missingcolours

Imagine that AI is just another tool. Imagine it's a weed whacker. Who is liable if a weed whacker destroys property?

Well, it depends. If I'm using it according to the instructions and the tool does something it's not supposed to, the manufacturer is liable. If I take a weed whacker and start cutting fiber lines with it, I'm liable.

So in accordance with standard liability practice, AI labs ought to be liable for unintended actions their products take, and the user for illegal actions they direct the product to take.

The slight caveat is that a model just produces text. It can't hack anything on its own. I would say that the product in this scenario would be the harness/agent - if you build a system that takes untrusted input (LLM output) and takes potentially dangerous action based on it, well, that sounds dangerous? Possibly negligent?

The corporate dilemma between "I hired some less competent engineers and they're dragging the team down" and "I hired only highly capable engineers and now I have to give 10% of them bad reviews in stack ranking and later let them go"

I'm curious what the trend for whites in the US is over time. That would be a good first approximation on how much of this decline is due to culture or the educational system (including COVID disruptions) and how much is demographic change/immigration.

Doesn't seem to have incentivized cryptocurrency enthusiasts away from being in that last category for years...

That doesn't seem untenable at all. It's perfectly consistent with legal precedent and current practice.

There are MANY cases where the government legitimately possesses personal information that could potentially cause harm if revealed, for operational as well as public safety purposes. Vehicle registration and drivers license records, for example.

This personal information is typically exempt from FOIA laws. I can't go FOIA vehicle registration records to look up where a certain driver lives, but the state and police officers do have access to that information.

A search involves access without consent. Police don't need a search warrant to perform a search with the consent of the owner. In the case of cell phone data the search is without the consent of the data owner (the phone company). In this case the municipality is collecting public-view data themselves, not without the consent of the collector, so there's no search warrant issue.

Common complaint is that government enforcement of laws is insufficient, i.e. fines are only x% of their worldwide revenue. Sounds like this way has teeth and might force companies to actually obey the law?

I was wondering about that. I've taken a number of different cold medications over the years, one active ingredient at a time, to check their effectiveness while sick. Phenylephrine, as widely reported, had zero effect. But Dextromethorphan seemed to work for my kids - give it to them, stopped coughing. Started coughing again 6 hours later, etc. n=1 and it's a 4 year old, but in one of those cases my results clearly matched the studies, and in the other it didn't, at least as reported by this article.

What I've been doing recently:

1. Vibe code a codebase that does what I want at a high level.

2. Iterate with LLM on this codebase to add features, fix bugs, improve performance, and address issues until it basically does what I want, but the code behind it is often a toxic waste dump.

3. Take lessons learned from vibe coded version and implement by hand. For challenging areas (writing a complicated algorithm) that would require a lot of thought or brainpower, I'll sometimes ask the LLM for a reference implementation and then modify it to suit.

This is a big speedup on manual code because you've figured out all the question-marks ahead of time and have a functioning blueprint you can refer to.

It's not as fast as having LLMs do all the code of course, but I find it to be a considerable improvement over doing everything by hand, while still letting me write code I'm comfortable with and understand deeply.

The other angle is to be very specific with the prompts and then dive deep into the code output and keep asking the LLM to change the code structure in various ways to ensure you get code you like. I found that to be frustrating and painful. Maybe in the future I'll write some really good prompts and future models will be better at following direction, but I haven't been happy with the results of that approach to date.

The people doing the voting are mostly talking about how much water datacenters supposedly use.

As I have snarkily observed at work: if I go $100 over the meal allowance on my business trip, I'll have to have an unpleasant conversation with my manager or finance. If I use $500 in AI tokens unproductively I'll be recognized for being a top AI adopter.

Certainly disagree with "AIs are not good at writing Rust". We can discuss the pros and cons of AI coding in general but in my experience they do just as well with Rust as any other language. If anything I'm impressed with how seamlessly the models can work with Rust's ownership model.

Yeah, same in the US. I would not leave my laptop in a cafe in SF. I would in the Austin suburbs. I might or might not in Chicago depending on the neighborhood and cafe.

In many of those examples, there is payment to the creator of the works that others are learning from. Authors are paid for their books, when we listen to music on the radio the musician is paid royalties, etc. When you lead a team and mentor junior engineers you're being paid for your time.

The nature of the source material matters though. Training a model on open source software seems perfectly fair - it has explicitly been released to the public, and learning from the code has never been a contested use.

IMO the questions around coding models should be seen as less about LLMs and more as a subset of the conversation about large companies driving immense profits from the work of volunteers on open-source projects, i.e. it's more about open source than AI.

If they can't tell their suppliers not to depend on Anthropic models and services, then... yes, potentially?

Right, usually law enforcement gets chat logs from a participant (search warrant for a phone, informants, undercover FBI agents, etc) and uses the metadata to connect messages to a real person's identity.