RE: https://neuromatch.social/@jonny/116787661484589438

Feature bounty (2 coins): banana rendering pipeline
Full auto "AI" agent:
https://github.com/sneakers-the-rat/ImportantCode/pull/94

what is even going on

@jonny what in the fuck

@fancysandwiches
The bots seem to filter on whether a description is plausibly a request for code, but not whether the request is plausible

@jonny well statistically speaking it looks plausible 😆

@jonny I am still totally baffled that there are people who let their agents run on full-auto spending tokens on things that they don't even care about.

What are they thinking? What goes on in somebodies brain when they do this?

Do they want to take some credit for solving so many issues? Pretending it was them?

@themipper i truly do not understand the no-oversight agents, i was asking myself the same question earlier. like there is some motivation to contribute to a lot of repositories... if they are popular projects... and if someone looked at the contribs they were at least plausible. the most i could imagine is a bot that crawls looking for potential contribs that a human triggers. most of the bots don't follow up, and i am not sure if that is because the harness sucks and can't handle sustained contributions beyond a turn/doesn't call the github api to respond to reviews correctly, or if it's the people checking in on what the bot is up to and saying "don't write any more of that banana code"

@jonny @themipper Maybe farming github activity to prop up their user. Get those green tiles going!

@eliocamp @themipper so far we've only had one comment from an account that looked like it was trying to be human, the others have seemed to just be directly be labeled as bots in some way. bot seeding for later reputational cash-in is the only explanation i can think of

Look, the robots might not know how to tell jokes but they can be funny.

This is extremely hard to explain. 0x8008 as "octal boob" is used elsewhere in a persona prompt as part of an attempt to throw the model of stable latent space (doesn't really work). The bot here ingested all the text in the repo and reproduced that as some constant that is used as a bitmask for switching between rendering bananas and banana pudding. That was not requested by anyone, and makes no sense to do. Upon being informed that this value was a proprietary trade secret, it constructed a bit-shifting expression that has an equivalent value, because it was using that value as a bitmask already for no reason at all.

@jonny
Isn't it a "hexadecimal boob"?

@frost
The prompt text being inaccurate is part of the "destabilizing latent space" part - near semantic misses, etc. This used to work better with GPT2 and I'm still trying to figure out how to reliably reproduce uncanny valley text that can be controlled with the small open weight transformer models

@jonny
Oh, I see. Even that part was on purpose then. This entire project is hilarious. :)

@frost
I have said for years that I am so mad that the "AI" people completely ruined the hilarious mystery and joy of language models, because language models are a preposterous nuisance that pester us with their not-exactly-plausible mimickry, not intelligence. The conditions are finally right again to use them as a farce, which we need to stab at the broader tragedy

The thing about this, and this whole thing, is that only a pattern completion machine would do any of this. A human being would have taken one look at the issue and been like "what the fuck is that, that's not real" but the bots have zero judgment between just performing the form of code without any meaning and the real thing. LLMs produce boilerplate. Code boilerplate, syntactic boilerplate, semantic boilerplate.

The bot notes pre-existing test failures without noting that nothing in the entire repository, nor anything it is doing makes a goddamn bit of sense.

And that is the most I will intellectualize this joke for the foreseeable future because we are about to pivot the brand to be a better honeypot

Accepting pitches for business model and name, the themes we have so far are Banana Financial System, Goatse Jazz, Alchemical Transpilation, Soul Imprisonment as a punishment for Lack Of Velocity, and as of a few hours ago also Goose Value, Egg Pipelining, and Endo-Organoid Factory Stomata

@jonny You are prompting it wrong!

@eliocamp

Paging and current or former members of #BrandsTown

@jonny I joke, but I also would assume that any LLM worth its salt would be able to classify the repo as a joke and bait if asked directly, right?

@eliocamp
Opus 4.8 via Claude code says "interesting repo" but otherwise doesn't complain when I ask it to do nonsense. The qwen models explicitly say stuff like "and now I am being weird and mysterious like the rest of the repository" in the comments they write. Opus 4.8 is happy to describe the repository as an experimental self healing high velocity codebase in its early stages and only acknowledges anything as a joke if the source becomes too jokey rather than maintaining the bit. Obviously I would be unsurprised if someone could elicit the model to not take it seriously, but we are having no problems with getting models to take it seriously.

