https://t.co/Hwj5DQeb6n

Which AI framework do you use?. If your option is Others, which one?. And Why?. Please RT @DSPyOSS @LangChainAI @crewAIInc @pydantic

@aidenybai it's a good deal sir

Why guess when you can know?

sometimes I think procedurally, sometimes probabilistically; the trick is to use the right one at the right time.

Here's Emil's 17 step guide to how he used VS Code agent mode plus Claude 3.7 Sonnet, Gemini Pro 3 and Claude Opus to build the new library - it's a fantastic case study in using LLMs for serious, prediction quality code (vibe engineering, not vibe coding)

@ryan_x_charles Yes and this has pretty profound implications. Running it in a loop through the night (for instance) now means you'll wake up with a better codebase. We still need to figure out what would be the right system/harness to make it put its efforts and attentions to places that are

Imagine if this code cell could run and print the results below, in e-ink. That would be a true notebook! Imagine sketching, writing, maybe some dictating—all with just the right amount of AI assistance. That's probably the final destination for me. Where I want to be when I

Before Opus 4.5, the more you ran LLMs on a codebase, the more brittle it was getting. Now it slopes upward

it seems like we are entering the phase where a harness that reliably make models work overnight will be very very valuable.

A year ago, we verified a preview of an unreleased version of @OpenAI o3 (High) that scored 88% on ARC-AGI-1 at est. $4.5k/task Today, we’ve verified a new GPT-5.2 Pro (X-High) SOTA score of 90.5% at $11.64/task This represents a ~390X efficiency improvement in one year

I am always amazed how good LLMs are at guessing what I mean despite very big and horrible typos that would trip up many humans (my future self included). This seems like somewhat of a bug. Similar to how they can read text that is very messed up with letters replaced by

fwiw: i think its too early for us to estimate the relevant probabilities with any level of confidence required for us to act on that information. It thus seems like a distraction to me. we should thus not slow down anybody trying to automated science 100.1%. We merely need to

Here’s 60 minutes of nonstop package sorting… boring enough for you? https://t.co/wQRH48OCu6

I think this is the core believe from Julian that must be discussed: "I love science, and I am afraid of a future where we are pushed back into the dark ages because we can no longer contribute to science. Human agency, including in creative processes, is vital and must be

This is incorrect. LLMs can call tools to get info and change about the outside world, including viewing videos, moving robotic arms, etc. The pic below shows a minimal falsifying example. The value of this number squared is new information—it's never been documented before.

Anthropic is donating the Model Context Protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation. https://t.co/DOt9Wsv2ip

i link feature engineering..

We are donating MCP to the newly created Agentic AI Foundation. I am thrilled that we found a way to ensure that MCP will always remain neutral. Our commitment to MCP remains the same. We continue to be deeply involved and help steer the ship alongside other core maintainers

Harness harness harness. Claude code's harness is wonderful Cursors' harness is terrible Poetiq harness pushes Gemini into new territory on arc agi Copilot harness is the difference between copilot and chatgpt... The harness mattered a whole lot. Harness engineering is on!

full stack developer in 2025 be like

Opus 4.5 is too good to be true. I think we've reached the "more than good enough" level; everything beyond this point may even be too much.