Suppose you run a platform, and you run a couple thousand different queries of different types throughout the day. It would make sense to have an auto-optimizer that would read long queries, ponder over them with an LLM, come up with some good plans, and store them as hints. This seems like quite a good idea? Is there a product for this?
It really doesn't sound good. A loss leader is meant to get you into the store; the gas station is usually far away, in the back of the parking lot.
Their $5 chicken now, that's a great "loss leader" (if it even is; that isn't confirmed). Because that forces you not just into the store, but all the way to the back. So you go in for the $5 chicken and leave with $20 of blueberries, $20 of cereal $20 of milk $20 of mayo.
The optimizers are lifted from convex optimization, but the point above was that they are applied to highly non-convex problems. They work for finding local minima, but a lot of the deeper literature does not translate (e.g. the conjecture being discussed in this post).
Encrypt the data with an asymmetric key; and keep the decryption key somewhere offline. You can get a hardware token to store the key on (I think a yubikey can do this).
Comparing the US and Sweden, it's also useful to know that the proportion of refugees accepted by these two countries is wildly different. Sweden has historically taken in many refugees (including draft dodgers from the US). In 2015 (an outlier) they accepted rouhgly 1 refugee (163k) for every 60 people (~9.4m) in the country. At its peak in 2024, the US admitted 100k refugees, significantly fewer than the Swedish peak. The impact of refugees is much more visible (also in budgetary allocations) in Sweden than the US because of this difference.
Why are you just comparing refugee numbers? That's incredibly disingenuous. A very large portion of the US's population is here illegally, and it would be significantly larger if we didn't have birthright citizenship.
> The hallucination problem is the difference between a clever toy and a system a hospital or a bank or a court can actually rely on. It is the whole ballgame for enterprise AI.
It... isn't? Hallucinations are surely a limitation of LLMs, but I haven't heard people worrying about it as some kind of existential question for a long time. You accept it's a non-deterministic system. You build appropriate safeguards or deterministic checks around it. And you accept it's not perfect, there will be occasional mistakes. No large enough organization can claim determinism for any sufficiently large system.
This is cute and all. But for anyone coming here for real-world advice: just use a regex, normalize to lowercase, and surface any errors to users so they know if their email got rejected. This will avoid 99.9% of issues and work for 100% of real human users. This is what everyone else does, and if you have a user with an esoteric email, they will still be able to furnish another one that passes this validation.
Verify all email address entries before you start using it... I absolutely HATE how much garbage I get because a few people don't understand you actually have to get an email address before you start using whatever you like.
Right, maybe I worded my response a bit vaguely. Of course you need to do an opt-in and verify them.
My point was rather that if you are operating a service with human users, there is no need to deal with quoted local parts, mixed-case, non-ascii, etc. You will just run into bugs (oh, the user signed up on an iPhone and the email field was auto-capitalized and now they can't find their account?) for almost no marginal benefit.
The argument is that rate of return only regresses to the mean if you diversify. It's like saying "I hold stock in a single company, so if I hold long enough, my return will average the return of the S&P 500". Your returns will look like the mean only if you diversify across investments, not over time.
reply