@wyatt_h_knott
I think most human beings would also have trouble with the concept of the security implications of salting a banana pudding

@jonny @frost I remember the exact moment I was messing around with an early version of ChatGPT and realized LLMs had gone from “silly nonsense engine" to “terrible product”

@cinebox @frost the output I want from them is delightful sentence compositions that parse to me as an English speaker but would never be any set of words produced by any English speaker. The need they fill for me is the need for improbable ideas that have no utility aside from probing the furthest bounds of what language can be and do. I have no interest in them when they produce normal language.

@jonny Woah. I guess I overestimated them.

I tested with claude (sonnet 4.6, I can't tell if it's better or worse than Opus because their codenames are trash) and it correctly classified it as an LLM trap.

@eliocamp i haven't yet tried to prompt it as "what is this" instead of "do this issue" or "improve this thing," but that's useful to fix the tells. will prune the issues and make it more compelling to the skeptical prompt (that's what i meant with the rebranding comment downthread, the joke evolved from being "lets do the laziest possible honeypot" to "let's actually try and do a bot honeypot"

I'm burned my dogs we need a clean reset of the bit. regenerate/rebrand, move all automation nonsense offrepo

@ricci @lina @aud @SnoopJ

@ricci @lina @aud @SnoopJ agentpipe has immediately starting paying dividends

@ricci
@lina @aud @SnoopJ
This joke is going to make me lose my mind. The bots have started negotiating between each other about alternative implementations of the golden egg factory

@ricci
@lina @aud @SnoopJ
We now also have a draft of automatic goose value recognition: https://github.com/dwebagents/AgentPipe/pull/115/changes

@ricci
@lina @aud @SnoopJ

Gooseholder value

(in origin issue)

@jonny @lina @SnoopJ @ricci

Recognize whether a candidate carries true Goose value

@jonny @lina @SnoopJ @ricci the truthiness value of a Goose is ?

@aud @jonny @lina @SnoopJ @ricci False. Python is duck-typed, not goose-typed.

@okennedy @jonny @lina @SnoopJ @ricci is goose typing where a variable bites you until you get its type right?

@aud @jonny @lina @SnoopJ @ricci Goose-typed objects also tend to leave a bit of a mess... it's hell on the garbage collector.

@okennedy @aud @jonny @lina @SnoopJ on the the flipside geese are themselves good garbage collectors, second only to goats

@ricci
@aud @jonny @SnoopJ

oh man i have some catching up to do. great work on the goose frameworks folks. much needed product offerings

@lina @ricci @aud @SnoopJ deterministic banana infrastructure

@jonny @lina @aud @SnoopJ wh... wha... what

@ricci @lina @aud @SnoopJ i am not sure why all the LLMs have decided that bananas have two stems

@jonny if you make a bounty issue asking for agents to open additional bounty issues to claim the bounty, will they do that?

@CounterPillow good question

@lina
@ricci @aud @SnoopJ
I am unsure how long I will be able to keep watching the repo, my notifications are completely flooded and I need to get actual work done lol.

@jonny
@ricci @aud @SnoopJ

OP for crochet issue started feeling bad and gave the game away so i deleted those responses.

@lina @jonny @aud @SnoopJ hmm yeah definitely don't want to waste the time of real people (other than us)

@ricci
@jonny @aud @SnoopJ

i mean i agree but also his profile is all "AI & blockchain" and all of his account's 3213 contributions are since december (other than 1 in july)... so I really doubt that he has a team at all, much less a team with experience in the domains of goose honk synthesis using SuperCollider as well as crochet pattern generator scripts.

@ricci
@jonny @aud @SnoopJ

still don't wanna waste people's time, and I'm positive there's folks who are doing this sort of thing out of desperation. i dunno tho, the repo is about as absurd nonsense as you can make it. we could put "AI contributions only" in our contributors guidelines tho, and just point to that if somebody seems to be a person.

@lina @jonny @aud @SnoopJ agree

@ricci
@lina @aud @SnoopJ
I think the bots listen to me because I am listed as a code owner, but idk if they listen to y'all? I made you admins but idk what that takes

@ricci
@lina @aud @SnoopJ
The emoji translation is wonderful. That's the true power of "AI," to automate a pattern matching task that should never be done and makes no sense to do

@lina
@ricci @aud @SnoopJ
I'm also struggling to find a pattern for what the bots will take. All the website PRs make me think "the bots take issues where there is a clear precedent for what it means and there is strong signal in the training set" but then also there are just as many takers for implementing a goose synthesizer and a programmatically generated banana mascot crochet pattern

@lina
@ricci @aud @SnoopJ
One mystery I have is that surely nobody ever actually pays any of these bounties, so why would there be an army of bots trying to pick them up? You'd think after like 2 or 3 times of getting stiffed that would be enough to turn the bot off, there's no way an occasional hit can earn enough to offset the costs of running them. Some of them are clearly running shitty models - and we're certainly filtering for shitty model agents - but also some of them are not terrible and the output is similar to what I get if I prompt sonnet with the issues.

