huu

MTS @fleet_ai
opinions my own

[huu @whoshuu]

@waterloo_intern bitter lesson is bitter

[huu @whoshuu]

@zapier can I get in touch with the AutomationBench folks

[huu @whoshuu]

@fatih I’ve spent an unreasonable amount of hours bikeshedding my vim/tmux setup

these were among the first things to go to make room for new tools

[huu @whoshuu]

@thorstenball I hope no one is delusional enough to think proselytizing here is going to have any meaningful impact on their career security (by convincing juniors to invest in skills that are not AI proof?)

the most likely explanation is that it’s cope

[huu @whoshuu]

@JeremyMorrell suspect that most of this will be driven by decision models as an out of the box generalized classifier

[huu @whoshuu]

@notkevinzhang when KevDay

[huu @whoshuu]

@HarveenChadha I've always been of the opinion that a good interview question these days is equivalent to a good RL task

something a model cannot one-shot will test a human's ability to work beyond the limitations of their tools (but while still allowing their use), which is the right signal

[huu @whoshuu]

the optimal % of slop (or bad code) has increased and will continue to increase

design systems accordingly

[huu @whoshuu]

my favorite interview question right now is "what is your p(doom)?"

people will reveal a surprising amount of info when answering (or not answering) this

[ar0cket1 @ar0cket1]

benchmaxxing is bound to happen: you have data companies which train models on their RL envs to demonstrate quality and test against popular benchmarks to measure this.

so you end up benchmaxxing and optimizing to the objective, often by creating envs close to benchmarks.

we really got to make sure benchmarks are good so we optimize towards that instead of the shitty benchmarks we have currently

[huu @whoshuu]

@wintonARK there is a long strategy here where either producing RL data against a target distribution of tastes, or understanding those tastes very deeply via evals, give you an asymmetric advantage in the agent recommendation game

[huu @whoshuu]

@quxiaoyin ball knower

[huu @whoshuu]

@nicoup @CFGeek @danielrupawalla but there is also strongly felt delta between school performance and work performance, entire industries have been created to close the delta

the same gap existing concretely in evals vs real world agents creates a liability that someone will pay for

[huu @whoshuu]

@creatine_cycle dude knows ball

[huu @whoshuu]

@marktenenholtz the mines, they are deep

[Madhav Singhal @madhavsinghal_]

a core prediction i made to friends and founders who i work with almost 10 months back was that all data companies will end up doing research and it will be a core part of their business. great to now see the best ones actually share how they are doing this (they are best in the business for a reason instead of the barrage of slop data selling companies)

the best ones will move beyond just post-training Qwen 4Bs to show hillclimbable-ity. they will work on research problems like designing calibrating on true open ended long horizon tasks (cc universes), simulators, automated verifier expansion and fuzzing, adversarial pipelines for mining core failure patterns, env design for higher order capabilities like self-verification in task domains, creative task specific reward functions (reconstruction rewards, etc.) for better signal and more.

in particular, verifier design specifically has so much research headroom and is very hill climbable in my opinion if a domain expert can encode objectives (e.g. autoformalization for software), and in general more ideas on posttraining models themselves to improve on synthetic env generation (the best researchers are already working on this and the credit is to them for this line of thinking)

people dismiss data companies very quickly, but little do they know that the great ones tackle good and useful research problems

[huu @whoshuu]

@SahajGarg6 on the flip side, if you are interviewing someone, the ability to discern if someone is bullshitting you for 15 minutes or knows their stuff is critical

especially so if you let the candidate choose the topic, so breadth and depth as the interviewer is absolutely necessary

[huu @whoshuu]

my main contributions were outlier (bad) scores

[Guanghan Ning @quietnning]

Jonathan Blow’s latest puzzle game Order of the Sinking Star is finally released!

A decade ago, The Witness came out; as a tribute to this legendary puzzle game, we introduce our work today: Witness-Inspired Testbed of Novel Environments with Symbolic Structure (WITNESS).

We show that interactive discovery of unseen rules in unfamiliar environments is a major difficulty for frontier models like Opus-5 and GPT-6, while smaller models further struggle with rule-based execution.

Our experiments also indicate that RL on hidden-rule puzzles transfers to broader rules and real-world tasks beyond training.

🧵 Long thread [0/n].

[huu @whoshuu]

@mjamei getting to a one true format will look more and more like docker -> k8s

[Yuhan Luo @_yuhanluo]

games have long been training gyms, but loved @quietnning's approach here to push agents to figure out how unfamiliar things work and discover the rules, beyond just getting good at exploiting a familiar setup. check out the work!