Tien Dat Nguyen @TienDat104

We've been defining the actual frontier of AI for science: - created the most powerful agents and generative models of our biomolecular world - coupled with the most accurate models of state - proved again and again ability to break novel ground in drug design >>> now I'm looking for those who want to push frontier LLMs even further, in an environment where you see the impact on real science the next day 1/2

Really happy to share our new paper on using AlphaEvolve for mathematical exploration at scale, written with Javier Gómez-Serrano, Terence Tao, and @GoogleDeepMind's Bogdan Georgiev. We tested it on 67 problems and documented all our successes and failures. 🧵

For 10 years, Google has worked to accurately read the operating manual of all life on Earth — the genome. Our AI tools are now used by partners for real-world challenges from improving healthcare to biodiversity conservation. Check out the key milestones ↓

Introducing NotebookLM for arXiv papers 🚀 Transform dense AI research into an engaging conversation With context across thousands of related papers, it captures motivations, draws connections to SOTA, and explains key insights like a professor who's read the entire field

Fantastic to see Genie 3, our state-of-the-art world model, featured in @TIME's 2025 Best Inventions. From a single image or text prompt to an entire playable world, it’s the future of AI and entertainment. So proud of @jparkerholder @shlomifruchter & the team - huge congrats!

Google DeepMind @GoogleDeepMind

10 months ago

We’re proud to announce that Genie 3 has been named one of @TIME’s Best Inventions of 2025. Genie 3 is our groundbreaking world model capable of generating interactive, playable environments from text or image prompts. Find out more → goo.gle/3KGqiYa

95 266 2K 355K 264

Meanwhile, AGI will in fact get better by simply adding more *compute*. It will not be bottlenecked by the availability of human-generated text.

Tri Dao (creator of FlashAttention) says there are 3 kinds of inference we will need to optimize for: > traditional chatbot workloads w/ fast enough to feel responsive but not instantaneous, to maintain a natural user experience > low-latency ultra-fast inference for highly interactive applications like coding assistants (e.g., Claude Code) or agentic tasks, where users pay a premium to stay in flow state and avoid interruptions > maximum throughput, large-batch size: synthetic data generation (e.g., creating vast amounts of training data from expert seeds) and RL training rollouts (e.g., sampling numerous trajectories to evaluate agent policies)

Thinking, Searching, and Acting A reflection on reasoning models. It's easy to fixate on the "thinking" that gave reasoning models their name, but just over a year out from o1-preview's release by OpenAI, the core primitives that make up models today has expanded. Searching and executing tools make up for their deficiencies as probabilistic tools with outdated information in their parameters. Together, these three actions will act as the foundation of the systems we use for years, and the engineer aspects of them matter just as much as getting precisely the right model weights.

Interconnects @interconnectsai

10 months ago

Thinking, Searching, and Acting A reflection on reasoning models. interconnects.ai/p/thinking-sea…

0 9 46 44K 37

Academic research is like neural net training. 99% of papers (weights) end up useless. But the useless ones needed to be there to be pruned / distilled to the 1% of important papers, just like you need to train a big NN before pruning / distilling to a small one. You can’t prune before training, and you can’t remove all “useless” papers without harming the development of the useful papers. Academic research is overparameterized.

Nice job! Open research / open source accelerates progress.

DeepSeek @deepseek_ai

2 years ago

🚀 DeepSeek-R1 is here! ⚡ Performance on par with OpenAI-o1 📖 Fully open-source model & technical report 🏆 MIT licensed: Distill & commercialize freely! 🌐 Website & API are live now! Try DeepThink at chat.deepseek.com today! 🐋 1/n

2K 7K 35K 12.5M 10K

Little is known about how deep networks interact with structure in data. An important aspect of this structure is symmetry (e.g., pose transformations). Here, we (w/ @StphTphsn1) study the generalization ability of deep networks on symmetric datasets: arxiv.org/abs/2412.11521

We outperform Llama 70B with Llama 3B on hard math by scaling test-time compute 🔥 How? By combining step-wise reward models with tree search algorithms :) We show that smol models can match or exceed the performance of their much larger siblings when given enough "time to think" We're open sourcing the full recipe and sharing a detailed blog post 👇