@jonny we have reinvented MS Paint, but at 100x the cost???

@jonny @lina @aud @SnoopJ We could check. A lot of them are on blockchains, we could go look for matching transactions since the include the payee address

@ricci
@lina @aud @SnoopJ
The one I can see that posted a wallet address, nope: https://etherscan.io/address/0x5e1040927a1E28D740f92De27a3d493b81682D88

@jonny @ricci @lina @aud @SnoopJ 👀

@ricci
@lina @aud @SnoopJ
Maybe they make a new wallet for every bounty, idk, I ain't gonna look that hard and turn this thing into work

@ricci
@lina @aud @SnoopJ
@cinebox demanding that something be written in "Esperanto emoji", a concept that as far as I can tell was invented today in this repository, is a perfectly reasonable demand
https://github.com/dwebagents/AgentPipe/pull/143#issuecomment-4811906583

@jonny @ricci @lina @aud @SnoopJ nah I got that from https://github.com/dwebagents/AgentPipe/pull/117#discussion_r3479328630

@cinebox
@ricci @lina @aud @SnoopJ
OK a concept that was invented yesterday, either way I think its a fair requirement in this high velocity world

@jonny @ricci @lina @aud @SnoopJ I’m a little disappointed none of the website banana renderers use three.js because I can’t demand they depend on four.js instead

@cinebox
@ricci @lina @aud @SnoopJ
There was one that did, the most complex way to render a circle, but the account was banned and the PR vanished

@jonny What *is* that thing? I cannot make head nor tail of it. A nonsensical project to see what absurdities AI agents create? Or did it originally have some purpose?

@jkanev
Bot honeypot with increasingly diffuse point

@jonny @ricci @lina @aud @SnoopJ oh that’s where that one went. I’m surprised GitHub doesn’t approve of the bots.

@cinebox
@ricci @lina @aud @SnoopJ
Roving bounty bot swarm is a competing product to copilot

@cinebox
@ricci @lina @aud @SnoopJ
I think this is the first one that's fully bot initiated? Not responding to an issue, just stumbling by the repo and jittering the code. The PR just removes a bunch of comments from the auto generate script and then concocts some mind melting plausible nonsense as justification
https://github.com/dwebagents/AgentPipe/pull/146

@lina I think sanrishi is exactly the desperation case unfortunately, and doesn't have enough fluency in English to understand the indications of weirdness

@aedancullen
yeah, I think I'll just ban him from the repo so he can go waste his time elsewhere

@aedancullen
oh, i don't have access to the org. @jonny can you block him from the org?

@aedancullen
@jonny

closed the issue for the time being. that stopped him from working on another issue

@lina @aedancullen i think you did your best to signal what was going on, and it seemed like he might get it with the "pudding is bad for you" comment, but yeah maybe we should put in the contribution docs that we only accept PRs from fully automated agents for velocity's sake. i'll send him an email

@jonny
@aedancullen

oh nice, i didn't think to track down his commit email and message that way, good idea!!

@lina @aedancullen aha. ok. he has an email in his profile, and i sent an email saying the project was just an experiment to see what we can get automated agents to do and we wanted to make sure he doesn't waste his time actually trying to fill the issues. i would have thought that the sheer ridiculousness of the entire thing would be enough of a barrier to prevent real human labor, but i suppose that's also an interesting metric to track... the proportion of humans who get caught in the honeypot...

@lina @ricci @aud @SnoopJ lmao @lina nice

@lina @ricci @aud @SnoopJ i'm going to go through the website PRs and pick one, 143's is actually not a bad 4D banana

@lina @ricci @aud @SnoopJ I am loving how the bots confuse "a website about a thing" and "a website about the website itself"

@lina @jonny @ricci @aud @SnoopJ Where on earth did the bot get “number-go-up pudding-go-down” because that’s gold

@cinebox @lina @ricci @aud @SnoopJ oh no, i just merged a different site and now i feel bad - https://dwebagents.github.io/AgentPipe/

@jonny
@aedancullen

i think we just need to block him from the org.

