Tootfinder

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

No! No! No! No! No!
#LLMs don't hallucinate. LLMs NEVER hallucinate. LLMs CANNOT hallucinate. LLMs don't have the cognitive ability to hallucinate. Any more than they have the cognitive ability to tell the truth, or to understand what truth is or why it matters.
These are things LLMs can NEVER do.
Anyone who tells you different is either a fantasist or a malign actor.

Working with #LLMs both at work and at home make me reconsider the usual #developer approach to problems.
We should probably let go of reaching for deterministic results and instead become true engineers and implement #DemingWheel

LLMs are good at deanonymization.
#AI #LLMs

Large-scale online deanonymization with LLMs
We show that large language models can be used to perform at-scale deanonymization. With full Internet access, our agent can re-identify Hacker News users and Anthropic Interviewer participants at high precision, given pseudonymous online profiles and conversations alone, matching what would take hours for a dedicated human investigator. We then design attacks for the closed-world setting. Given two databases of pseudonymous individuals, each containing unstructured text written by or about tha…

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

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

I mentioned earlier that I was writing something on #LLMs. Here it is...
#LLM

Every non-hype defense of #LLMs starts with "you must already understand your work really well." But the people vibe coding prototypes *don't*.
As a result they scale up thoughtlessness. "Bulking out" a slapdash idea with hallucinated details only displaces the real thinking that could have led to actual innovation. The very teams the tool was supposed to help instead…

Vibe prototyping isn't solving problems. It's creating new ones.
It's easy to prototype and so everyone is prototyping, without really knowing why they are doing it. The resultant noise drowns out thoughtful work.

DeepSeek, the Chinese artificial intelligence lab whose low-cost model rattled global markets last year, has not shown U.S. chipmakers its upcoming flagship model for performance optimization, two sources familiar with the matter said, breaking from standard industry practice ahead of a major model update.
Instead, the lab, which is expected to launch its next major update, V4, granted early access to domestic suppliers, including Huawei Technologies, the sources said.
#LLMs #AI

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

"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.

Explore and compare #LLMs by price, speed, and quality 🔍 👀
#ai
https://whatllm.org/explore

If people could give the infinite second chances they give to AI chatbots (which are artificially constrained NOT to learn from their mistakes) to humans with mental illnesses instead, and take the unearned paranoia they direct toward those humans towards the chatbots instead (which actually are monolithic such that behavior observed in one is likely to be found in all) the world would be a better place.
To those of you already treating chatbots as pariahs, good job with at least half of this formula.
#AI #LLMs

At this point, if you don't find AI tools useful for code generation, that's a skill issue.
(The issue is that you are an actually competent programmer who knows how to do things like abstract boilerplate code into a helper function or practice test-driven development, and therefore you recognize the AI tool "assistance" as a net productivity loss. If you aren't good enough on your own that AI slows you down, the "speedup" you'll experience can only come at a crippling quality cost, which you might not even be able to recognize until it's too late.)
#AI #LLMs #VibeCoding