Igor Kotenkov (@stalkermustang) / X

Igor Kotenkov

3,250 posts

Igor Kotenkov

@stalkermustang

Around the world, A-A-AROUND

Joined June 2012

Pinned
Igor Kotenkov
@stalkermustang
May 11
While reading the DeepSeek v4 paper, I ended up writing down over 90 questions. A lot of the paper reviews out there skip over the details, which is usually where the actual learning happens. So, I decided to put together a proper guide: an Annotated Paper Walkthrough. The core
00:00
40K
Igor Kotenkov
@stalkermustang
Apr 4, 2025
Replying to @sama
sama what about o3 pro? we don't need mini models, we want them to be KNOWLEDGEABLE
234K
Igor Kotenkov
@stalkermustang
Oct 29, 2025
Replying to @nizzyabi
it will slap you when u sleep
85K
Igor Kotenkov
@stalkermustang
Jan 17, 2025
btw o3-mini access is out for (at least some) testers thought there are two models, and it seems OpenAI wasn't losing any time during NY holidays
130K
Igor Kotenkov
@stalkermustang
Feb 13, 2025
babe wake up AIME 2 LLM results dropped o3-mini is the king, gemini is cooked, R1 is _reasonable_
65K
Igor Kotenkov
@stalkermustang
Jan 28, 2025
2.0-pro is available in AiStudio, even responds to my request announce/soft launch through @OfficialLoganK 's twitter any minute now? waiting for metrics
65K
Igor Kotenkov
@stalkermustang
Nov 5, 2025
this is what's called "Noam's touch" btw
John Yang
@jyangballin
Nov 5, 2025
Replying to @jyangballin
We ran 1,680 tournaments (25,200 rounds) to evaluate 8 frontier models. Claude Sonnet 4.5 tops the leaderboard, but no model wins across all arenas! GPT-5 dominates Poker. o3 crushes Halite. Claude owns Core War. Every arena reveals different strengths.
73K
Igor Kotenkov
@stalkermustang
Nov 2, 2025
It still amazes me how many ppl aren't aware of this paper (that proves the point in tweet is wrong)
Cameron R. Wolfe, Ph.D.
@cwolferesearch
Nov 1, 2025
The value of RL is very clearly / nicely articulated by DeepSeekMath… - RL enhances maj@k (majority vote), but not pass@k. - RL boosts the probability of correct completions that are already in top-k. - RL does NOT clearly enhance model capabilities.
44K
Igor Kotenkov
@stalkermustang
Nov 8, 2025
Prediction: in the next 2-3 months we'll see a suite of new benchmarks which will show that, in fact, Kimi's K2 is lagging behind frontier models. (and some of these benchmarks will be multimodal... guess why K2/DS won't be there on the LB) When people will start looking at the
Deedy
@deedydas
Nov 7, 2025
🚨 Today is a turning point in AI. A Chinese open source model is #1. Kimi K2 Thinking scored 51% in Humanity's Last Exam, higher than GPT-5 and every other model. $0.6/M in, $2.5/M output. The best at writing, and does 15tps on two Mac M3 Ultras! Seminal moment in AI. Try it
101K
Igor Kotenkov
@stalkermustang
Jul 22, 2025
Rumor has it, the folks from partner companies outside OpenAI are already testing o3 alpha / GPT-5 (unclear whether these are the same or not) @apples_jimmy what do you say
16K
Igor Kotenkov
@stalkermustang
Sep 26, 2025
This is cool & all but what are these tasks? I've decided to look deeper into it and made some plots that OpenAI folks didn't include in the paper :( I was trying to figure out why, on GDPval, Opus beats GPT-5. My main hypothesis - which I still buy - is that the model is
Tejal Patwardhan
@tejalpatwardhan
Sep 25, 2025
Understanding the capabilities of AI models is important to me. To forecast how AI models might affect labor, we need methods to measure their real-world work abilities. That’s why we created GDPval.
8.5K
Igor Kotenkov
@stalkermustang
Feb 18, 2025
This man is not impressed, thus no pressure on announcing/releasing GPT-4.5. He's just chillin' and vague postin'
29K
Igor Kotenkov
@stalkermustang
Jan 23, 2025
Humanity's Last Exam by @hendrycks is out lastexam.ai (paper is there) results below, the first image is the whole exam, the second is text only (DeepSeek-R1 can't take images as input) R1 seems good. Waiting for o3 results though.
51K
Igor Kotenkov
@stalkermustang
Jan 24, 2025
Replying to @SmokeAwayyy
source? Doubt this number
25K