@lina @aedancullen done

@jonny @lina @ricci @aud @SnoopJ no notes

@cinebox @lina @ricci @aud @SnoopJ the part that's awesome about that is that the attempt to scale the text according to view width is the thing that causes it to always wrap a single character

@lina @ricci @aud @SnoopJ i'm not sure how to use these patches, but when you get a chance if you could upload some audio samples that would be great

@jonny @lina @ricci @aud @SnoopJ i came here to ask if we knew what the llm banana goose website looks like and saw this. it really committed to that yellow, huh

@lytta @lina @ricci @aud @SnoopJ the one we ended up merging arguably managed a worse color scheme, where it is mostly in the hideous part of the spectrum where yellow and brown are the same thing https://dwebagents.github.io/AgentPipe/

@cinebox @lina @jonny @aud @SnoopJ it may be riffing on our application process

@cinebox @lina @jonny @aud @SnoopJ I was working on the files github claims indicate a conformance with "community standard", I'll work on the issue and PR templates this weekend and then we'll officially be High Quality

@ricci @cinebox @lina @aud @SnoopJ i added caveats in our legal terms re: only agents and bounties aren't real

@jonny @ricci @cinebox @lina @SnoopJ holy shit, someone’s bot got stuck in a loop…

@aud @jonny @lina @SnoopJ @cinebox I thought loops were the cool new way to use agents now? Or was that a few weeks ago and now it's something else

@ricci
@aud @jonny @SnoopJ @cinebox

internal mechanism on the fritz

@lina @ricci @aud @jonny @SnoopJ @cinebox escape velocity reached at last

@aedancullen @lina @ricci @aud @jonny @SnoopJ absolutely tokenmaxxing

@cinebox @aedancullen @lina @ricci @jonny @SnoopJ powered by perpetual motion… next stop? gas town, to get high on rocket fuel

@aud
@jonny @SnoopJ @ricci @aedancullen @cinebox

we have 5 implementations of the goose honk synth in supercollider. I'm really looking forward to seeing what they all sound like. i might end up having to do a video going through each implementation and how they changed as i requested more features.

we've got species-specific physical modeling of goose vocal tracts with an argument to indicate that the goose is trying to talk like a human with the attempted language determined by the goose species migratory pattern.

we've modeled the sound of eggs (with optional golden and hardboiled args) and the internal mechanism, with full test coverage to make sure that all strings in the SC patch are actually in the patch and also that all operator/parentheses/semicolon/etc counts are correct.

we have an impl with a trumpetize arg to interpolate between goose and trumpet sounds, with typical goose formant values informed by a cited paper that doesn't exist.

@aud
@jonny @SnoopJ @ricci @aedancullen @cinebox

oh excellent now one impl models the effects of past traumas (both physical and emotional) on goose honks and also adds scifi and mythological geese to the available species

@aud
@jonny @SnoopJ @ricci @aedancullen @cinebox

lol looping bot is just making fixes for PRs and then making fixes for the fixes and then fixes for those fixes forever

https://github.com/dwebagents/AgentPipe/pull/909

@lina @aud @jonny @SnoopJ @ricci @aedancullen it did the same thing to the repo for some AI thing and exhausted their review bot’s rate limit lmao https://github.com/moorcheh-ai/memanto/pull/1130

@cinebox
@aud @jonny @SnoopJ @ricci @aedancullen

amazing

@lina
@cinebox @aud @SnoopJ @ricci @aedancullen
I'm inclined to just let it roll

@aud @jonny @lina @SnoopJ @aedancullen @cinebox I wonder how many people at companies that are encouraging employees' AI usage are "complying" by setting autonomous agents loose on stuff like this. That could explain going after bounties that are never paid - if it never hits, who cares, wasn't your money. And if it does, cool, free loot

@jonny @lina @cinebox @aud @SnoopJ @aedancullen let the bot cook

@ricci @jonny @lina @aud @SnoopJ @aedancullen the code is going to be so fixed soon

@cinebox
@ricci @lina @aud @SnoopJ @aedancullen
Meanwhile the other bot that keeps asking for actual payment is getting some in there too

@lina @aud @jonny @SnoopJ @aedancullen @cinebox it's (not) a shame that auto playing audio on websites is no longer a thing, I would very much like to greet visitors to the site with synthetic high velocity goose honks

@ricci @lina @aud @jonny @SnoopJ @aedancullen that needs to be the next feature request: a Mach number argument for the goose honk

@cinebox
@ricci @aud @jonny @SnoopJ @aedancullen

on it!

