David Cramer

@dexhorthy @0xblacklight @_ChrisCovington

@zeeg @0xblacklight @_ChrisCovington no david ur not supposed to like it tell us you hate it

1) not surprising whatsoever 2) this is exactly what I keep saying about models not being powerful enough today the fact that they can do so much with lossy compression is amazing, but there's no magic here imo (for transformers) context windows need to be 1-2 orders of magnitude larger for the future people keep saying is reality, and even then the compute is probably not worth it

@0xblacklight @dexhorthy @_ChrisCovington I like it

@dexhorthy @zeeg @_ChrisCovington 'zeeg go fix this sentry issue'

@_ChrisCovington @0xblacklight @dexhorthy Careful around anyone who names themselves after one of my open source projects 😅

@0xblacklight @zeeg @dexhorthy we seem to be on the same page alot lol

Too many founders these days think the game is showbiz. It’s grit.

all the delve tweets are gonna look particularly spicy as evidence if any of this is true Pro tip don’t flex unless you actually go to the gym

Hey. Sentry supports @EffectTS_ now. 😎 github.com/getsentry/sent…

@jiaweiou @miguelbetegon @sentry We should def detect TTY here - I think it’s fine for humans but when it’s not an interactive we can collapse it

@miguelbetegon @sentry I know. without instructions/skills, the first "instinct" of a harness is to call `help`, so that logo is in the context window. maybe partly my issue. I don't use claude code and am too lazy to write skills.

oh man. I know when Claude Code does that, it seemed cool, but @sentry please don't do this in your CLI. I like @sentry, but when you are not a harness, you just waste my tokens and pollute my context window for no reason.

Gonna boot up the telescope again tonight. Need to get some calibration frames and fix some of the software. should be live here when it’s up: cra.mr/astro

@kneath kstack

Way back when, I created The Zen of GitHub to direct the taste of a growing software company. I'm still really proud of it. Fast forward twelve years. Now it's the basis for my agent instructions (with some edits). What a wild end-run. warpspire.com/posts/taste

@dhh Bring it back in one piece!

We are back in Sebring for the 74th running of the 12 Hours this Saturday.

The meme survives

Very disappointed to learn that they used a python templating engine to generate SOC2 reports vs Opus 4.6

ok fine ill try it

@thdxr gitlab lol

opencode 1.3.0 will no longer autoload the claude max plugin we did our best to convince anthropic to support developer choice but they sent lawyers it's your right to access services however you wish but it is also their right to block whoever they want we can't maintain an official plugin so it's been removed from github and marked deprecated on npm appreciate our partners at openai, github and gitlab who are going the other direction and supporting developer freedom

@djgrant_ its definitely not _just_ larger context windows, i agree with that and larger might not be the solution. its very possible transformers cannot solve this at all. either way the amount of information it can manage is insufficient

@zeeg I don't think larger context windows solve this. You give an LLM all the information and it can be completely in the dark about what a thing is ontologically. It's like a DVD player that has no idea what the film is about.

At the Agents Anonymous SF meetup last night we did another 🙋 AI usage survey, here are the est. numbers: Usage stats: - 90% Claude Code - 60% Codex - 30% Cursor - 20% OpenCode - 10% Conductor - 10% Own agent/Pi 80% have prompted a coding agent from mobile 50% have not handwritten a single line of code this year 99% think they're more productive now vs. pre agentic coding agents Parallel agent usage: - 90% 3+ - 70% 4+ - 50% 5+ - 5% 10 Also want to give a ginormous thank you to our incredible speaker lineup: - @jonas_nelle & @alexirobbins from @cursor_ai - @southpolesteve from @Cloudflare - @LewisJEllis from @ycombinator - @aidandcunniffe from Git AI - 🦞 @steipete from @openclaw Hope to see you all at the next one! 🫡

@0xblacklight Yeah I agree it’s both recall and quantity and neither problem improved much. Codex skill calling is fairly impressive on recall but it still struggles more as time goes on just like everything else

yeah to be fair if we can get 1M-2M context AND an OOM better-quality instruction adherence at that length then we are really cooking with gas I think right now folks have the mindset of 'longer context will save us' but longer context really doesn't help without better instruction following e.g. I have so far been generally unimpressed with Opus 1M, even at shorter lengths it seems to attend less to instructions that regular opus would handle fine

@deepdrpt They can be totally helpful and also not capable of the impossible ;)

@zeeg the counter argument to that is that the models are still incredibly helpful for in distribution task, ie coding that’s not in brainfuck

@rickc42069 Yeah tho that’s very narrow scoped. Code is unfortunately the opposite of narrow

@zeeg For ex. These Yolo image detection models are so tiny and so specifically trained for their use case. That they can run image detection well even on n150 intel cpus which consume 5w on load. #get-started" target="_blank" rel="nofollow noopener">yolov8.com/#get-started