oleg

co-founder @loglogai; prev quant trading, asana, yandex

[oleg @_fjolne]

the new "my code is right, must be a compiler bug" excuse is "my eval shows <open weights model> is better than opus"

at this point it is likelier the problem is in the eval

[oleg @_fjolne]

the bitter lesson i keep getting:

telling yourself you are capable of something is not the same as believing it

subconsciousness is hard to trick

either be in environment where people like you do hard things

or find someone to believe in you

until you get used to it

[oleg @_fjolne]

3 weeks ago I was cto of another rl envs startup.
it raised a lot of money.
at some point the founder decided he doesn't want to do rl envs.
but i still wanted to:

[Ted @notmiramurati]

Meet Yvligo, our open-weights model for running X ads.
(3x better than Grok 4.7)

At https://t.co/pAn5O703dD, we built an X simulation and post-trained DeepSeek V4.1 Flash, outperforming frontier models

Live results:
Yvligo: 383 relevant website visits for $200
Grok: 123 visits

[oleg @_fjolne]

@gortron how do you rerank with Jev: pairwise comps or per-item score? i guess latter given comparison with BM25

[oleg @_fjolne]

@MParakhin https://t.co/OSvYMLImgx

[Diogo Almeida @CompleteSkeptic]

you might be the first person talking about the data over the architecture! 🥲

we consider ourselves a data research lab! the vast vast vast majority of research was on making data that is truly general (ala a cognitive core) and 100% of our data is synthetic (but not the type of crap that is just spit out from an LLM obviously)

[oleg @_fjolne]

@guanlan results can be misleading without multiple tries per combination aka confidence intervals

[oleg @_fjolne]

@composio is it pass@1? models have huge variance, so numbers without confidence intervals can be misleading

[oleg @_fjolne]

@danrobinson you can seed your run with OWM

[oleg @_fjolne]

well, technically it didnt run it

[oleg @_fjolne]

@Hangsiin how app-server is different from `codex exec` for this? the only diff i can think of is if it supports direct tool calls instead of agent calling it, but it seems like you still send a prompt to app-server

[oleg @_fjolne]

@MatternJustus do you control for models stopping prematurely? gpt tends to stop early and simple “continue” makes it go further

as per your stats opus clearly spends way more time on a problem which might explain why it wins

[oleg @_fjolne]

@FactoryAI gpt 5.4 fast is allegedly the same model with request queue priority so confidence intervals must be pretty wide here

[oleg @_fjolne]

@badlogicgames that matches my private synthetic bench: opus 4.7 passes it on any effort, opus 4.6 only on max

[oleg @_fjolne]

@danrobinson @paradigm hey, applied on the weekend but haven’t received a confirmation email yet, is it expected?

[oleg @_fjolne]

@HamelHusain at some point they may come to a solution without sacrifices, however complex and unnecessary it is

tho it seems to be fixable by asking it to leave code comments with tradeoff reasoning

[oleg @_fjolne]

@HamelHusain it often happens when the decision has tradeoffs: 1st pass implements solution which sacrifices A, 2nd pass notices the sacrifice A and suggests to sacrifice B instead, 3rd pass notices B and suggests A etc

[oleg @_fjolne]

@awnihannun cant agree thats the core idea of RLM paper tho: u could impl RLM by storing prompt in a file and letting any existing agent cli slice and dice it and call subagents on it

unless u consider agent cli + filesystem a REPL, which it is

[oleg @_fjolne]

@awnihannun Agentica achieved SOTA on ARC-AGI-2 by giving it a REPL

[oleg @_fjolne]

@Muhtasham9 @claudeai also i was wrong about default subagent model, the default is to inherit from main conversation

[oleg @_fjolne]

@Muhtasham9 @claudeai okay now i see that it uses haiku and sonnet for subagents, and looks like ur screenshot is also from a subagent