Tweets

"The cost of intelligence has been falling dramatically over the last couple of years.

The best way to think about that is that the cost for each tera of intelligence has been dropping. One fact on that is that you can get intelligence at the level of GPT-4 for over 100x cheaper than GPT-4 was at launch."

Micah breaks down one of the big trends he has been seeing at Artificial Analysis on the cost of intelligence significantly dropping.

@_micah_h

"On 𝜏²-Bench Telecom, it's cheaper to run more expensive models like a GPT-5 compared to some smaller open source models because GPT-5 for instance got to the answer faster and so was able to resolve the customer's query faster in fewer turns.

I think that's what we're getting to. I think number of turns is gonna be a metric we're going to be talking about a lot more."

George breaks down the non-obvious reasoning why more expensive models (on a per token basis) may actually be cheaper to run than smaller open source models.

@grmcameron

Agent memory is still largely discussed as a retrieval problem.

“Memory in the Age of AI Agents” paper argues that framing is incomplete.

The paper formalizes agent memory as an evolving state that lives inside the agent loop, not just as an external store to be queried.

Memory is governed by three operators:
1. formation, which decides what gets written;
2. evolution, which determines how memory is consolidated, rewritten, or discarded
3. retrieval, which decides what gets surfaced back into the model at decision time.

Many existing agent memory systems place most of their emphasis on retrieval and rely heavily on static recall mechanisms. Without meaningful evolution, memory keeps growing, due to which redundant entries accumulate, and conflicts remain unresolved.

The result is often a growing record of past interactions rather than a curated and adaptive memory base.

One of the more subtle points in the paper is how it treats short-term versus long-term memory. These are not presented as separate architectural components. Instead, temporal roles emerge from how frequently formation, evolution, and retrieval are invoked.

If formation happens frequently while evolution is rarely applied, the agent becomes verbose but brittle. If evolution is overly aggressive, stability can break down. This perspective helps explain many long-horizon failure modes observed in practice.

The direction suggested is not simply better retrieval. It is the design of memory evolution policies that decide when experience should be compressed, merged, rewritten, or forgotten. This is where memory moves beyond storage and begins to shape agent behavior over time.

Paper: arxiv.org/abs/2512.13564

🚀 Automate your X posting with Agent Skills!

  Just built a skill that lets Agent post directly to X (Twitter) - regular posts with
  images AND long-form Articles (Markdown support!).

  ✨ Uses real Chrome browser via CDP - bypasses anti-bot detection
  📝 Preview mode before posting
 🖼️ Auto-uploads images from your Markdown to X Articles

  Based on @wshuyi's Playwright MCP implementation:github.com/wshuyi/x-artic…IIT

  Install:
npx add-skill jimliu/baoyu-skills

  Then just ask Claude: "Post this to X" or "Publish this article on X"

  GitHugithub.com/JimLiu/baoyu-s…aDE

We are starting to test ads in ChatGPT free and Go (new $8/month option) tiers.

Here are our principles. Most importantly, we will not accept money to influence the answer ChatGPT gives you, and we keep your conversations private from advertisers.

It is clear to us that a lot of people want to use a lot of AI and don't want to pay, so we are are hopeful a business model like this can work.

(An example of ads I like are on Instagram, where I've found stuff I like that I otherwise never would have. We will try to make ads ever more useful to users.)

We’ve acquired Torch, a healthcare startup that unifies lab results, medications, and visit recordings. Bringing this together with ChatGPT Health opens up a new way to understand and manage your health.

We're excited to welcome the Torch team to OpenAI @IlyaAbyzov, @elh_online, @jfhamlin, and Ryan Oman.

Answering a top request from our users, we’re introducing Personal Intelligence in the @GeminiApp. You can now securely connect to Google apps for an even more helpful experience.

Personal Intelligence combines two core strengths: reasoning across complex sources and retrieving specific details, e.g from an email or photo, to provide uniquely tailored answers.

It’s built with privacy at the center. You choose exactly which apps to connect, these connected app settings are off by default.

For AI to be truly useful, it needs to understand you. 
With Personal Intelligence, we’re beginning to solve this. With your permission, Gemini can now securely reason across your own data to answer questions that generic models simply can't - like suggesting plans based on travel dates in Gmail or your hobbies found in Photos. An exciting step towards a digital assistant that’s uniquely helpful to you.

To make a bit of an excuse for Microsoft: the world is just waking up to the fact that coding agents are general agents.

It’s bitter lesson adjacent: Writing and executing code will likely outperform years of handcrafting vertical-specific agents with expert knowledge.

Actually it might exactly map in bitter lesson: Program synthesis is a form of scalable search.

I did office hours with a startup yesterday that I ended up advising to (a) decrease their growth rate and (b) spend the money they've raised faster by hiring more people and paying them high salaries. Possibly the first time I've ever told a startup this.