Tootfinder

Interesting LLM nuance: why does using phrases like "you're a pen-tester" cause chatbots to emit substantially different predictions, and basically "follow" that instruction? Because of how LLMs work, this implies that the training data has plenty of examples where real humans told each other that they were some role and the humans just immediately jumped into that role without question or intervening dialogue. But that's not something people do in normal conversation. Even in playing-with-kids contexts if you drop that out of the blue you're probably going to get "no I want to be a robot" or "but you were the elephant last time!" rather than immediate assumption of the assigned role.
It's possible that training LLMs to predict immediate role-assumption is something the big models spent a lot of manual effort on. But what I think is more likely is: it's the legacy of role-play forums! All those reams of pages of teenagers (yes, often horny) pretending to be Captain Kirk or their own incredibly cringe "cool" character (but honestly, why call it cringe, let kids be kids and have fun)...
So next time you "tell" a chatbot "you're a..." to get it to do what you want, I'm pretty sure you have an RP forum teen from the past to thank :)
#AI #LLMs

The thing I love about computers is that I can tell you basically anything about how a computer works, not because I know everything but because I know how to figure it out. I can walk you, over the course of an hour, two, there, maybe more, though every step of typing something into a web form, from the electrical signals that get turned into digital via an ADC, to the USB controller memory, to the kernel driver, to user space, through the application stack, back down to the kernel, to the network driver, through routers, up the server stack, TCP/IP, key exchanges, etc.
I don't mean I have the time to dig into these things. I used to, and it was fun. I've given more than my fair share of interviews talking though variations of this. What I'm talking about isn't pure knowledge, but that, given relatively simple theory, and the right tools, every action of a computer can be understood down to the limits of physics.
The thing I hate about #LLMs is that take something comprehensible and make it something almost completely opaque. Even with a solid understanding of the theory, literally no one understands what's happening. That is shit. It makes playing with technology not fun anymore. The way in which companies are making things even more opaque by running stuff in the cloud is everything I hated about closed source on steroids.

Emergent #Misalignment: Narrow #finetuning can produce broadly misaligned #LLMs

“To waste what little bandwidth we have left – when 750 million people worldwide lack access to electricity – assisting some of the richest men ever to hone their plagiarism bots would be a historic idiocy that future generations are unlikely to forgive today’s leaders for.”
#StochasticParrots
#LLMs

You can't make this stuff up (or can you?):
#LLMs #AI #GenAI

On the one hand, LLMs are planet-destroying bullshit machines, gaslighting our sense of reality by spewing an endless supply of hallucinations.
On the other hand, they can provide a brief yet overwhelmingly mild sense of satisfaction when you convince them to do something silly.
#LLMs #DonnieTurnip

Are you in tech and outraged about generative AI? Is it being forced down your throat at work?
Here's a nice vindictive way to get a little revenge if you want:
1. Find a project that contains slop code.
2. Optionally, identify specific files or functions that are LLM-generated. I guarantee you that on average, this code has not been adequately tested/inspected, even/especially if it contains LLM-generated test cases.
3. Make up a reason the code could be flawed, bonus points if it's subtle or hard to test. Don't put effort into this or try to actually find a flaw. Just make something up at random.
4. Report your made-up defect as a bug.
That's it. If anyone ever questions you on the incorrect report, just say "oh I used an LLM and it said there was a bug so I reported it." (Don't actually use an LLM, that would be feeding the bubble.)
Note that you are showing the creator of the code the exact same amount of disrespect that they've shown you by publishing slopcode in the first place. I'd bet odds are 50:50 or better that if a human actually follows up on the report, even though they'll find out that the bug report is wrong, they'll find and fix some other subtle flaw in the LLM-generated code, so this is actually helpful in a way.
For step 3, try to get creative. Like "logic in decideUVParameters can cause state to be inconsistent in some cases." If asked for a steps to reproduce, either make one up if it's easy to do so, or say "I forgot how I triggered this." Surely they can ask an LLM to figure out conditions that would trigger the bug ;).
#AI #LLMs #GenAI

Dear generative AI enthusiasts,
Look, I know the tokens you're burning right now don't actually use *that*much energy (even though it's somewhat substantial already and disastrous when we take into account the quality of the crap it's being used for) but what's more important is the appearance (or not) of that token spend on the quarterly earnings report of OpenAI/Anthropic/etc. lays the foundation necessary for those companies to go ahead with their plans for datacenters on a truly ridiculous scale, and those datacenters, if built, ate indeed a climate nightmare which *my kids* will have to live through even if they never benefit from any of it at all. That's (one of many reasons) why I personally need you to stop using generative AI right now.
The fact that the output is crap, the way it erodes your intelligence, and the ways in which it plagiarizes and actively undermines good citation practices are among many other practical reasons not to use it, but what's personal to me is the way that your frivolous sloperation is making the future worse for the baby I'm feeding blueberries to as I type this, and half the time I interact with people like you the conversation begins with some form of "putting aside the ethical issues..."
#AI #GenAI #LLMs

"Using #AI to check the output of AI for errors is a method that is historically prone to errors"
No shit, Sherlock.
#LLMs
#StochasticParrots
AI Translations Are Adding ‘Hal…

AI Translations Are Adding ‘Hallucinations’ to Wikipedia Articles
AI translated articles swapped sources or added unsourced sentences with no explanation, while others added paragraphs sourced from completely unrelated material.

I have some sketches of an essay that I need to write, but I think it's worth brain-dumping a bit more in the mean time.
#LLMs are an attempt to make tech grow forever. But like, how many "your mom/a friend, but done by a precarious worker instead" apps do we really need? Everything right now is in the AI grift hole, but there's almost nothing of interest (even if you ignore the ethical concerns). Like, no, I don't fucking want a robot to lie to me about my groceries. That doesn't sound like a useful feature. There's a lot of useless shit being pumped out to prop up the bottom line, and a lot of people just want to be able to use their old phone for more than a couple of years.
No one is happy with this. No one wants this. Except the billionaires who are forcing us all to drink the capitalism koolaid, because they'd rather exterminate life on earth than live in a world where they experience consequences.
Nothing grows forever. That's not how literally anything in reality works, or has ever worked, at all in history. Some people think that the universe itself may work like that, but that's only an educated guess. Finite things don't grow forever. Every organism, every society, every technology, every dynamic and adaptive system we have ever known goes through a growth phase and then goes in to a stabilization phase. Or, following a Malthusian pattern, grows until it reaches a catastrophic point and collapses. Like lemmings. Or reindeer. Or cancer.

Was just made aware of this pretty good but entirely predictable study:
#LLMs