Sam Knight

Currently fine tuning a language model for an enterprise and I can tell you - GPT 5.6 is better than fable in EVERY way.

I could spend all day making codex and claude code skills. Too addicting

Opus has become mentally handicapped. Legit terrible.

If someone could create a product that let cowork and claude code work together in real time... Would be massive.

The models are smarter in the cloud and execute better in the harness. Why do I always get better thinking output via GPT or Claude cloud even though the harness has my second brain with way more info about me?

ChatGPT and Claude are not better than one another for knowledge work. It literally varies prompt to prompt

Am I the only one that has Claude Cowork and Codex open 24/7 and pressure test all their work with the other?

Opus limits SIGNIFICANTLY improved with 4.8

Don't let anyone tell you to use anything but xhigh/the highest level compute you can for every AI task. It is a psyop. Use compute like a homeless person uses WiFi at a Starbucks.

Claude has dropped about 50 IQ points this week it seems

James Blake is criminally underrated

👀 @OpenAI

Telling Claude Code "Idk how to do that, you do it" for every suggestion it gives. Works 75% of the time.

AI music absolutely slaps 😂. Cannot believe how good it has gotten.

Early GPT 5.5 thoughts. They FINALLY fixed the word vomitting that 5.4 did. The answers are much more succicent and on point, very similar to how human Claude feels. My biggest gripe is tool calling still sucks. I ask it to review a call in Granola? Cannot find the call. Find something in slack? No luck. Claude still much better for agentic use, however GPT much stronger reasoning.

Nano Banana is much better than the new GPT image gen v2. Follows instructions much more closely.

Microtransactions will make millionaires via AI agents. Come back to this tweet in 10 years.

Ending prompts on opus 4.7 with "Use your Xhigh setting to think deeply" works every time for me. It thinks way longer than if I just have it use adaptive thinking...

I have never been more bearish on non engineer vibe coders. Shipped enough multi-agent systems this year to spot the pattern. The companies doing it wrong are building isolated agents per workflow. The companies doing it right are building one orchestration layer that sits above their existing tools. Think about it from different verticals.. Finance alone has 100+ workflows. One agent per workflow is a maintenance nightmare. Nobody can trace where data went wrong. Nobody knows which agent owns which output. If you are building agents inside a big organization you MUST have a coordination layer that routes work to specialized agents, passes context between them, and surfaces what matters to humans. Your existing software stays untouched. Your agents live above it.