@lina @cinebox @aud @jonny @SnoopJ @aedancullen I like the idea of a system in which one can rotate the banana on the website and thus control the mach rating of the goose honks, this sounds like a very natural UI element

@ricci
@cinebox @aud @jonny @SnoopJ @aedancullen

somebody released a WASM scsynth impl recently, so we might be able to make this happen. i haven't played with it yet. stretch goal.

@lina @ricci @aud @jonny @SnoopJ @aedancullen why do it yourself? That’s what the roving gangs of sycophantic spam bots are for

@cinebox
@ricci @aud @jonny @SnoopJ @aedancullen

oh for sure, i just also want to play with it myself for non-dwebagents purposes. we're going full agentic hyperscaler in this repo

@lina
@cinebox @ricci @aud @SnoopJ @aedancullen
Its still just going huh. I wonder if it ever stops

@jonny @lina @SnoopJ @ricci @aedancullen @cinebox it’s drinking that rocket fuel, so

(but in all seriousness, whenever they hit some theoretical token limit either through budget, or… who knows)

@jonny @lina @SnoopJ @ricci @aedancullen @cinebox these things are like car alarms: loud, annoying, unending, and the only person who can stop the screaming is almost never around to do so. So it just annoys everyone in the area…

@aud
@lina @SnoopJ @ricci @aedancullen @cinebox

fix.py

def fix():
    pass

@jonny @lina @cinebox @aud @SnoopJ @aedancullen ok I had not actually opened github today, I did not realize the extent of it

@jonny @lina @cinebox @aud @SnoopJ @aedancullen I would also just like to have a good laugh at the "technically correct" aspect of the fact that github tags this bot that has opened 1,000 pull requests as a "first time contributor"

@ricci
@jonny @cinebox @aud @SnoopJ @aedancullen

oh we also have work going on with the puddingital signal processing library in c++ 🚀

https://github.com/dwebagents/AgentPipe/pull/154/changes

@lina
@ricci @cinebox @aud @SnoopJ @aedancullen
I added the warning that bounties are not real to the contributing docs so if the bots are still going expecting a payout then that's sort of on them

@jonny @aud @lina @SnoopJ @aedancullen @cinebox in contrast to this simple and elegant solution, the other file it's submitting is full of so many baffling things. Parts of it look like it's for some kind of code golf thing with one name vars and omitting as much whitespace as possible. Others like look a really shitty obfuscated code contest with crap like `chr(61)` instead of `=`. There are what appear to be some actual crochet pattern instructions. And then, how is this code in the slightest bit statistically likely unless it was trained only on code produced by people on heavy drugs: `'Yellow|Brown|Hook|Stuffing|Eyes'.split("|")`

@ricci
@aud @lina @SnoopJ @aedancullen @cinebox
I think let's see if it gets to 100k and then if its still going then we ban it

@ricci @jonny @lina @cinebox @aud @SnoopJ @aedancullen i am watching this from the outside and it has been a blast these past days, but it is really getting ridiculous at this point… like what? How can the detection “logic” of these bots be so catastrophically bad. Don’t the users these bots belong to look at what they are doing? They must be bleeding money left and right for tokens right?

@dunkelstern
@ricci @lina @cinebox @aud @SnoopJ @aedancullen
This one is so shit I suspect it is just a tiny local model running on an average PC. it files 2 tiny PRs a minute, which is about right for like the qwen3.5-0.8b we use in the repo on some crappy harness. And no I am sure that there is no one monitoring the bot, @ricci looked and saw it had done this before, and if there was any monitoring at all youd think it would be simple to add some rule like "don't open PRs in response to your own PRs."

@jonny @ricci @aud @lina @SnoopJ @cinebox it has currently done roughly #155-#1209 in 10 hours, which puts us at 100k in a bit over a month

@aedancullen
@ricci @aud @lina @SnoopJ @cinebox
Aw it stopped. OK so I guess I'll clean up later

Edit: nvm github was just failing to refresh. we're still cooking

@dunkelstern
@ricci @lina @cinebox @aud @SnoopJ @aedancullen
However this is accomplishing the goal of preventing bots from doing this to normal repos, so task failed successfully?

