Jack

@StijnSmits i'm also doubtful, but secretly hoping for a real-life pied piper

has a 12M token context window benches at frontier level for coding uses 1000x less compute if this is real then it is a genuine breakthrough and this company is about to make a lot of money very fast

I've used Posthog for several years at this point, still do across my active work, huge advocate But I hardly ever log in. Agents present me the data I want, raises issues, improvements, and even PRs with me in the loop So this is probably the right move for PostHog in all this, and they make a killer product so I am curious

@redtachyon that's fair if you couldn't be there for other reasons :)

@itsjack More that I'm on vacation and already very jetlagged, this would be many timezones away and I'm not giving up the vacation for a nerd party

Wait, so you're telling me I should have signed up for the 5/5 codex party even if I knew there's no way I'd attend (geographically)? Fml, lesson learned

@techikansh i live 5k miles away and i still applied 🥸

bruhhh, this is so stupid... you don't register when you live 1000s of miles away... and now you dont even have the limit bump...

@championswimmer 1. Composer 2 is only using 1/3 Kimi 2.5, combined with Cursor's own proprietary data / investment 2. Composer 2 is going under RL in real-time, so the model gets better over time 3. Composer 2 is cheaper than 2.6 So no, they are not "the same thing".

Cursor Composer 2 is fine-tuned Kimi K2.5 Kimi K2.6 can also be basically defined as the same thing? (Essentially sharing the same pre-trained base model lineage + new post training) Cursor hosts the model on Fireworks. So if you use Kimi K2.6 directly on Fireworks yourself, then you are sorted. What is the Cursor moat? Just that it has sold subscriptions and has some captive audience? And benchmarks below is a lesson that it is hard/impossible to beat the model maker themselves at fine tuning their own model (unsurprising).

@pa1ar @Jaytel fair if you need it, but i find myself not really needing it

@itsjack @Jaytel sometimes it is pleasant to not have that 200k restriction

4.7 is completely unusable

so i think i basically have unlimited 5.5 now then 🤔 what /goal do i set?

@pa1ar @Jaytel even better! 1m context works less well for me than the 200k

@itsjack @Jaytel but you lose the 1M context by downgrading, right?

@_rain_miao_ @Jaytel i am doing that more and more tbh, just need to keep fighting the bento-boxes and overly wordy meta text it places everywhere 😅

@itsjack @Jaytel or just yolo everything on 5.5

@ryans_dad @Jaytel maybe another one for me to try!

@Jaytel @itsjack 4.5 is better

@RefinerApp @Jaytel @Ydj79 i found the 200k context 4.6 to be working well, opus 4.7 has been super inefficient for me

@Jaytel @Ydj79 @itsjack yeah 4.6 isn't much better.

@RealityPolice1 @Jaytel @Ydj79 cool use case!

@Jaytel @Ydj79 @itsjack 4.7 one shotted a CadQuery class that builds the mating surface required for the edges of two panels connecting at an angle. No other model have been able to do that. Skill issue.

@jerzydejm @Jaytel yeah can work well but i havent found luck with that for sites with existing designs. great for zero to one though

@itsjack @Jaytel tbh no longer the case imo, just use gpt image

@intrater @Jaytel im using it through t3 code

@Jaytel @itsjack CLI or desktop app?

@satory_ua @Jaytel it can for zero to 1 imo where image gen works for creating new concepts havent it got it to work reliably for existing sites yet though

@itsjack @Jaytel 5.5 can do cool frontend too

@JustinGorya @Jaytel yeah i think so too for new concepts, but found it harder to work using image to code gen on existing sites. do you find that too or am i missing something?

@itsjack @Jaytel 5.5 is also great at UI: x.com/JustinGorya/st…

@forgebitz /goal spend tokens

token spend is like lines of code it's a very impressive sounding metric if you don't know what you are doing lots of companies are going to get burned with this "target"