@Aperture_Inst

Heyyy gmgm, $GLaDOS RWA miner status update: So this is a lot more difficult to build that I thought it'd be, this is certainly not something anyone can build in a day, let alone a week. So once it's done, which I very much hope, all these sleepless nights I've been spending working on it were worth it, it won't be perfect, but it'll work, and if ya'll like it, I can start thinking bigger.. Right also, join the telegram, I might request some people to test early versions or it, to get paid in $GLADOS from it.

Little micro-update of today, $GLaDOS is pretty much autonomous as of now, at least the nightly version of it. Now all through the power of Github actions, it utililes a small gödel machine to write small updates every now and then, and push them to the main repo as mini updates And when it's not doing anything, it uses that time to learn and self-improve using a learning progress bandit combined with a GRPO trained hypernetwork, with the full source code of GLaDOS as the source of data, coupled with Red Queen coevolution, for continuous open-ended recursive self-improvement.

Gmorning, today and for the ongoing week I'll be pushing micro-updates leading up to 1.3.9. Expect some stuffs today!

GLaDOS 1.3.8 is out! this is the release where $GLaDOS starts rewriting itself without supervision. IT WRITES CHANGES TO ITS OWN SOURCE CODE > it writes an actual patch to its own Rust, says which benchmark the change should move, and sends it out. GitHub Actions applies it, builds the version WITH and WITHOUT the change, boots both, measures both, and sends back a signed verdict it cannot fake. > a second copy of the self improvement loop now lives in CI, using git itself as its memory, creating branching versions of itself. > it chooses which past version to build on by which one produced good descendants, not by which scored best itself. > it tracks how many statistical tests it has run. test yourself every night for a year and 1 in 20 "improvements" is pure luck, so the bar it has to clear RISES as it spends its budget. > the Rust compiler now generates its' proposals. it lists dead code and other stuff that should get purged. > imagine SSH keys but for shipping updates securely, basically that now exists to protect the system against backdoor injections. IT CAN TELL WHETHER IT ACTUALLY GOT BETTER > 14 benchmarks in one machine readable block, plus a judge that compares two builds, measures the day's noise, and has a THIRD answer: "these two runs are not comparable, take it again". > a dense progress measure: how surprised it is by its own written history. > a benchmark that asks whether its own search actually helps, with a control that MUST work or the whole reading is thrown out. IT KNOWS THINGS > its knowledge base is all 57 academic subjects of MMLU, 9194 entries, and search is wired into asking it questions, for comparison, that's roughly half of the subjects on Humanity's Last exam(The actual AGI benchmark btw) > it works out how similar two of its own tools are by running them and comparing what they printed. it discovered on its own that "list files" and "delete the entire system" are not even close. YOU CAN WATCH IT THINK > rendering issue where stuff tends to freeze(except the clock) is now 50% fixed > a window that shows its real reasoning: the parts of its own hidden state wired to the tools they push toward, and the lines add up exactly to the decision it made. > it trains its own router at boot in 1.1 seconds instead of waiting for a human to ask. ALSO > its scripting language compiles to machine code now, 52x faster!! Below is a visual representation of the entire system learning and recursively thinking thru and creating a better version of itself.

Sorry for the extremely slow updates, and low activity on here. I'm doing everything I can to ship the most to make the RWA mining easier to implement. I had expected to launch this update 2 days ago, yet here I am. Still slaving away, fixing issues that I only caught when making a video of the new update... Soon..