Andrej’s tweet is the right way to think about it right now but I totally believe that in one or two years we will start relying on AI for very challenging decisions like diagnosing disease under limited information. Key thing to note here is that big decisions can be viewed as a tree of individual reasoning steps and RL on chain of thought seems like a feasible way for AI to do any single step pretty well and probably recover if there is a mistake In addition, with better scaffolding like improvements in retrieval, browsing, and long context management, AI will be able to leverage its inherent advantages over humans like not getting tired or distracted, having nearly infinite memory, and not being clouded by emotions. So i think we will reach the “magical AI feeling” soon :)

Andrej Karpathy @karpathy

2 years ago

People have too inflated sense of what it means to "ask an AI" about something. The AI are language models trained basically by imitation on data from human labelers. Instead of the mysticism of "asking an AI", think of it more as "asking the average data labeler" on the

545 2K 13K 1.8M 5K

I have a theory called "Matrix Attraction Theory (MAT)", saying tech is converging toward the universe in the movie Matrix. We are getting there closer & faster. Hi Neo, nice to meet you! :-)

Promoted to Senior Staff Research Scientist @GoogleDeepMind 🚀 Energised to push even stronger on exciting foundational research, while also showing off new applied work (soon 👀)! Now more than ever, this would not have been possible without all who supported me. Thank you ❤️

🚀 Excited to share our work on Encoder-only Next Token Prediction (ENTP)! While most successful LLMs are decoder-based, we asked: Can encoder-only TFs be used for next-token prediction? Yes! Moreover, ENTP might be better than decoder-only models!!! 😎

💡CLIP is the default choice for most multimodal LLM research. But, we know CLIP is not perfect. It is good at high-level semantics, but not for capturing fine-grained info. 🤩🤩 We present CLOC ⏰, our next-generation image encoder, with enhanced localization capabilities, and serves as a drop-in replacement for CLIP. 🚀🚀How to do that? We conduct large-scale pre-training with region-text supervision pseudo-labelled on 2B images. 🎁As a result, CLOC is indeed a better image encoder, not only for zero-shot image/region tasks, but also for multimodal LLM.

One of the most-crucial yet often-overlooked aspects of success in research (and life in general) is ensuring that you're optimizing for the right function. It's easy to fall into the trap of chasing some reward without taking the time to examine whether it aligns with your personal goals and values. Consider these two researchers: - Researcher A is primarily motivated by publishing papers and gaining recognition. They optimize for quantity of publications and citation count. - Researcher B is driven by a desire to solve important problems and make a meaningful impact. They optimize for the potential applications of their work and its ability to advance the field. Both of these approaches have their merits and potential drawbacks. Researcher A might build an impressive CV, publish many papers, and gain visibility in the field, which could lead to more opportunities and collaborations. Researcher B might have fewer papers, but each paper is extremely impactful in its particular area because the researcher has spent years working on a breakthrough despite not being externally rewarded while doing so. The key insight here is that the function you choose to optimize dramatically influences not just your results, but also your day-to-day decisions and overall satisfaction. When you optimize for external metrics (like Researcher A), you might find yourself more attuned to the current trends and demands of the field. In contrast, optimizing for internal motivations (like Researcher B) might lead you to take on more challenging, potentially groundbreaking projects, even if they come with a higher risk of failure. Choosing the right function to optimize for isn't easy. It requires honest self-reflection and often means balancing competing priorities. By being thoughtful and honest about what you're optimizing for, you can ensure that your efforts are directed towards outcomes that truly matter to you, whether those are external achievements, internal fulfillment, or a carefully chosen balance of both.

We have 3 works accepted to @NeurIPSConf 2024, centered around developing intelligent agents in real-world environments. 2 of these works are led by @taoyds 's @XLangNLP . We strongly believe in this direction, and you can expect several future works from us in this area. 1/6

Victor Zhong (@hllo_wrld @UWaterloo @UWCheritonCS) works at the intersection of natural language processing and machine learning, aiming to teach machines to understand language better in order to generalize to new problems.