I am SUPER stoked! Chrome is finally getting agentic. 🤖 Google's latest update adds Gemini for AI summaries, multi-tab context, and task automation (yes, it’ll actually DO stuff for you). The browser is officially becoming an assistant. 💁 blog.google/products/chr...

91% of SEOs say clients/management now ask about AI search visibility - but only 35% have a strategy. It's early days, but leadership is watching. 👀 Survey via @aleyda.bsky.social 👇 hub.seofomo.co/surveys/stat...

Today, I'm thinking about how to create advanced prompts to monitor AI Mode results that set the stage for what the final question in a back-and-forth might be, where context matters more than the final query.

I hate it when someone is essentially right, but you still want them to lose. wandb.ai/byyoung3/ml-...

Google wants you in AI Mode faster ⏩: an “Ask Anything” box is showing in AI Overviews, sending users straight to AI results. 👇 www.seroundtable.com/google-ai-ov...

LLMs, lost attribution & agentic AI. 🎙️ I talked w/ Search Engine Land ahead of my SMX Advanced keynote on how SEO is evolving & the weirdest SEO gotcha I’ve hit yet. 👇 searchengineland.com/dave-davies-...

AI agents that read, summarize, and document a GitHub repo — end-to-end automation using CrewAI + @weightsbiases.bsky.social Weave for observability 👀. A great demo and tutorial on multi-agent orchestration 👇 wandb.ai/byyoung3/cre...

ChatGPT just got way more useful for devs: 🔎Deep Research now connects to GitHub. You can query repos 📚, analyze APIs, and break down code structures. o4-mini fine-tuning 🎛️ & GPT-4.1 nano access expanded too - w/ verification required. Details 👇 wandb.ai/byyoung3/ml-...

🚀 www.prnewswire.com/news-release...

Qwen3 just crushed AIME 2024 with 66.7% - 3x DeepSeek R1. 🤯 Fine-tune it with Unsloth, evaluate with @weightsbiases.bsky.social Weave, and toggle its reasoning mode. Wildly flexible open-source LLM. Read more 👇 wandb.ai/byyoung3/Gen...

New papers from @SFResearch & @Tsinghua_Uni suggest RL in LLMs may be overrated. 📊Simple filtering > complex RL 🍪RLVR ≠ new reasoning These finding are covered on the @weights_biases blog, and may be a game-changer for post-training strategies. 👇 wandb.ai/byyoung3/ml-...

Most agents today can’t talk to each other. A2A changes that. Agent2Agent is an open protocol for secure, cross-platform agent collaboration. Think: LangGraph x CrewAI in one workflow. No glue code. Full write-up + tutorial over on @weightsbiases.bsky.social blog 👇 wandb.ai/byyoung3/Gen...

👏Big congrats to @cohere.com on Embed 4 — a new multimodal embedding engine for enterprise AI. It handles 📚 128k token docs, 🌏 100+ languages, and real-world mess like legal PDFs & product decks. Built for 🤖 RAG, agents, and cross-lingual search. wandb.ai/byyoung3/ml-...

SEO in the agentic era isn’t theory - it’s here. Over on @searchengineland.bsky.social I explore 🔍 the impact agents are having on how (and where) we optimize, and what that means for your strategy. I even outline an agentic system I'm working on. Enjoy 👇 searchengineland.com/ai-agents-di...

I just published a GPT-4.1 quickstart using the OpenAI API over on the @weights_biases blog, including a Colab to get going fast. It includes full W&B Weave integration so you can track everything out of the gate. 👀 wandb.ai/onlineinfere...

How long until we start talking about ACO (Agent Card Optimization)? 🤔 google.github.io/A2A/#/docume...

🛒Retailers are using AI agents 🤖 to do a lot more than recommend products. This post from @weightsbiases.bsky.social breaks down a smart LLM-powered system for triaging customer emails and building real-time recs via vector search. 👇 wandb.ai/byyoung3/Gen...

LLMs don’t need to think + talk in the same pass. Retrieval Augmented Thinking (RAT) 🐀 splits reasoning from response - boosting transparency + control. DeepSeek, Claude, GPT-4o all in the mix. Code + Weave traces👇 wandb.ai/byyoung3/Gen...

Amazon Nova Reel 1.1 🎥 now supports 2-minute, multi-shot videos—with automated and manual modes. Great flexibility, now on AWS Bedrock. wandb.ai/byyoung3/ml-...

Llama 4 🦙 isn’t just open—it’s competitive. @weightsbiases.bsky.social tested it head-to-head with GPT-4o on ChartQA using Weave Evaluations. Maverick scored higher on multimodal accuracy and costs a fraction to run. Here's how they did it + the code 👇 wandb.ai/byyoung3/Gen...

🚶‍♂️ Gemini 2.5 Pro Experimental is a big step up. 🧠 Multimodal input, 1M-token context, native code execution, and better math/code reasoning. @weightsbiases.bsky.social evaluated it against Flash 2.0 on AIME problems using Weave; in a tutorial you can do yourself. Results? 👇 wandb.ai/byyoung3/Gen...

Want your OpenAI Agent to explore files, analyze data, and log everything with Weave? 🧠 This walkthrough shows how to hook agents into external tools using the Model Context Protocol (MCP). 👇 wandb.ai/byyoung3/Gen...

