Multitask Unified Model (MUM) — our latest AI milestone — has the potential to transform how Google helps you with complex information tasks. #GoogleIO

Today is the start of a new era of natively multimodal AI innovation. Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality. Llama 4 Scout • 17B-active-parameter model with 16 experts. • Industry-leading context window of 10M tokens. • Outperforms Gemma 3, Gemini 2.0 Flash-Lite and Mistral 3.1 across a broad range of widely accepted benchmarks. Llama 4 Maverick • 17B-active-parameter model with 128 experts. • Best-in-class image grounding with the ability to align user prompts with relevant visual concepts and anchor model responses to regions in the image. • Outperforms GPT-4o and Gemini 2.0 Flash across a broad range of widely accepted benchmarks. • Achieves comparable results to DeepSeek v3 on reasoning and coding — at half the active parameters. • Unparalleled performance-to-cost ratio with a chat version scoring ELO of 1417 on LMArena. These models are our best yet thanks to distillation from Llama 4 Behemoth, our most powerful model yet. Llama 4 Behemoth is still in training and is currently seeing results that outperform GPT-4.5, Claude Sonnet 3.7, and Gemini 2.0 Pro on STEM-focused benchmarks. We’re excited to share more details about it even while it’s still in flight. Read more about the first Llama 4 models, including training and benchmarks ➡️ go.fb.me/gmjohs Download Llama 4 ➡️ go.fb.me/bwwhe9

Highlight reel of Charlie Munger spitting one banger after another.

BREAKING: Meta's Llama 4 Maverick just hit #2 overall - becoming the 4th org to break 1400+ on Arena!🔥 Highlights: - #1 open model, surpassing DeepSeek - Tied #1 in Hard Prompts, Coding, Math, Creative Writing - Huge leap over Llama 3 405B: 1268 → 1417 - #5 under style control Huge congrats to @AIatMeta — and another big win for open-source! 👏 More analysis below⬇️

We are excited to share Reka Flash ✨, a new state-of-the-art 21B multimodal model that rivals Gemini Pro and GPT 3.5 on key language & vision benchmarks 📈. We've trained this model from scratch and ground zero with a small (but amazingly capable team 🧙‍♂️) and relatively finite resources. We're amazed at how strong it is 🦾. I'm proud of our financially optimal LLM team. Abandoning one's comfort zone is surely difficult and having to redo things from scratch is often scary & daunting. Many things in the wilderness don't work from the get go and it was often a huge pain in the neck 😢. I should write a separate post someday of how much we have we've had to rebuilt (and suffered 🤣). Everything from robust training infra, proper (human) evaluation pipelines and proper RLHF setups. I am thankful of the crazy talented team we have here ☺️. Meanwhile, our largest most capable model Reka-Core is finishing soon and we're already very excited by early results 📈. More to come very soon! 9 months in. Excited to be back at the frontier 🔥. Check out our blogpost here: reka.ai/reka-flash-an-effici…

New in tf-nightly: the NumPy API. - GPU and TPU-accelerated NumPy code - Interoperable with the rest of the TF ecosystem Documentation: tensorflow.org/api_docs/pyth…

Daniel Dennett (28th March 1942 - 19th April 2024) RIP Dennett on how to criticise wisely

Today we launch Asimov. Asimov is our code research agent that is best-in-class in codebase comprehension. It is built for teams, built for enterprises, and built to remember. We use it everyday to accelerate our velocity and streamline distributed ops. Link below to sign up for waitlist.

Scaling Up Models and Data with and deepai.org/publication/scali… by Adam Roberts et al. including @afrozenator, @ChenZhuo19 #NeuralNetwork #OpenSource

How to Work Hard: paulgraham.com/hwh.html

Today we're releasing all Switch Transformer models in T5X/JAX, including the 1.6T param Switch-C and the 395B param Switch-XXL models. Pleased to have these open-sourced! github.com/google-research/t… All thanks to the efforts of James Lee-Thorp, @ada_rob, and @hwchung27

