BREAKING: OpenAI introduces new o-series models

o3 and o4-mini

OpenAI claims that these are models that can produce novel and useful ideas.

Here is all you need to know: They are rolling out starting today on ChatGPT and APIs.

These reasoning models have gotten better at using internal tooling to solve very complex tasks.

And they are getting way better at it.

NEW: Google announces Agent2Agent

Agent2Agent (A2A) is a new open protocol that lets AI agents securely collaborate across ecosystems regardless of framework or vendor.

Here is all you need to know: Universal agent interoperability

A2A allows agents to communicate, discover each other’s capabilities, negotiate tasks, and collaborate even if built on different platforms. This enables complex enterprise workflows to be handled by a team of specialized agents.

Llama 4 is here!

- Llama 4 Scout & Maverick are up for download
- Llama 4 Behemoth (preview)
- Advanced problem solving & multilingual
- Support long context up to 10M tokens
- Great for multimodal apps & agents
- Image grounding
- Top performance at the lowest cost
- Can be served within $0.19-$0.49/M tokens LMArena ELO score vs. cost

"To deliver a user experience with a decode latency of 30ms for each token after a one-time 350ms prefill latency, we estimate that the model can be served within a range of $0.19-$0.49 per million tokens (3:1 blend)"

Prompt Engineering is NOT dead!

If you develop seriously with LLMs and are building complex agentic flows, you don't need convincing about this.

I've built the most comprehensive, up-to-date course on prompting LLMs, including reasoning LLMs.

4 hours of content! All Python! Check it out if you're building AI Agents or RAG systems -- prompting tips, emerging use cases, advanced prompting techniques, enhancing LLM reliability, and much more.

All code examples use pure Python and the OpenAI SDKs. That's it!

NEW: OpenAI announces new tools for building agents.

Here is everything you need to know: OpenAI has already launched two big agent solutions like Deep Research and Operator.

The tools are now coming to the APIs for developers to build their own agents.

A Few Tokens Are All You Need

Can you cut the fine-tuning costs of an LLM by 75% and keep strong reasoning performance?

A new paper from the Tencent AI Lab claims that it might just be possible.

Let's find out how: The First Few Tokens

It shows that all you need is a tiny prefix to improve your model’s reasoning—no labels or massive datasets are required!

Uses an unsupervised prefix fine-tuning method (UPFT)—only requiring prefix substrings (as few as 8 tokens) of generated solutions.

Say goodbye to Chain-of-Thought.

Say hello to Chain-of-Draft.

To address the issue of latency in reasoning LLMs, this work introduces Chain-of-Draft (CoD).

Read on for more: What is it about?

CoD is a new prompting strategy that drastically cuts down verbose intermediate reasoning while preserving strong performance.

NEW: Sakana AI introduces The AI CUDA Engineer.

It's an end-to-end agentic system that can produce highly optimized CUDA kernels.

This is wild! They used AI to discover ways to make AI run faster!

Let's break it down: The Backstory

Sakana AI's mission is to build more advanced and efficient AI using AI.

Their previous work includes The AI Scientist, LLMs that produce more efficient methods to train LLMs, and automation of new AI foundation models.

And now they just launched The AI CUDA Engineer.

NEW: Google introduces AI co-scientist.

It's a multi-agent AI system built with Gemini 2.0 to help accelerate scientific breakthroughs.

2025 is truly the year of multi-agents!

Let's break it down: What's the goal of this AI co-scientist?

It can serve as a "virtual scientific collaborator to help scientists generate novel hypotheses and research proposals, and to accelerate the clock speed of scientific and biomedical discoveries."

BREAKING: xAI announces Grok 3

Here is everything you need to know: Elon mentioned that Grok 3 is an order of magnitude more capable than Grok 2.

Introducing... Agent Leaderboard!

Many devs ask me which LLMs work best for AI agents.

The new Agent Leaderboard (by @rungalileo) was built to provide insights and evaluate LLMs on real-world tool-calling tasks—crucial for building AI agents.

Let's go over the results: 1️⃣ Leader

After evaluating 17 leading LLMs across 14 diverse datasets, here are the key findings:

Google's 𝗚𝗲𝗺𝗶𝗻𝗶-𝟮.𝟬-𝗳𝗹𝗮𝘀𝗵 leads with a 0.94 score at a remarkably low cost.

OpenAI Introduces Operator & Agents!

Here is everything you need to know: Operator is a system that can use a web browser to accomplish tasks.

Operator can look at a webpage and interact with it by typing, clicking, and scrolling.

It's available as a research preview. Available in the US for Pro users. Available to Plus users later.

Goodbye web scrapers!

Say hello to /extract by @firecrawl_dev

Just write a prompt and get the web data you need!

It doesn’t get any simpler than this. The /extract endpoint is simple to use. Provide a prompt and a schema and retrieve any data you need from a website.

I’ve added the /* to the URL to find and extract information across the entire website.

The endpoint can return up to thousands of data points at once.

The DeepSeek-R1 paper is a gem!

Highly encourage everyone to read it.

