Mike Vegeto

Almost every topic is at least two orders of magnitude easier to understand than it seems from the outside.

@MiTiBennett Boston

Where to move base? Tax is a consideration, but mostly want a fun tech / research scene.

I wonder if there's anything interesting you could do with a variable-aperture nozzle for a 3d printer

"When the Yogurt Took Over"

@zanehengsperger they're just more into codex

we are so incredibly early

@thdxr it was the best of code, it was the worst of code, It was the age of wisdom, it was the age of foolishness

you can use LLMs to produce the best code of your life you can use LLMs to produce the worst code of your life

@bcherny @DavidKPiano We have drift tests that show clearly that Claude is regressing in quality, particularly in tool calling. It's visible in Claude Code anecdotally, but clearly and verifiably in testing.

@DavidKPiano Hey, Boris from the team here. What can we do better?

Been using Codex much more than Claude Code lately It's wild that Claude became popular *because* it had the best-in-class models for coding... and all it had to do was hold the lead

@realhasanshoaib @DSPyOSS Yes, larger models do better. However I assume it's still more taxing and thus less optimal

@ZECTBynmo @DSPyOSS this might be the case for smaller less capable models, not so much for the frontier models

a fascinating dynamic in LLMs: structure of context is INCREDIBLY important. Having the right content is not enough. If you feed a model the right info in the wrong format, it will reliably not use it correctly. That is, I assume, why @DSPyOSS works so well.

It makes sense when you think about the details, but I definitely wouldn't have expected it to be such a strong, reliable pattern. I think it has big implications to things like harness setup, RAG impl, etc.

This is an extremely reliable and repeatable experiment. For any question where the model fails the MMLU question you can give the correct answer in question: answer format it gets the right answer (just missing world knowledge) The answer in human format almost always fails

@Celmaun @theo this you?

the fall of @theo. from the highest signal to noise ratio on dev youtube to absurd and useless AI ramblings youtu.be/iFdTLLlby9E

@basedjensen "it's web scale"

Only one right answer btw

do it

the best quality in a developer - now and forever - is that they're stubborn. Find the ones who never quit

Hutchins' Law: Whatever an LLM can mess up it will mess up

@photoncmndr can't wait till his agentic CAD product is released

Non-engineer buyers be like:

@ryanvogel yes.

is there really a better AI memory solution other than "if you think something is important keep it in MEMORY.md"

@therealcarlin esp-idf is so great

I LOVE ESP32 > Tons of power > Cheap > Widely available > Bright future But to be CLEAR, if you are using ESP32 with ANYTHING other than ESP-IDF, YOU ARE DEAD TO ME. I HATE ARDUINO!