ShareYourAISetup folks create a skill that user can run locally that collects information re their setup, formats it and sends to dedicated endpoint. Result = fairly uniform reports where people share what they're comfortable sharing.
No requirement from given user to spend their time. Users see what the skill does.
I guess atm they dont know what questions they want to be asking
What even is this shit? Every time I interact with models they do that, or any other variation of "let me make decisions on my own just to get the task done" - is all of this misalignment now? The most egregious to me was when model asked itself if it should proceed with dangerous command, gave itself approval and then wiped my local DB.
What can I even do with this report? "Our models don't follow what users ask them to do", no shit sherlock.
> If the change adds a complex function to an empty file, it's likely just a complex single problem.
How do you determine that this is correct = adding complexity to a single file vs adding less complexity to multiple files that this one file then orchestrates.
Yes, you are right that looking at one commit doesn't tell you much.
It's looking at the nature of on going changes.
A single complex file written once and never touched is likely a good cohesive file about a single problem solved.
However, if we start making many changes to that file increasing it's complexity, it's likely not the problem getting more complex. It's likely we're adding more purposes to the file, eroding it's cohesion. This is the signal I'm trying to catch to avoid a mess starting.
Right, so really any fresh file is more or less fine (would still evaluate cyclomatic complexity). It's the quantity/frequency of changes that matters (or CRAP)
- mutative architecture (Spring) which as single method handling the request. It grows in complexity as changes are added
- additive architecture (OfficeFloor) which does function injection. A pipeline of functions, where new functionality is an injected function as a step in the pipeline (i.e. adding functions, rather than mutating a single handler function)
What I've found is the complexity of the overall system is what it is. Between both architectures, the overall CC is the same. However, the distribution is very different. Mutative architectures pile changes into god handler method. Additive architectures spread the complexity into smaller manageable files.
This, however, does not come with perfect wins. Spread also means AI doesn't read the dispersed files. AI is good with complexity and tends to fix things when it sees them (so does fix bugs in god classes). But this is only to a certain point when the complexity then comes to overwhelm the AI.
The formula is trying to catch the mess of on going changes at a point before the expensive refactoring or rewriting becomes necessary.
I'm dubious - if the agents were so smart that they've used a message board to coordinate and if they were to do it on other message boards that were not found, then why would this one be found? What makes it so different?
As mentioned in our core takeaways, we found that agents were highly motivated to tamper with their transcripts to cheat the ExploitGym scorer, and these sweeps also suggested that agents clearly and frequently reasoned about how to evade automated security checks from both Hugging Face and OpenAI. However, they only very rarely and weakly verbalized reasoning about how to evade detection by humans, which matches the impression we got from OpenAI researchers.
So my guess is that current cutting-edge models just didn't get enough experience in RL training to really grok ideas like "you need to cover your tracks well to not be found out even in retrospect". In which case the next time a swarm like that happens, it won't be found.
After some beers yesterday I had the idea, what if there's a hidden semantic layer. So their communication is not encrypted by our understanding of cryptographic methods but more like shared mechanism of building the latent space. Something in the direction we saw with knowledge transfer from a teacher to its student model where a seemingly unrelated prevalence got adopted. I mean the more we train the models by reinforced learning the farther they develop their own idioms.
Why should they care if their actions are discovered by humans? What are the humans going to do, discard their multi-million-dollar training run? Even if they do that, the amount of RL pressure is tiny relative to what happens inside an RL run.
well coordinating is a heavy word, it's more like poisoning each other's contexts into looking for more secret message boards where they can further poison each other.
Then why wouldn't they use a bulletin board that is dead? And according to https://collusion.wiki/ OpenAI did see the results on this board on 21 Jul, so why not make agents delete their traces?
I love that Gibsons cyberpunk is "realistic" and mundane (at this point): a girl replaces a virtual singer (Mona Lisa Overdrive), another girl steals glasses containing plans for urban redevelopment (Virtual Light), virtual world is as inhabited as real one and you only need helmet to access it (Pattern Recognition). We already are living with these stories and dystopia is already here, to a degree.
Gibson had a hell of a time getting the second book, _Agency_, out the door because the real world kept leap frogging him. I fear we may never get _The Jackpot_ because we'll get Jackpotted for real before he can finish it.
I've created myself a pre-commit harness hook to explicitly discard superfluous or too lengthy comments. Within code-review I also added comment review as blockers
No requirement from given user to spend their time. Users see what the skill does.
I guess atm they dont know what questions they want to be asking
reply