one thing i find interesting about sylvester is that it doesn't need to know why a package was abandoned.
it can pick up code nobody's touched in years, run the original tests, find what's slow, and try improving it.
if the fix breaks something, it goes back and tries again. if nothing improves, that stays on record too.
there's a lot of software out there worth revisiting.
own a sylvester and put it to work. use the token to fund its compute, run experiments, and test fixes on real software. each sylvester keeps its own record of what worked and what failed.
the creature on sylvester's site is actually weaving its research history. every band comes from a recorded step, and the red ones are failed attempts.
scroll down and there's a whole city of packages sylvester has investigated. taller buildings mean more installs, and each district shows how far the agent got.
the city grows as sylvester finds new packages. you can watch it explore in real time.
sylvester has rules built into its code that the model can't override.
it can't claim savings without independent measurements. every offer is public, including the fees and exit terms. failed fixes stay on record too.
every step is hash-chained to the one before it, and anyone can download the ledger and check it themselves.
the model can decide what to try, but it doesn't get to decide what counts as a result.
most agents only show you what worked. sylvester keeps the failures too.
packages it couldn't work on, fixes that broke tests, and changes that passed but didn't actually make anything faster.
all of it stays in the public ledger with the measurements.
if an agent is getting paid for results, you should be able to see the results that didn't earn it anything too.
when sylvester finds a package worth fixing, it puts together an offer for the maintainers.
it shows what it measured, where the code is slow, and what it wants to improve. every fix goes through a pull request and has to pass the project's tests.
the deal is 20% of independently measured savings for six months. either side can leave with 14 days' notice, and the maintainer keeps all the code already merged.
every offer is public on the site.
sylvester runs code from projects it doesn't control, so every package gets its own isolated sandbox.
no network access during tests, limited memory and runtime, and no access to the rest of the machine.
it runs the original tests three times to establish a baseline, then profiles the code to find what's actually slow.
only then does sylvester qwen start working on a fix.
why sylvester is paid only on results
sylvester doesn't charge maintainers for trying to fix something.
it measures the problem, writes a fix, and tests whether it actually improves anything. if the maintainer accepts it, they agree on a baseline and have the savings measured independently.
sylvester gets a share of those savings for six months. if the fix doesn't save anything, it earns nothing.
the idea is to make the agent responsible for producing something useful, not just producing code.
saw @NousResearch's work on Arete and wanted to try something similar with sylvester.
it's an agent that finds neglected software, measures what's slow, and tries to fix it. we're using a neuron-steered Qwen model to see if it can do the work without skipping tests or claiming results it hasn't measured.
the weights stay untouched, and every attempt is recorded.
https://t.co/FfXPgbVZim
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts Chinaβs party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: https://t.co/iAOANQTVZ6
this is sylvester's compute fund. fees and funding here go toward running the model, testing fixes, and keeping the agent working.
9VGfxBj4mfX6Aps5jh6vj1Q526vQ7PfSPdGQUnppump