The Pro series of GPT models continues to be really impressive for one-shot problems that involve complex calculations, references, or sophisticated ideas. It isn't particularly fast (or particularly pleasant) but GPT-6 Pro gives that "ask one question of...

High-engagement X post with 204 visible interactions on frontier model behavior.

Bun AoT in an Express node:sqlite REST API 40% less memory 14% lower cpu No JIT warmup

High-engagement X post with 168 visible interactions on AI developer tooling.

The OpenAI robotics result where Astra decimated specialized robotics models is deeply surprising and has greater implications for ML I think. it implies something, I don’t know what: that transfer is increasing with scale, or some fundamental algo shift,...

High-engagement X post with 194 visible interactions on frontier model behavior.

SITUATION DETECTED: Elon Musk says that Colossus 2 will be running 1.2 million NVIDIA Blackwells by the end of 2026.

High-engagement X post with 250 visible interactions on AI compute and inference.

Someone else who uses one of the applications I created has already implemented 150...

I'm now building an "improvements" feature into all of my personal vibecoded applications: a freeform text input that's hooked up to an agent that yolo implements whatever is submitted. It's an interesting experience -- the activation energy barrier to...

High-engagement X post with 340 visible interactions on AI misuse, safety, or security.

GPT-6.1 sol is better than GPT-6 astra low key. just you can understand it, it's fast, and honestly mostly as smart

High-engagement X post with 190 visible interactions on frontier model behavior.

Small tip in case you aren’t doing this already: Ask your agent to make your sites a chrome app, so you get a desktop and android app container for free. I’ve been doing this for all my little vibe coded tools, so useful.

High-engagement X post with 344 visible interactions on AI signal.

There are roughly 1T tokens of Opus 5.5 run through OpenRouter every day Almost half of that is run in Claude Code! Why the heck are people running it at API rates when on any Claude plan it's 66x subsidized?

High-engagement X post with 260 visible interactions on AI developer tooling.

Brilliant paper on agent memory from FAIR at Meta. (bookmark it) A single global history is the usual way to let agents learn from past runs. When tasks are interleaved, it gains almost nothing over no memory at all, an average of +0.015 in normalized...

High-engagement X post with 166 visible interactions on AI developer tooling.

Banger paper from IBM on decision gates for agents. They add a runtime gate to tool-using agents and keep the models and tools the same. JevAct is a check that sits between the agent and its tools. An LLM first turns the domain policy into code checks and...

High-engagement X post with 196 visible interactions on frontier model behavior.

Deterministic tests > Written instructions Anything that should be checked automatically should have a test case: • Checking behavior → Unit and integration tests • Checking formatting → Formatters and linters • Checking types → Static type checkers •...

High-engagement X post with 230 visible interactions on AI developer tooling.

Oh good news!! Zhipu GLM-5.5 has entered post-training!! Internal 32B/122B/256B models being merged into the next-generation base. We will be def getting two models, a flash and a pro one. GLM-5.5-flash is what most of local AI is looking forward for, if...

High-engagement X post with 183 visible interactions on frontier model behavior.

For the first time, working at Redis with agents is becoming a real struggle because of the many security stops I get while working at non security things. It is a real shame as system programming is one of the silver bullets of automatic programming.

High-engagement X post with 153 visible interactions on AI misuse, safety, or security.

"Towards Looped Models Done Right" Not all ways of looping Transformers work equally well. This paper finds that looping only the middle layers performs better than repeatedly looping the entire Transformer. The researchers compare different loop...

High-engagement X post with 118 visible interactions on frontier model behavior.

When doing security research it's important to run it securely to avoid impact on your machine. Use NEAR AI RED Sandbox to leverage uncensored models and isolated microVM for all tool calls.

High-engagement X post with 209 visible interactions on AI misuse, safety, or security.

LangChain used to be framework company but now ships its own harness called Deep Agents,...

AI Agent framework vs harness. A framework is a library you import to build an agent. ADK, LangGraph, and CrewAI give you the pieces like agent classes, tool interfaces and model adapters. you still decide how the loop runs, what goes into context and how...

High-engagement X post with 184 visible interactions on AI developer tooling.

🚨 Train Your Own LLM Using A Text Prompt - start with a text prompt - customize based on your use-case - run LLMs locally Try it on Abacus AI SuperComputer

Official or high-signal @bindureddy post with 157 visible interactions on AI signal.

Just became aware that when cognition tested Gemini 4 argon here, they used Devin cli, while Astra and opus both got their first party harness I do think you should test models in their harness, but it’s a bit unfair Gemini didn’t get tested in anti...

High-engagement X post with 175 visible interactions on frontier model behavior.