternative strategies, Google DeepMind also has an amazing robotics team with lots of fantastic work for real-world robotics - including multi-robot generalists, showing positive effects when co-training one agent or model on multiple environments/bodies. Their prior work was very inspirational to us in SIMA! https://deepmind.google/discover/blog/scaling-up-learning-across-many-…

It's worth remembering that AI is more than LLMs. DeepMind is still doing big stuff: https://deepmind.google/discover/blog/millions-of-new-materials-discov…

https://en.wikipedia.org/wiki/Frank_Herbert_bibliography#WorShip_novel… The whole series is worth experiencing as another view on many of his common themes. Post is prompted by https://www.deepmind.com/blog/rt-2-new-model-translates-vision-and-lan…

totally! MuZero is my favourite[1] The super cool thing about MuZero is that it learns the dynamics of the problem, i.e. you don't have to give it the rules of the game, which makes the algorithm very general. For example, DeepMind threw MuZero at video compression and found that it can reduce video sizes by 6.28% (massive for something like YouTube)[2][3]. Curious if anyone else knows examples of MuZero being deployed outside of toy examples? [1] https://arxiv.org/pdf/1911.08265.pdf [2] https://arxiv.org/pdf/2202.06626.pdf [3] https://www.deepmind.com/blog/muzeros-first-step-from-research-into-th… (edit s/Google/DeepMind)

The board game Diplomacy is the classic example of multiplayer competition, where there is explicit negotiation of alliances and the possibility of trust and betrayal at every turn. P.S. Recent success with AI for Diplomacy at DeepMind: https://www.deepmind.com/blog/ai-for-the-board-game-diplomacy

There was also a separate DeepMind announcement made by Hassabis: https://www.deepmind.com/blog/announcing-google-deepmind It seems that DeepMind has now gone from what had appeared to be a blue sky research org to almost a product group, with Google Research now being the primary research group. Jeff Dean's reputation has always been as an uber-engineer, not any kind of visionary or great leader, so it's not obvious how well suited he's going to be to this somewhat odd role of Chief Scientist both to Google DeepMind and Google Research. How things have changed since OpenAI was founded on the fear that Google was becoming an unbeatable powerhouse in AI!

This is a more up to date version of David Silver’s course and judging from the few lectures I watched, it is very good but if you haven’t, I recommend having a look at Silver’s lectures, too. He is an amazing teacher and it is a joy to watch him explain something. https://www.deepmind.com/learning-resources/introduction-to-reinforcem…

I doubt it: they may be behind in the LLM space but LLM's aren't everything. Don't forget Google has DeepMind: https://www.deepmind.com/. Its board/video game AI surely beats GPT, and its interactive agent / AlphaCode AI at least compares. They also have Imagen (https://imagen.research.google/) which is among the top image generators and one of the couple video generators out there. Also, it's only been a few months since ChatGPT and the LLM explosion. They are surely doing research on LLMs, they just haven't released it as a consumer tool like OpenAI has.

Introduction to Reinforcement Learning with David Silver https://www.deepmind.com/learning-resources/introduction-to-reinforcem… This classic 10 part course, taught by Reinforcement Learning (RL) pioneer David Silver, a popular resource for anyone wanting to understand the fundamentals of RL.

Yes, see DeepMind RETRO: > In our experiments on the Pile, a standard language modeling benchmark, a 7.5 billion parameter RETRO model outperforms the 175 billion parameter Jurassic-1 on 10 out of 16 datasets and outperforms the 280B Gopher on 9 out of 16 datasets. https://www.deepmind.com/blog/improving-language-models-by-retrieving-… Though, there hasn't been much follow-up research on it (or DeepMind is not publishing it). Annotated paper: https://github.com/labmlai/annotated_deep_learning_paper_implementatio…

Do not nerd snipe me! There's tons of juice left on the CPU side. Again I recommend Metamath Zero (which is somewhat of an opposite direction as HoTT), as it's able to check the entire Metamath proof database in a few hundred milliseconds. This is orders of magnitude faster than most other proof systems. If you did want to manipulate trees on the GPU, then my stack monoid work[1] might be a good basis for it. Of course, a whole nother direction is to use AI to develop proofs, which is a fascinating and growing area (see Holophrasm[2] and DeepMind work[3]). That of course runs on the GPU. [1]: https://arxiv.org/abs/2205.11659 [2]: https://arxiv.org/abs/1608.02644 [3]: https://www.deepmind.com/publications/proving-theorems-using-increment…

New ways of using books are coming up - you put 1TB of books on your SSD, index it for search and use a neural net on it. Ask a question or pose a task, the model retrieves passages and then computes the reply. For this to work what is needed is a blanket license for search and snippet use. Can't buy a whole library for such a sparse use case, ain't gonna read it in the classical sense. This approach can guarantee the generated text will not replicate the source material, just use it for reference, with citations. - DeepMind RETRO - https://www.deepmind.com/publications/improving-language-models-by-ret…

Look into RETRO, which greatly reduces the model’s tendency to confabulate by teaching it to query a document database known to be truthful, and justify its answers with specific references: https://www.deepmind.com/publications/improving-language-models-by-ret…