Ishan Ramrakhiani

cpo @ shepherd (yc s26), @stanford BMC

[Ishan Ramrakhiani @ishanr07]

Life update: graduated high school and moved to SF for YC S26.

[Josh @jjpcodes]

i'm kinda sick of 5.6 sol now honestly. yes it's a good model, no it doesn't help me ship as much software as it should.

it feels too RL-fried and ends up in all these local minima where it just spins its wheels doing nothing, invents bullshit overly cautious procedures you didn't ask for, doesn't solve your problems, hyperfocuses on tiny details, adds theatrical bullshit like tests/ledgers/hashing/proofs/review loops to "ensure stuff works" (none of it works) and degrades context over long-running threads.

1/3rd of the time it can distil your intent kinda okay and keep it up for a few hours but it will eventually collapse into total slop cannon, no matter how much you steer it. the other 2/3rd of the time it will just misunderstand you and slop cannon.

oh yeah and the limits are cooked and it burns 5x the tokens of 5.5 whilst you don't ship.

bring on GPT 6.....

[Ishan Ramrakhiani @ishanr07]

@GaddipatiHarsha @rohandevs real

[Ishan Ramrakhiani @ishanr07]

@garrytan clav told us it's one x now

[Ishan Ramrakhiani @ishanr07]

hardware founders: I’ll DoorDash lunch to your team if you tell us about your documentation problems.

We’re trying to understand where context gets lost between files, docs, and people as the product changes.

20 min, no pitch. Reply or DM me.

[will depue @willdepue]

what's scary is i can tell you firsthand there's still tons of low-hanging fruit everywhere you look. it'd be freaky if this were the endgame, it's even freakier when our methods still feel weirdly nascent

[Ishan Ramrakhiani @ishanr07]

@olam_labs Wow amazing!!

[Ishan Ramrakhiani @ishanr07]

Compare it to 5.6-Sol. Kimi might lose on cost and speed. https://t.co/kwk1GLLTPZ

[Cline @cline]

We tested Kimi K3 and Fable on a real bug from the Cline repo, and found that while both models were able to fix it - Fable wins on speed & Kimi wins on cost.

- Kimi used 1.7x more tokens than Fable (1.2M vs. 730K)

- Fable finished 3.4x faster - 3.5 min and 18 tool calls vs. Kimi’s 12 min and 34 tool calls.

- Kimi cost 2.3x less ($0.92 vs. $2.13) thanks to its 3.3x per-token discount

Both runs used the same Cline harness, and the traces indicate that Kimi is RL trained to spend more tokens thinking and verifying before completing.

This is the first time we've seen an open weight model compete head to head with SOTA. Congratulations to the @Kimi_Moonshot team on this milestone!

[Philip Meng @philipxmeng]

Had a great time coworking at the Corgi offices today! Bullish on @UseCorgi @lauradang0

[Ishan Ramrakhiani @ishanr07]

@freesoloco Small models like this are going to be really important for things like context management. If you really want to make billions of tokens legible to your agents at your company, you need to be able to let larger models orchestrate quicker models.

[Ishan Ramrakhiani @ishanr07]

@seb_thunman @OpenAIDevs @work_louder We all do funny things sometimes lol
https://t.co/VK4Vu5geJp

[Sebastian Thunman @seb_thunman]

Bro thinks we can’t hear him

[Ishan Ramrakhiani @ishanr07]

@seb_thunman @OpenAIDevs @work_louder Nah we bought 3

[Ishan Ramrakhiani @ishanr07]

Codex subscription tps will increase very soon...
https://t.co/bXgXlLP5wz

[Ishan Ramrakhiani @ishanr07]

For anyone using codex, I recommend using the Codex API over subscription, but only for Terra and Luna! For some reason, the tokens/sec rate is just much higher for these models (not Sol!) from the API, and these models are pretty cheap anyway.

[Ishan Ramrakhiani @ishanr07]

No. When you make an API call to pro, the TPS is unusually high, but the visible tokens count is much lower than the total tokens generated. I think Pro mode just takes a best-of-n approach, making multiple calls in parallel. https://t.co/NLSvYhkFSh

[zek @zekramu]

looks like gpt 5.6 on cerebras quietly launched lol.

luna pro max is basically ASI. this is the most important launch in the space in last year and perhaps since gpt 4…

[Ishan Ramrakhiani @ishanr07]

@sensho @blakeandersonw having my agent bench these rn

[Ishan Ramrakhiani @ishanr07]

@sensho @blakeandersonw wait what!? can I use these in codex?

[Ishan Ramrakhiani @ishanr07]

@sensho @blakeandersonw I've been using Devin 1k tps but the model is a bit jester

[Ishan Ramrakhiani @ishanr07]

@sensho @blakeandersonw Makes sense, yeah. At this point I’d pay 10x for higher tps rather than more intelligence.

[Ishan Ramrakhiani @ishanr07]

@sensho @blakeandersonw Yes, makes sense. But I heard that they had ~40% margins, which indicates costs pretty far from numbers blake estimated. Not sure where that's going if pricing is >10x the cost to the model provider.

[Ishan Ramrakhiani @ishanr07]

For anyone using codex, I recommend using the Codex API over subscription, but only for Terra and Luna! For some reason, the tokens/sec rate is just much higher for these models (not Sol!) from the API, and these models are pretty cheap anyway.