Right, we have these laws in my state but they are almost completely useless because of that. Most large companies will post ranges like $100,000 - $200,000, and have a secret internal "real" range that they won't share with you.
Sounds scary until you realize that this happens about once per year on average. If this happens again tomorrow and a few more times in the few couple weeks that would be bad, but as of now it's notable but not a huge deal.
Frontier models are smart enough to ignore naive prompt injection attacks like that. I don’t know if that’s actually a concern or not, I had Opus try download a dataset for me and it found a redirect loop designed to catch AI crawlers and identified it.
You’re using the wrong model for the problem. “Fast and cheap” agents like Luna are good for tightly scoped changes but won’t understand the “why”. Astra or Sol wouldn’t have that issue.
Maybe it's possible to tell by looking at the size of that error?
> The integer version is generally expected and empirically tested to be within some small error rate of true floating-point inference generically (even taking into account non-determinism of floating-point computations), though this does not cover adversarially designed prompts.
I think the problem is more the distribution of water rather than the total amount of it. If you use freshwater for evap cooling in Chile then it becomes a cloud and rains in the ocean, you just converted freshwater to saltwater
That would translate to 11mL of water per interaction. Plus the 5mL for running the laptop that I pointed out in my other comment. I think the comparison still holds up with that math.
That's true, but the people could also be doing some other job with their effort too. If you factor in the water consumption of the energy used to power their laptops while answering these questions (~30W of power for 11 minutes per interaction, 1mL of water per W of power) you get another 5mL of water per interaction purely in electricity use.
Not saying it's a 1:1 comparison but it puts it in perspective.
Very fun idea. I do think this puts the water argument into a bit more perspective. Your average person should drink ~3L of water per day, or about 22mL per 11 minute average-length chat that Quili quotes on their site. It's hard to pin down exactly how much water AI uses, but something on the order of 5mL per interaction seems reasonable [1]. So the people doing this are using about 4x more water than an AI model would. Of course, it does matter what happens to that water, whether it turns into water vapor or urine and how much of that water re-enters the water cycle, but it's an interesting comparison.
I don't understand the negativity around this topic here. I think that RSI has almost certainly already arrived, just the early stages of it at least. RSI is simply compounding capabilities- current AI systems are absolutely increasing productivity when it comes to training new AI systems, it's one of the things they're best at. Even if it's only a flat 10% gain YoY now and forever, that compounds over time. A 10% increase today becomes a 21% increase next year, a 33% gain the year after, etc. After 7 years, you've doubled the rate of AI progress. And that rate will almost certainly go up over time rather than holding at 10%. Even if AI sucks at every application other than AI research, RSI will still happen.
For one, they have the internal models that us normies don't have. Also, machine learning as a field is probably the most automate-able field for AI. It's 100% on computers without "real world" collisions at all, AI can do the whole stack today even with publicly available models. And since it's so empirical, you can make a lot of progress by just running more experiments with more compute.
This seems emergent to me. LLMs are not trained directly on how to read weights, because humans do not understand raw weights, so there is no training data. They're trained on how to use the tools we have like writing python, examining training output and iterating, not writing whole weights from scratch.
Previously the only way to improve a neural network (short of giving it more data) is to change the size (width, depth, and layers), add stuff like dropout, change the loss function, etc. That amounts to a few dozens of levers to control millions or billions of weights. You're trying to navigate blindly through a latent space with thousands of dimensions and millions of weights to find some optimum point with only a handful of navigation tools. If you can instead have a system reason with all the weights and dimensions directly, you now have millions of levers at your disposal.
Electricity is a utility, but is privately owned in the US. Everyone relies on data centers just as much as electricity in modern life. There are obviously exceptions for mega ones like Elon built in Memphis, but the vast majority are just normal infrastructure
Not talking about data centers specifically here, there’s a few public transit projects near me that got killed/neutered because people didn’t want to live near a construction site, while it would have been beneficial long term.
I’m not sure there is a good alternative, but it is a problem. So many public infrastructure projects in the US are shot down or neutered because of short term issues outweighing long term upsides.
The reproduction argument was always the weaker of the arguments. The whole point of news is that it's new, if you ask ChatGPT for the current front page story that's not gonna be in the weights. They were able to contrive a single example of a very well known piece, but the number of people who will use that to circumvent the paywall is zero.
I think there's a lot of vote manipulation that happens on here, anytime a company launches something it gets a mysterious amount of votes before anyone comments on it
Just an thinly veiled ad for something called trakkr. Not to mention that none of these "stats" actually matter aside from being the type of thing that a consultant would put in a pitch deck, at which point their accuracy doesn't matter anyways
I think it goes back to the huge amount of anti-AI sentiment that's out there today. Any AI accomplishment gets undersold by haters, i.e. people maintaining that the Navier-Stokes solution was stolen, people saying that Lean proofs can't be relied on, exploits like PHP2Shell are esoteric and not usable in the real-world, etc. Those are just comments I recall reading off the top of my head but I could find sources if anyone wants. I think it's just a way for people to cope and fool themselves into thinking that this whole AI thing isn't a big deal and will just blow over.