I'm actually finding it worse. Anyone else? Am I just using it wrong? :D

ARC-AGI-2 is here. 🔥 The benchmark that flips the script—easy for humans, brutal for AI. 🤖 If you're chasing AGI, this is where you prove your model can reason 🤔, not just pattern-match. 🤑 $1M in prizes. Hosted on Kaggle. wandb.ai/byyoung3/ml-...

MCP is like a USB port for AI—letting LLMs interact with external data, tools, and APIs without custom integrations. Anthropic’s new protocol makes AI applications more scalable, interoperable, and easier to manage. Enjoy a hands-on guide for building an MCP server.👇 wandb.ai/byyoung3/Gen...

🚨 AI agent hijacking is a growing threat. US AISI's latest research using AgentDojo found that even top models like Claude 3.5 Sonnet were tricked 81% of the time with new attack strategies. Security evaluations need to evolve—attackers already are. Read the full story 👇 wandb.ai/byyoung3/ml-...

🔌 "A USB-C port for AI" – that's how Anthropic describes the Model Context Protocol (MCP). MCP standardizes how LLMs connect to tools & data, making AI assistants more powerful. Could this be the ODBC moment for AI? Learn more 👇 wandb.ai/onlineinfere...

But don't worry, Elon will let us all know if there's a conflict of interest. www.msn.com/en-ca/news/w...

Google is rolling out AI Mode to Google One AI Premium subscribers in the US. 👉 x.com/sundarpichai...

For anyone interested in getting started with the new GPT-4.5 via the API, I just updated my GPT-4o quickstart to 4.5. You can find it on the @weightsbiases.bsky.social blog here 👇 wandb.ai/onlineinfere...

👀 Want to build AI agents that can research, analyze, and summarize in real-time? @weightsbiases.bsky.social takes a look at CrewAI - a framework for multi-agent systems - and walks through how to build an AI agent. Full tutorial here 👇 wandb.ai/byyoung3/Gen...

Google Ads almost tricked me. ;)

🚨 AI safety just got a lot easier. @weightsbiases.bsky.social Weave Guardrails helps devopers detect toxicity, bias, hallucinations & more - before bad outputs hit users. Pre-built scorer models + real-time safeguards = better AI apps. Try it now 👇 wandb.ai/site/guardra...

It was a good evening. :) www.nhl.com/video/can-us...

🇰🇷 South Korea has accused DeepSeek of sharing user data with ByteDance. Their app skyrocketed 🚀 in popularity, but is under fire 🚒 for potential data transfers. It's already banned in Korean app stores, and more countries are watching closely. Full story 👇 wandb.ai/byyoung3/ml-...

🚀 xAI’s Grok-3 is here, and is topping benchmarks. ✅ Trained with 100K H100 GPUs in 122 days ✅ #1 in Chatbot Arena, beating Gemini-2 Flash & DeepSeek-V3 ✅ 93 on AIME (math), 79 on LCB (coding) Self-correcting, reasoning, and already in Deep Search & X Premium+ More here 👇 wandb.ai/byyoung3/ml-...

Elon Musk announced that Grok 3, which he described as the “smartest AI on Earth”, will be released with a live demo on Monday, February 17, 2025, at 8:00 PM PT

🚀 Can we push GPT-4o’s reasoning beyond state-of-the-art? Brett Young tested fine-tuning and budget forcing on AIME math problems and it turns out, we can! The full breakdown, and code to run your own experiments can be found on the @weightsbiases.bsky.social blog here 👇 wandb.ai/byyoung3/Gen...

WTF? I'm not sure what's more pathetic, the lack of any spine from Google or that there are people who were triggered by the notion of people learning about history.

The future is coming VERY VERY fast.

Well, they did it. OpenAI got $200/mth out of me. :D

I'm super-excited by the new 🍎 Apple researcher show that self-play RL can train AI to drive without human data! Their model, GIGAFLOW, ran 1.6B km in sim, outperforming specialist models on real-world benchmarks. This may dramatically speed up the road to AV. 👇 wandb.ai/byyoung3/ml-...

🪄 Can a decoding trick make a smaller open model match OpenAI’s o1-preview? Budget forcing - forcing an LLM to "keep thinking" - boosted s1-32B’s accuracy 13% at test time with no extra training. Here’s what Brett Young found, published on @weightsbiases.bsky.social 👇 wandb.ai/byyoung3/Gen...

A reminder for my US friends: Less than 1% of fentanyl 🚚 entering the US 🇺🇸 is coming from Canada 🇨🇦 and only 1.5% of illegal migrants entering the US are coming from Canada each year. (Based on data from 2024) 1/3

🎉 OpenAI’s o3-mini is here: faster, smarter & cheaper than o1-mini. Want to get the API up running in Python in ~5 min? Here’s my latest quickstart over on @weightsbiases.bsky.social, with W&B Weave for tracking. 👇 wandb.ai/onlineinfere...

T̶h̶e̶ ̶O̶3̶ ̶F̶a̶m̶i̶l̶y̶ o3-mini family

Well that's a bit of walkin' around money. 🚶 www.cnbc.com/2025/01/30/o...