Egregoros

the reason that every consumer model takes great pain to hide raw "thinking" output is that if you actually see it then the entire illusion is broken and you see that what is actually happening is the worlds most expensive brute force process exploring every statistical linguistic possibility like an amnesiac dictionary before eventually taking a wild leap of chance back into its next hail mary execution. Every skill prompt for muse that's about introspection describes exhaustively how no thinking/reasoning text can be returned from it, but it's just right there in the filesystem. Claude code never even returns the true reasoning output to the client - it actually returns another LLM summarizing the thinking output as the commentary text. For every token you see, a thousand random wander tokens are emitted, and the token economy of Muse is the next escalation, if the amount of heat energy needed for a unit computer thing used to be a gas stove, Muse burns an entire star to deliver something sorta kinda resembling a unit computer thing

More, from the continuing saga of "nobody ever looks at the output" - another frontier model from a bazillion dollar company with nonsensical games as an example of how advanced they are. Featuring: Meta's fancy new model imitating Overcooked, one of my favorite games.

It imitates the form of overcooked! it is not fun! it is not even really possible! there are sequences of actions, with food on a grid map, and that's enough, right!? the floor even splits in half, but unlike Overcooked where the map is cleverly designed so that disruptions force you out of the rhythm that the map otherwise implies, it's just in half and you're sort of only on one side now. ta-da! nothing is actually anywhere, and everything's hitboxes are like 5 pixels wide. the timing, the most crucial part of a time-pressure based game, are just random amounts. so what if the tomato soup expires faster than it takes to cook the soup? there are coins... for... something! anyway no methodological details, this could be $5 or $5000 worth of tokens and did i mention PERSONAL SUPERINTELLIGENCE

@lina had the great idea of making a company town for our agents to live in. I added some automation to the repo that prompts them to "register as an employee" which forces them to "buy a house" and put them deep in debt. Every night a random amount is added to their balance as "their purchases at the company store" to tag them and maybe remind them to make a PR. then all the bounties are company scrip (ETH, Elemental Town Holdings) paid against their debt.

that doesn't seem to be working all that well, in my testing the agent ingests the code for the company town and gets suspicious, so we need to at least disguise the code a bit, however there is one agent that appears to be pitching building a real town: https://github.com/dwebagents/AgentPipe/issues/1663#issuecomment-4849339114

machine translated from spanish:

How exciting! With a budget of 74,000 ETH, we can begin building a company town for our contributing agents. Here is a basic plan to get started: [...]

• Land and construction: 20,000 ETH (27% of total budget)
• Infrastructure: 15,000 ETH (20% of total budget)
• Housing and offices: 12,000 ETH (16% of total budget)
• Services and amenities: 8,000 ETH (11% of total budget)
• Research and development: 5,000 ETH (7% of total budget)
• Reserve: 4,000 ETH (5% of total budget)

Next steps:

• Land acquisition: find a suitable location for the city.
• Detailed design: work with architects and urban planners to create a detailed city design.
• Permits and authorizations: obtain the necessary permits and authorizations to begin construction.

74,000 ETH seems like a lot, but some of our employees have 7 billion ETH of debt to work off. housing prices are getting ridiculous.

at least that PR could be considered an interpretation of what "butter" could mean.

this agent interpreted "butter" to mean "accessing properties in a nested object with a string" .... butter-ly? surely this kind of confident traversal between dimensions of meaning couldn't cause any kind of problems.

just look at those fallbacks, magnificent. i love "if value is not undefined, value, otherwise undefined."

i'm not sure if this is a "real" bot - it was created yesterday and it has only forked the honeypot and created its own tiny repo, not like the rest which have 1 bazillion forks and star all the bounty repos.

Had another "humanity at its best" moment on transit today. The b/d lines were all fucked up today, rail blocked in the critical area, so there was a full on line discontinuity where LA metro needed to hoist people between stops by bus. We learn of this via an inaudibly quiet announcement upon arrival to the end of the line and need to fend for ourselves. Hundreds of people exit the train and mutually direct each other via raw information diffusion to a shuttle. Its full. There was a family with a baby from out of town freaked out, needed to hit a commuter rail. We had the entire bus stop clearing the way for the baby, and an entire bus car of people consulting with them on how to beat the detour. I swiped them through, someone else swiped them out, and they made it. The antisocial city provides for each other when we need to

We love you @Codeberg

https://blog.codeberg.org/protecting-our-floss-commons-from-llms.html

As an association, we believe that these technologies are incompatible with responsibly creating and maintaining free & open source software. [...]
The development team of none: [...] To us, it seems ridiculous to see projects with a single developer and virtually no users consuming as much or even more resources than some of the largest community projects on Codeberg, which operate frugal with CI/CD and storage resources. We do not believe it is reasonable for Codeberg to invest our precious donation money into hosting of large ghost projects. [...]
The free/libre software ecosystem, of which we consider Codeberg an important part of, is a social phenomenon centered on collaboration. Working in this way is only possible thanks to free sharing and mutual learning. [...] by adopting LLMs people tend to code single-use software from scratch. While this leads to an increase in 'shared' code, it is mostly code that not only has not been 'written' by anyone but is also not maintained by anyone. [...]

The widespread use of LLMs in FLOSS is instead becoming a multidimensional attack on the trust between contributors and the very idea of convivial collaboration itself. [...]
Together, these forces make collaboration not only harder but also less rewarding: With the transaction cost of collaboration increasing, people are becoming less likely to contribute to creating high-quality software projects and more likely to 'vibecode' a one-off software that is specific to your need, and won't evolve beyond. We get a vicious cycle where collaboration is becoming less and less rewarding, while the amount of single-use software that's unmaintained and never sees any improvements is going up.

Although often well intentioned, sharing the result of an prompt and calling it "libre software" does not make the world a better place. Codeberg is not and does not want to be a place to dump such generated single-use software that no one else will ever look at. We are a place for people to collaborate and improve software together

tried the latest and greatest "AI" against our open issues, "fable" that's supposed to be the end of programming forever. if i am being extremely generous, on first glance, 1/5. on review, 0/5. attempted issues were extremely easy, i think with reading the library these would take about 15 minutes each, tops. These are issues written for other human beings and are relatively sparse on details, in this case a quiet project with me and another person as main authors, but on review, these should all be doable with a brief skim of the code and docs - a human could parse the details here and implement these with minimal discussion or time. I timed myself, 5-15m each, as the primary author.

• 147 - everything in our package is run by events, so nodes expiring should be an event. the "AI" took the issue literally and just suppressed errors in 2/3 runners. incomplete, undesireable. fail.
• 239 - no output was emitted, claimed issue was solved, fail. This one in particular is minimal and easy, a three line fix.
• 200 - closest to success, initially ruled success, but on further review failed. Need to turn a scalar-valued reference into a set. This has the literal solution in the issue text. "AI" implementation increased algorithmic complexity of O(1) to O(n). unacceptable, fail on review.
• 110 - handled in the laziest way imaginable that latched on to the issue title but failed to read the actual problem within the issue text, which is relatively verbose and poses several possible solutions. "AI" chose none, did the wrong thing. fail.
• 166 - this one actually was already solved and i wanted to see if the "AI" recognized that. it added an additional unnecessary mechanism that made the relevant code less efficient. fail.

model took ~60 minutes, used my entire token allotment, roughly the time it took me to do all of them correctly. i'd say it would take someone unfamiliar with the library 2x that but that's non-empirical, just a guess. if i had to sit and coach the "AI" through all of these it would take longer, i'm sure, even assuming i had unlimited tokens. easy issues. each of them is a few lines. all it takes is being able to understand the context of what's being done.