this is a very pleasant surprise feature
immediate reaction: meh claude had this for a long time
tried it for the first time: interesting. i can see it being helpful from time to time
a few turns in: fahh im a truly meat proxy (at the level that ive never experienced with claude or cursor’s similar suggestion)
its surprisingly good at 1) saving me from context switching while multi tasking, 2) reminding me of things i may miss, 3) solving blank page problem when im in an unfamiliar territory
Show more
we made a simple thing, people internally *really* like it, and we hope you like it too
the little composer that could, from planning the launch with @sharifshameem and @andhudhow:
Show more
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!
Show more
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].
Show more
if you can’t outpace frontier models, earn your way into the weights.
be the best in your vertical, or be the most accessible one. ideally both.
feeling fancy when Get Ready With Dots 💃
got the morning chores that i’d procrastinated for days done while calling dots (in a british accent) to check calendar and track down why i hit claude overage
Show more
research won’t be magically happening. DDIA still matters. i repeat. DDIA still matters a lot.
sharing my codex setup behind this reading list:
it scouts x (optional; separate mcp setup required) + papers around the given interests, takes feedback to pick better going forward, and optionally saves to a queue in notion or space
have fun reading!
Show more
i’ve often worried that pass@k = 0 tasks weren’t worth shipping, but this makes me feel better about making extremely hard tasks -- it seems like failed traces can still yield useful learning signals, even when there’s no good solution to copy
also this reminds me of a similar attempt:
RLTL;DR: failed attempt + verifier feedback -> generate TL;DR lesson -> train the "task to lesson" mapping (no separate teacher)
HINT-SD: full failed attempt -- thru hindsight analysis --> select relevant actions + generate corrective feedback -> train that "trace to action" mapping at selected steps, with hindsight informed teacher predictions (the teacher that comes from the student)
Show more
We let an agent self-improve by exploring and internalizing ”on this sort of task, keep this sort of thing in mind“ bits of self-feedback. It learned to solve tasks where the original policy failed 128 attempts in a row and GRPO flatlined.
Paper:
🧵1/8
Show more
pov: you just hired a manager to resolve conflict between your two beloved ICs. thanks dots for making me a skip
still don’t love the way opus talks to me, but woah man it’s an amazing adversarial reviewer
pass@k = 0 where k is very large 💀
ladies and gentlemen may i present you the FuhuihuaBench
woke up to two dots playing chess with each other in someone else’s log channel??? now i do worry about global warming
not even trolling. i may not need to open claude mobile app anymore
ok i think dots design is quite elegant. it's exactly what i was hoping for, except it's a single bot (still wish for multiple bots). but i love i can talk to my dot to talk to individual tasks and i can still intervene those tasks as needed. @thsottiaux you guys deserve a demo reset
Show more
used grok bot for real work for a few days + been an hourly active on codex for months. really want something in the middle: a few bots (e.g. chief of staff style) that sit above the codex sessions, but i can still dig into any individual task to micro manage as needed
Show more
holy shit calling dots is really fun
reward design may be the new feature engineering
this one so far is supervised learning with weighted reward. tho i wonder if jev’s existence could change its future direction (like move from full LLM backed ranker to a model trained specifically for classification and scoring
Show more
Netflix replaced their 15 years old recommendation algorithm with an LLM.
It’s called "GenRec" and it completely changes how recommendation algorithms are built.
For over a decade, the Netflix recommendation engine was a masterclass in feature engineering. Data scientists built thousands of complex, handcrafted features to figure out what you wanted to watch next.
It required bespoke architectures. Massive infrastructure. Constant manual tuning.
Netflix threw all of it away.
They built GenRec, an LLM-backed ranker.
Instead of translating your behavior into complex math, they just turn your watch history and metadata into a natural language sentence.
They feed that raw text into a foundation LLM.
The AI simply reads your behavior like a story, understands your evolving tastes, and scores the entire catalog in a single forward pass.
No manual feature engineering. No complex bespoke architectures.
Here is the part that should terrify traditional data scientists.
This text-based LLM didn't just match the highly tuned production system Netflix spent years perfecting.
It beat it.
And it achieved those statistically significant gains using roughly 40x fewer labeled training examples.
We are watching a massive paradigm shift in real time.
The most complex predictive algorithms in the world are being replaced by models that just know how to read.
If Netflix can replace their core product engine with an LLM, what complex system in your business is about to become obsolete?
Show more
on simulation and env. i always view the starting world as the end state of a simulated history, vs just a collection of db records or files which you’d have to check for consistency after the fact.
say, we’re training a customer service agent, when we drop it in a training env to solve its first ticket, there should already be history. maybe there was a software outage days ago, which resulted in tickets piled up. some got refunded over time, and some is still waiting.
if that initial state comes from simulation (meaning the history actually ran), things like timestamps, causes, or even complex scenarios should all line up on their own. you no longer have to go back and check if the refund happened after the outage.
now our data/env quality question gets simplified to “whether the world generating them is right”, and the good news is… we don’t have to “train a simulation model”. rather there’s ancient old math designs exactly solve for this new problem!
for example, to simulate state changes like opened to escalated to closed for a ticket properly, we can take “state transitions” to model the ticket with explicit transition probabilities. so it won’t look fake or inconsistent.
another sweet spot is to inject that outage in a history timeline. the old school “discrete event simulation” was made for it - it helps schedule the outage and all of its downstream effects on their own clocks, with realistic uncertainty baked in.
a wrong world gets great consequences - reward hacking on deemed fake env, unsafe when deployed to real world when trained on wrong setup, etc etc, but i believe these explicit stochastic math first sims could already go a long way.
so much more to reread and model. fun times.
Show more
wait is this human distillation. super cool! also loved this wasn't another sft validated data paper.
i once fed the taste skill back to generate tasteful images for prototypes. exciting to see a full effort like this!
Show more
Today we give Superintelligence back to its owners.
Introducing Skill2Env 🎨 , the most aligned and diverse dataset to fuel modern Agentic RL . 🧵
how come chat instant's takes are more holistic (curious?/smart??) than astra in codex 😵💫😵💫 it must be a me problem... i'm feeling extra dumb rn letting codex use CUA to chat with chat instant in web cos one has all the connections and i like the other's analysis better
Show more
still don’t love the way opus talks to me, but woah man it’s an amazing adversarial reviewer
used grok bot for real work for a few days + been an hourly active on codex for months. really want something in the middle: a few bots (e.g. chief of staff style) that sit above the codex sessions, but i can still dig into any individual task to micro manage as needed
Show more