It's clear that LLM reasoning capabilities can be learned in different ways.

RL, if applied correctly and at scale, can lead to some really powerful and interesting scaling and emergent properties.

There is more to RL than meets the eye!

Here is my breakdown of the paper along with a few tests: youtu.be/3GlFd3doO3U?si…

The multi-state training might not make sense initially but they provide clues on optimizations that we can continue to tap into.

Data quality is still very important for enhancing the usability of the LLM.

Unlike other reasoning LLMs, DeepSeek-R1's training recipe and weights are open so we can build on top of it. This opens up exciting research opportunities.

About the attached clip: the previous preview model wasn't able to solve this task. DeepSeek-R1 can solve this and many other tasks that o1 can solve. It's a very good model for coding and math. When DeepSeek said "on par with OpenAI-o1" I thought they were just hyping. But based on my tests, it's clearly not so.

Wanted to add that DeepSeek-R1 got all of the hard tasks from the OpenAI LLM reasoning blog post correct for me. This is wild and totally unexpected! The only task where it failed (i.e., crossword puzzle) o1 also fails.

Agents Overview

Great write-up on Agents by Chip.

Here are my takeaways: 🤖 Agents Overview

An AI agent is made up of both the environment it operates in (e.g., a game, the internet, or computer system) and the set of actions it can perform through its available tools. This dual definition is fundamental to understanding how agents work.

Google recently published this great whitepaper on Agents.

2025 is going to be a huge year for AI Agents.

Here's what's included:

- Introduction to AI Agents
- The role of tools in Agents
- Enhancing model performance with targeted learning
- Quick start to Agents with LangChain
- Production applications with Vertex AI Agents

Great place to start learning about AI Agents. kaggle.com/whitepaper-age…

Summary of today's OpenAI announcement (Day 9):

- o1 is launching out of preview in the API
- support for function calling, structured output, and developer messages
- reasoning_effort parameter to tell the model how much effort to spend on thinking
- vision inputs in the API is here too Visual inputs with developer message (this is a new spin to system message for better steering the model) inside of the OpenAI Playground

Summary of today's OpenAI announcement:

- introduces reinforcement fine-tuning (RFT) of o1
- tune o1 to learn to reason in new ways in custom domains
- RFT is better and more efficient than regular fine-tuning; needs just a few examples

1/n How it looks in the dev platform. Examples show how to select RFT on o1-mini

That's right! It's a huge week for small language models (SLMs)

Few new SLMs on my radar: Mistral NeMo

Highlights:
- Introduced by Mistral + NVIDIA
- Apache 2.0 license
- outperforms Gemma 2 9B and Llama 2 8B
- multilingual capabilities
- efficient tokenizer (Tekken)

https://x.com/MistralAI/status/1813947930455499200

JUST IN: Google DeepMind releases Gemma, a series of open models inspired by the same research and tech used for Gemini.

Open models fit various use cases so this is a very smart move from Google.

Great to see that Google recognizes the importance of openness in AI science and technology.

There are 2B (trained on 2T tokens) and 7B (trained on 6T tokens) models including base and instruction tuned versions. Trained on a context length of 8192 tokens.

Commercial use is allowed.

These are not multimodal models but based on the reported experimental results they appear to outperform Llama 2 7B and Mistral 7B.

I am excited about those MATH, HumanEval, GSM8K, and AGIEval results. These are really incredible results for a model this size.

Excited to dive deeper into these models. The model prompting guide is dropping soon. Stay tuned! Blog:

Technical report: blog.google/technology/dev…
storage.googleapis.com/deepmind-media…

Gemini is here!

Google DeepMind just announced Gemini, their largest and most capable AI model.

A short summary of all you need to know:

1) What it is - Built with multimodal support from the ground up. Remarkable multimodal reasoning capabilities across text, images, video, audio, and code. Nano, Pro, and Ultra models are available to support different scenarios such as efficiency/scale and support complex capabilities.

2) Performance - The results on the standard benchmarks (MMLU, HumanEval, Big-Bench-Hard, etc.) show improvement compared to GPT-4 (though not by a lot). Still very impressive!

3) Outperforming human experts - They claim that Gemini is the first model to outperform human experts on MMLU (Massive Multitask Language Understanding), a popular benchmark to test the knowledge and problem-solving abilities of AI models.

4) Capabilities- Gemini surpasses SOTA performance on a bunch of multimodal tasks like infographic understanding and mathematical reasoning in visual contexts. There was a lot of focus on multimodal reasoning capabilities with the ability to analyze documents and uncover knowledge that's hard to discern. The model capabilities reported are multimodality, multilinguality, factuality, summarization, math/science, long-context, reasoning, and more. It's probably one of the most capable models by the looks of it.

5) Trying it out - Apparently, a fine-tuned Gemini Pro is available to use via Bard. Can't wait to experiment with this soon.

6) Availability - Models will be made available for devs on Google AI Studio and Google Cloud Vertex AI by Dec 13th.

blog:

technical report:
Here is the model verifying a student's solution to a physics problem. Huge implications in education. Will be taking a very close look at applications here.