Now we have Terence Parr speaking about the role of visualization in ML research

Our @RekaAILabs Tech Report / Paper is out! 🔥 Tech reports with completely no information are kinda boring so we’re revealing some interesting information on how we train our series of Reka models including tokens, architecture, data & human evaluation workflows. 😃 We tried our best to give a behind-the-scenes experience 😊. In particular, if you enjoyed my previous blog post about training LLMs in the wilderness, there’s a dedicated section on that in this report! 🌴 We can’t disclose literally everything but we tried our best to make it interesting, I promise. 🙏 Here’s a rundown summary of some of the highlights. 🔹Edge and Flash are outrageously strong 7B and 21B models. They are trained on 4.5-5T tokens in total. Also, they have been improved significantly since their first public appearance! They outperform many popular faces. Some data mixture information is in the report. 🔹We discuss our internal human evaluation workflow, prompt distribution, and how we use Core for model development and automatic evaluation. 🔹We describe our infrastructure setup for training large models, quantifying node failures, and report loss curves for training our models. 🔹Aside from the hardware lottery, we also show how this affects node stability across time. Once we were told our cluster became less stable because there were "big guys" moving things around the data center. 😅 On performance which you might have already seen on other threads. 🔹Core approaches frontier-class models like Claude3 Opus and GPT4-V. It outperforms Claude3 Opus on third-party blind human evaluation for multimodal chat, outperforms Gemini Ultra on video QA, and is quite competitive to other frontier models on core text metrics. It also matches GPT4-V on MMMU! 🔹Core ranks #2 on our internal multimodal chat leaderboard, right after GPT4-V. On text, it ranks #3 just behind Claude Opus and GPT4 Turbo. Core outperforms GPT-4 (0613) on this ranking. This has been a focused and concentrated effort of a small team of ~20 people in the past 4 months (yes, we got access to 90%+ of our compute only late December last year! 🚀). This tech report tells our story. Enjoy! Happy to answer any questions in replies or DM! PS: it was nice writing in latex after one whole year! PPS: I had quite some fun writing this 😊. There's some puns and easter eggs and interesting tidbits in there. Trust me. 😏 Link: publications.reka.ai/reka-co…

🔥JAX meets Transformers🔥 @GoogleAI's JAX/Flax library can now be used as Transformers' backbone ML library. JAX/Flax makes distributed training on TPU effortless and highly efficient! 👉 Google Colab: colab.research.google.com/gi… 👉 Runtime evaluation: github.com/huggingface/trans…

The most dramatic optimization to nanoGPT so far (~25% speedup) is to simply increase vocab size from 50257 to 50304 (nearest multiple of 64). This calculates added useless dimensions but goes down a different kernel path with much higher occupancy. Careful with your Powers of 2.

Using our Multitask Unified Model, or MUM, as introduced at #GoogleIO, we were able to identify 800+ variations of vaccine names in 50+ languages in seconds, making it possible to provide timely, high-quality information about COVID-19 vaccines worldwide. blog.google/products/search/…

We’re excited and proud to be opening up the Character.AI beta to the public! Character lets you create and talk to advanced AI (language tutors, text adventure games, celebrities, talking animals + more).

At age 25, you give your hedge fund manager $100K to manage, and he produces an annual return of 8%. Assuming a 1.5% management and 20% performance fee, by the time you retire at 65, you will have $764K. But the manager will have $1.24M (at zero initial investment!)

My favorite thing that @TylerCowen has ever written

If google ever started selling TPU hardware and released internal tooling, they'd MOG nvidia so bad. Just a trillion dollar company waiting to be built. most people don't realize how good JAX + TPUs + (other stuff) really is.

The False Promise of Imitating Proprietary LLMs Open-sourced LLMs are adept at mimicking ChatGPT’s style but not its factuality. There exists a substantial capabilities gap, which requires better base LM. arxiv.org/abs/2305.15717

Google engineer: AI is a serious risk to our business Dec 26, 2018

1/ Michael Mauboussin recently gave a great talk to Columbia Business School, discussing the revised and updated version of his book with Alfred Rappaport, Expectations Investing: Reading Stock Prices for Better Returns. We’ll share some of the highlights here

Saying college is a scam to those from poor/working class backgrounds actually does more harm than you all think it does

Congratulations on Maverick, looks like a great model!!

Sam Walton: Constraints are your friend