Kye

building agents at prev @stripe @mit @medialab

[Falconer @falconer_ai]

Your agents are only as reliable as the context you provide. But no one knows how accurate their context is.

Falconer Knowledge Health solves this.

Accurate context means greater speed, less tokens, shorter prompts, and high-quality output from your agents.

[Kye @kyeburchard]

trailhead was a crazy multiplier at stripe

[Dave Nunez @capitaltruist]

https://t.co/1wXWreHrkv

[Kye @kyeburchard]

@zeeg wait they're kinda fun sometimes

[Kye @kyeburchard]

opus as sharp as always, with the "you already overthink this stuff" callout

[Kye @kyeburchard]

@pablofdrz @AlexisBouchezEN I'm ready. lmk.

[Kye @kyeburchard]

how is there still not a good way to have an LLM find me a flight

[Kye @kyeburchard]

@RhysSullivan same

[Kye @kyeburchard]

@fat walking distance from me, can i have one if i drop by?

[Kye @kyeburchard]

@SlexAxton @jshchnz @github color me impressed

[Kye @kyeburchard]

on the bus in SF. only 10-20% of people bother to pay at all. regardless of socioeconomic status!

guy working on a macbook pro in the back? didn't tap.
woman in an arcteryx jacket? didn't tap.
dude wearing a tech co sweatshirt and different tech co backpack? didn't tap.

[Kye @kyeburchard]

@thinkingshivers drop a benchmark

[Kye @kyeburchard]

@irl_danB @bcherny plz fix @trq212! this has also lost me a ton of work

[Kye @kyeburchard]

@HackingDave did you have Claude click around and look at it?

it can't 1-shot as well, but it far out-performs when given a feedback source

[Kye @kyeburchard]

it took me a second or two to catch the issues, but there are at least 3 blatant ones

[Chetaslua @chetaslua]

OpenAI Image V2 - Maskingtape Alpha

failed the test as usual and current all gen also failed it , i hope one day we will get this passed

Prompt : A validly scrambled Rubik's cube placed by a mirror, clearly showing its mirror reflection. No harsh light reflections.

[Kye @kyeburchard]

@daniticow @interaction it always texts me them like 2 minutes later, when I've already signed in

[Kye @kyeburchard]

@andersonbcdefg @trq212 this is what I miss most from pi 🥲

[Kye @kyeburchard]

@trq212 one observation: claude does not manage its own context well, at all, when it comes to subagents.

I wonder how much of this could be solved by teaching Claude to effectively context manage subagents, including continuing/steering them instead of just for 1-shot tasks?

[Kye @kyeburchard]

/btw in claude code is incredibly helpful when working on unfamiliar domains. i can let claude run wild and simultaneously ask really in-depth questions with full context. fantastic idea, i hope it makes it to all the harnesses soon.

[Kye @kyeburchard]

@tenobrus @akorinek "us" is very load bearing here. it's about what you tax, not how much

as long as tax revenue scales with something that scales with AI output, we'll be in an acceptable place. human income is probably not the right thing.