{evolutionary algorithms, reinforcement learning, distributed systems} \\ previous: HFT & exited startup founder
Joined August 2025
- Tweets47
- Following360
- Followers60
- Likes89
PPO is running within the foragers environment. The hardest population control band is right near extinction (unsurprisingly).
Lets see if i can a policy to keep a low population alive within a band without it going extinct!
Tested out Flint by @Recursive_SI on the RTX Pro 5000.
~240 tokens/s with 4-bit NVFP4 weights, so very snappy.
Anecdotally producing some fairly good Rust code - may become the local daily driver
More experiments with population dynamics in JAX.
Found some interesting baseline configs but now that the environment is fully set up I'll be training a control policy (probably PPO) and seeing how Opus 5.5 / Fable 5.1 perform against it.
github.com/raymogg/foragers-…
What tools are you all using for managing the technical content / papers you consume?
It's seemingly quite difficult to discern what papers are "quality" and should be read, papers that should be skimmed and papers that are not worth the time, especially for new research.
It seems likely to me that eventually everyone interacts with more powerful (custom fine tuned) models.
However as is current discourse, most people simply don't. They just use Googles default search AI or free chatGPT.
Difficult to tell which path we take:
a) Similar to previous technology cycles, cost comes down, access widens, frontier models enter average consumer products.
b) K-shape adoption - power users and institutions continue to use these models more and more heavily. Your average Joe is left behind.
I think its critical we hit path a
Building a "foragers" RL environment in JAX. Could stare at the simulated world all day - watching the population boom and bust.
Microduck training facility about to go live. Time to train my duck army.
Whats the most ridiculous policy I can train onto one of these @pollenrobotics ducks?
I love that you can take an RL concept and almost 1:1 apply it back to humans.
Take this paper "Learning to Reason at the Frontier of Learnability" arxiv.org/pdf/2502.12272
Want to learn a new skill? Get an LLM to produce tasks, questions and tests that it believes you have a high variability of being successful at. Its that simple - just keep applying this process iteratively
Some thoughts on the game theory behind pacing the frontier
- Guaranteeing other labs & nations don't defect is near impossible.
- The incentive for defecting is extremely high, however if you "win" but the civilisation as we know it changes for the worse, whats the point?
- The only rational move is for more resources, effort and focus to go into alignment and safety research. If cooperation can't be guaranteed, our next best bet is target the most critical problems. Ensure that the cost of defection on humanity is low.
We Must Pace the Frontier: I’ve written a new essay on why the AI industry should slow down, with a three-part plan for doing so.
Anthropic is unilaterally committing to the first of these steps. We’ll provide third-party evaluators with permanent, employee-level access to our systems, so that they can verify adherence to our safety measures, report on incidents, and assess models’ alignment during training.
You can read the full post here: darioamodei.com/post/we-must…
The thing I love about the term "side questing" is that it immediately takes the pressure off whatever you are doing and allows you to be a beginner again.
Learning a language -> side questing
Playing a new sport -> side questing
Career pivot -> side questing
inspired by @Recursive_SI results, I've been experimenting with a custom harness for exploring @karpathy autoresearch solutions.
From 250 minutes on an initial seed with a H100 the harness scored 0.985924 bpb - which would put this position 29 on the autoresearch-at-home leaderboard in very little run-time.
Have a few more ideas to test & open source soon.
One downfall of AI & vibe coding is that now my girlfriend knows I can create a decent looking CRUD app in < 2h, she wants a custom web app for everything.
This must be what tradies feel like when they get tasked with "small weekend projects"
Eye opening that a single atomic arb bot on mainnet made ~12.5m, while only paying ~34m to the builder.
I imagine most people have stopped playing the on-chain atomic arb game, those that stuck around ate good here.
correct me if I'm wrong but this seems like the largest ever single-block builder profit in ethereum history, ~$33m to titan
it also may be one of the largest MEV block rewards ever on eth, a 568 ETH proposer payment which falls just behind the SVB USDC depeg (had a 692 ETH payment), 2023 sushiswap whitehat hack (689 ETH), and 2023 curve whitehat hack (584 ETH)
others already commented on the original issues with the order (illiquid route + insane $155k AAVE limit price), but here's where the $50m went:
- $36k to the user's cowswap order (331 AAVE)
- $619k cowswap solver fee
- ~$9.9m to the MEV bot that backran the 17,957 ETH -> 331 AAVE swap (backrun was 128 AAVE -> 17,959 ETH)
- another ~$2.6m to the same MEV bot from backrunning the $50m USDT -> $37m WETH swap over multiple txs
- ~$34.3m fee to titan from the MEV bot (includes $1.2m to lido as the block proposer)
- ~$3.5m in dex swap fees + residual smaller arb txs
insane payday for titan, who sent their profits to coinbase, and this single MEV bot took the majority of the arbs in both the illiquid AAVE/WETH pool and the $13m slippage swap in the main USDT/WETH pool
Side by side - V4 quoting and multi route search with and without poolscout.dev on @base
Not only is it 2.5x faster, but you don't have to worry about RPC rate limits or brute forcing from known hook addresses.
This little sensor has surprisingly been one of the better qol additions I've made in the last 6 months.
Didn't realize how bad the home office air quality got with 2 people working side by side and minimal ventilation.
Highly recommend