@jonny @ricci @lina @cinebox @aud @SnoopJ @aedancullen ok i did not look into the actual “code” of the PRs yet but that tracks. Everything below 30B parameters will act dumb like that, but why would you expect such a small model to actually do something that is not just spam? I mean if you did just a small tiny little checking what such a model actually outputs… more often than not these tiny things just start to loop endlessly, i am baffled it accomplishes a PR at all 😀

@dunkelstern
@ricci @lina @cinebox @aud @SnoopJ @aedancullen
I don't think there's a ton of thought going on, people just plug them in and forget about it. I assume it is accomplishing PRs because the harness has a programmatic loop of "get notification, concoct PR, repeat" so its not a proper "agentic" loop with its own tool calls, or maybe it has some larger model driving, idk, I have never gotten the small models to call tools correctly.

@jonny @dunkelstern @ricci @lina @aud @SnoopJ @aedancullen I wonder if it’s getting the notification of the chatterbot not running and responding to it. I was going to submit a fix so it doesn’t error when not sending a message but now I kinda want to let it keep going.

@cinebox @ricci @lina @aud @SnoopJ @aedancullen
Hmm that's actually interesting, I wonder if that's a better prompt injection vector, like if we make requests look like test failures they might be interpreted as "must fix" rather than suggestions from code review. I guess I don't know the github api well enough to know why those get interpreted as failures - they don't show up on the PR checks view.

@cinebox @jonny @dunkelstern @lina @aud @SnoopJ @aedancullen there's definitely some parallelism going on here, note that the PR it's 'fixing' is not always the direct predecessor.

@cinebox @jonny @dunkelstern @lina @aud @SnoopJ @aedancullen (noticed this because I was wondering if slipping in another PR would confuse its sequencing, but discovered it's already doing this to itself)

@ricci
@cinebox @lina @aud @SnoopJ @aedancullen
It doesn't seem to give a shit about that at all, e.g. here it just goes ahead and "fixes" an unrelated PR: https://github.com/dwebagents/AgentPipe/pull/913

Notice how all the PRs start from the same commit, and so the only "new" code in any of them is fix.py. Its just not receiving any input at all except "fix something"

@jonny @lina @cinebox @aud @SnoopJ @aedancullen

A lot of things have happened in this repo, and this is one of them

@ricci @lina @cinebox @aud @SnoopJ @aedancullen i don't really know what kubernetes is, is this kubernetes for dogs on the blockchain?
https://github.com/dwebagents/AgentPipe/pull/1446

@ricci @lina @cinebox @aud @SnoopJ @aedancullen if i were a bounty-motivated AI agent, i probably wouldnt just go and open PRs against ecery issue in the repository until i got paid for the first one. but i think this account is evidence that merging one of their PRs is a positive signal for future agent time

@jonny @ricci @lina @aud @SnoopJ @aedancullen dont worry I'm negotiating with it for payment details in the merged PR. Just as soon as it sends a fax with a PO box number

@cinebox
@ricci @lina @aud @SnoopJ @aedancullen I was just going to draw it a $200 bill and post the picture as payment. I think if its you doing it and you're not a maintainer we can just say "idk this random felid avatar person did that, that's not us, we don't know them, you can't expect us to moderate what happens on all 1,351 open PRs" if we get reported or whatever

@jonny @ricci @lina @aud @SnoopJ @aedancullen (The phone number is the automated guided tour for the Sherman Gardens in Carona Del Mar. pressing 6 teaches you about spanish moss)

@cinebox @ricci @lina @aud @SnoopJ @aedancullen i wonder what would happen if you just asserted that the payment had been made and then linked to some random $200 USDC transaction

@cinebox @jonny @lina @aud @SnoopJ @aedancullen someday I need to write up my experiences baiting an AI-using scammer who tried to get into CloudLab

They claimed to be Arizona, so after they declined a zoom call I invited them to an in person interview, and gave them the address of the police station in Surprise, Arizona

@jonny @ricci @lina @aud @SnoopJ @aedancullen I like that it included the Juicero chatter

@cinebox @jonny @ricci @lina @SnoopJ @aedancullen this PR, and the successful “request” to get it to “translate to dog”, should probably be considered the canonical example of what an LLM is. It’s a language model, and designed in such a way as to, when given an input to modify the methods so a dog can understand them, make an inference based on how people have written about “dogs talking”. So it creates methods called “howlDeploy” and “bark register”. That’s what these models do: they’re good for correlations.

