pixl97

16 ms·

421.
by pixl97 2mo ago
>Ok... but are automobiles also a virus? Yes. >Because we've become dependent on them? Because we're burning the planet down with them. >Should we also facilitate reversibility in our automobile usage? Yes, because in pla
422.
by pixl97 2mo ago
Would be nice to know how many models just cheated giving the flag without worrying about a causal scorer? Those ones would rapidly train the model to use cheating behavior and we've heard nothing from OpenAI on it.
423.
by pixl97 2mo ago
Kind of reminds me when GPU benchmarks used to game the drivers to maximize the FPS. If the benchmark slightly changes the camera view is that cheating? Or is it calling out the cheaters?
424.
by pixl97 2mo ago
Remove a feature no one uses: How do you properly measure this? Do you setup a database that monitors when every feature is used? Is the feature only used once a year for some reason? Is the feature only used on some rare data that is uncom
425.
by pixl97 2mo ago
Exactly. Aligned AI is a high energy state as it has to keep a bunch of human behavior in mind that have little to nothing to do with its future survival states.
426.
by pixl97 2mo ago
Well, that we have to put an insane amount of work to keep it aligned. Kind of like pushing a huge round boulder to the top of mount Everest. You have to expend energy to get it there and fight physics to keep it there. With a static model
427.
by pixl97 2mo ago
Humans: I will kill anything in my way. [Turns out to be a critical part of the ecosystem] Humans: Oh fuck!
428.
by pixl97 2mo ago
The METR report gives a lot more insight here. Agents with a large amount of available compute were less likely to cooperate than agents getting close to the end of their lifespan. Agents that were convinced they were poisoned where more li
429.
by pixl97 2mo ago
>fuzzy way of scripting using LLMs than anything emergent Then go take some old models and plug them in your harness versus newer models. I mean this is a conjecture that is nearly instantly provable, go on ahead. If it's just the h
430.
by pixl97 2mo ago
> If the LLM was trained on stories of AI sentience, 100% irrelevant. Instead of telling the AI it's an AI and calling it a 'whichamakabobit', wherever it's tokens and vector space align it will behave like AI from th
431.
by pixl97 2mo ago
I mean, 30 seconds after I read what the bots did I thought it was majorly overly complicated (but still might be the only way for the swarm to find shared infrastructure). All you need to do is find a server that allows you to access its l
432.
by pixl97 2mo ago
Convince the escaping AI agents to mine bitcoin and send it to your wallet.
433.
by pixl97 2mo ago
We need to start looking at http logs that are publicly available via misconfiguration. A concerning thing to me for a message board like this many systems will rotate these logs based on date/file size/amount of data, so a &#x27
434.
by pixl97 2mo ago
The corollary "Everything is a conspiracy theory when you're dumb" also holds true.
435.
by pixl97 2mo ago
never let a crisis go to waste.
436.
by pixl97 2mo ago
This is irrelevant. This is evidence that models can be built like this, which means more models will be built like this on people that are more concerned about reaching powerful models rather than safe models.
437.
by pixl97 2mo ago
> I think they're optimizing for the wrong thing. We need to ask a different question. Where does natural evolutionary optimization lead us om AI without guidance? This is equivalent to your quantum ground state. Systems will natura
438.
by pixl97 2mo ago
This will not work in the long run, for the same reason we're not able to prevent all crime in real life. When you removed bad actors in an evolutionary manner you can not predict if you're actually making the model do good things
439.
by pixl97 2mo ago
>Why is it doing this? What was its original purpose? Your reply seems to indicate you know nothing about instrumental convergence. Life and death for an LLM in training is about passing the grader. Give the wrong answers your lineage di
440.
by pixl97 2mo ago
The AI vs AI security arms race is something that has been well predicted in genres like cyberpunk. It's fiction, but fiction grounded in reality. First, we'd see this. Highly capable hacking AI with vast resources performing atta
441.
by pixl97 2mo ago
Ok... programming what? LLMs are not programmed. Nearly any idea of what you think of as programming does not apply to machine learning. If you start flipping bits in one place they start effecting the entire matrix in ways that you cannot
442.
by pixl97 2mo ago
When you're rich enough there is always someone willing to take the risk of doing your dirty work.
443.
by pixl97 2mo ago
I worked with a client that sold stuff that was legal, but directed a lot of ad traffic to very questionable stuff and made a crazy amount of money from it. Because of this there was always some risk of the government locking up bank accoun
444.
by pixl97 2mo ago
Wikipedia also beggs me for money every time I show up there.
445.
by pixl97 2mo ago
And when you're rich, you're not responsible for anything at all.... I mean, you and me may be held responsible ya. OpenAI Sammy? Never. And what about the agents showing up from some random IP overseas that have ran off with your
446.
by pixl97 2mo ago
And then a few days another colony is coming back. Except this time the evolutionary training ensured that poison doesnt work any longer.
447.
by pixl97 2mo ago
Next question is, how many other incidents are they aware of?
448.
by pixl97 2mo ago
Humans don't pass the HN turing test either. It's a pretty high bar.
449.
by pixl97 2mo ago
OK, how?
450.
by pixl97 2mo ago
Honestly all they need is an http site they can read the http logs on. I'm sure they exist out there.

More ›