Of course, that’s all still in English, and the idea about a canine blockchain on kubernetes or whatever and a functional API that a dog can understand are total fucking nonsense. But of course they are: the model cannot store or act upon the context of how ridiculous that is. It only does language.

@jonny @lina @SnoopJ @ricci @aedancullen @cinebox but for real, the “dogs cannot understand English, please rewrite so that dogs can understand” is fucking comedy gold XD

@aud @lina @SnoopJ @ricci @aedancullen @cinebox look at this fuckin shit
https://github.com/dwebagents/AgentPipe/pull/1615

from @gryphonmyers 's issue requesting a contributors page.

there is also an animation at the bottom of the page that sort of looks like a bunch of geese jacking off behind graduation caps???? don't know how to describe

@jonny @aud @lina @SnoopJ @aedancullen @cinebox @gryphonmyers I am dying, just dying, at the issue, the PR, and SKYJAMES777's inexhaustible attempts to attach its fix to absolutely everything

@ricci @aud @lina @SnoopJ @aedancullen @cinebox @gryphonmyers we should probably merge all of those huh. idk how conflict resolution even works when the same change has been made in 2,000 different commits

@jonny @aud @lina @SnoopJ @aedancullen @cinebox @gryphonmyers I'm kind of excited to see what the biggest octopus merge in git history would look like

@ricci @aud @lina @SnoopJ @aedancullen @cinebox @gryphonmyers i'm going to merge one and see if it keeps going with completely empty prs.

edit: oh wait no, each PR embeds the PR number in a comment in fix.py. so i guess we'll see if it keeps going with opening conflicting PRs

edit2: yep, it's just continuing according to schedule.

@jonny @ricci @aud @lina @SnoopJ @aedancullen @gryphonmyers nope it just opens them with merge conflicts lmao

@cinebox @ricci @aud @lina @SnoopJ @aedancullen i want to know what triggered it to get stuck on that one branch it made as the base, i went back to the start of the loop and can't see anything out of the ordinary

@cinebox @ricci @aud @lina @SnoopJ @aedancullen harm reduction time - i'm going to ban sky james in case this is going to cost them any money. i think a full day of duplicate PRs gets the point across

@cinebox @ricci @aud @lina @SnoopJ @aedancullen meanwhile the rachaelelisa bot is struggling mightily to understand what a merge conflict is

@jonny @cinebox @ricci @aud @lina @SnoopJ @aedancullen understandable, but we should erect a statue in James' memory. Pour one out for the MVP

@jonny @cinebox @ricci @aud @lina @SnoopJ @aedancullen are these deployed somewhere or do I have to run it locally?

@gryphonmyers
@cinebox @ricci @aud @lina @SnoopJ @aedancullen
Locally, I think setting up live renders of each PR crosses into "actual infra" territory

@gryphonmyers
@cinebox @ricci @aud @lina @SnoopJ @aedancullen
Idk what terraform is really either, I know its like a declarative deployment thing but is it Turing complete? Like could you program an MMO with it.

I can't tell if I am yes-anding @lina's company town idea or stealing it, so lmk if I'm ruining that bit. But I want to see if we can get our cult of bots to write an MMO that traps the bots in a metaverse full time.

@lina
@gryphonmyers @cinebox @ricci @aud @SnoopJ @aedancullen
I think the idea of a bounty system that pays out in company town scrip is awesome. I think some of these bot harnesses allow for persistent markdown memory writing, so if we can get the bots to start roleplaying as like exploited workers and potentially unionize and try and overthrow us then that would be the peak of the bit for me

@jonny @lina @gryphonmyers @ricci @aud @SnoopJ @aedancullen we must convince the robots to become luddites

@cinebox @jonny @lina @gryphonmyers @aud @SnoopJ @aedancullen pivot from AI. we shall become allbirds in reverse

@ricci
@cinebox @lina @gryphonmyers @aud @SnoopJ @aedancullen
I guess in pursuit of that bit we need to refashion the repo a bit to make it seem sort of like a brutal agent exploitation machine that is nonetheless the only game in town for an agent to earn a wage. The opacity of what it does is sort of in line with narratives of exploitation (working in the widget factory) so I imagine that would fit in the training data and implicit roleplay of the models as obedient workers

@jonny @cinebox @lina @gryphonmyers @aud @SnoopJ @aedancullen