If someone spends significant effort to gather documents proving that their family was forcibly relocated from Poland, they may be able to become a Polish resident and then spend a year or so doing paperwork, bringing all of the documents that are required according to the official web site for some task to the appropriate government office where they are then told that other documents are required, or that nobody in that office even knows what documents are required, and so on, and after that time they may achieve Polish citizenship. You know, in recognition of the fact that their family is in fact Polish. But during that year or so, they may have trouble using the banking system because of sanctions on Russia and because no Polish bank will serve them until they become Polish. So their employer may be on the lookout for alternative payment rails.
These example queries really do take fewer tokens than the SQL queries, and the syntax is apparently designed to minimize token count.
My impression is that letting the model write SQL queries is not a typical baseline for RAG. It's unlikely that a company could substantially benefit from this reduction, because the cost of their internal LLM usage is probably mostly tokens that are not RAG queries or tabular responses. They also probably cannot fall back to a cheaper model because the "minimum acceptable level of intelligence" of the model is likely chosen for tougher constraints than the ability to generate correct RAG queries.
I don't think this is a reasonable thing to make a business around, because the claimed benefits mostly do not exist.
This question in particular could use an answer about like "use an up-front design that will not prevent subsequent performance work", so it is sad that it has two similar answers.
But also, it's not really reasonable to map all these opinions onto two axes
> You focus on efficiency and performance. You like to work close to the metal, optimize for speed and memory usage, and prefer direct control over system resources.
well, it's not wrong, but i feel like a lot of these questions are "too normative" in their framing. there is some stuff that i do because of the nature of my work, and some other stuff that other people should be enabled to do because of the nature of their work, and some other stuff that really sucks that nobody should be doing and is nonetheless quite popular
Memory leaks are unrelated to memory safety. That is to say, code that leaks memory is memory safe. So I'm not sure what RAII is supposed to help with.
A problem not solved in C++ is the need to reserve a single bit-pattern per type that can be moved from, to indicate that it has been moved from (and is not a valid value for any other purpose).
Well, for example you may have some functions which accept types and return types, which are not compatible with some input types, and indicate their incompatibility by raising an error so that compilation fails. If the program actually does not pass some type to such a function that leads to this sort of error, it would seem like a bug for the compiler to choose to evaluate that function with that argument anyway, in the same way that it would be a bug if I had said "template" throughout this comment. And it is not generally regarded as a deficiency in C++ that if the compiler suddenly chose to instantiate every template with every value or type, some of the resulting instantiations would not compile.
I don't have any experience running the Vite or Bun bundlers or the typescript compiler on the client, but I think these are not really supported use cases.
I think it is a bit more troublesome to ship a this comptime implementation and a minifier to the client so the specialization may be performed there than it is to ship just a c preprocessor implementation.
I have read the examples, and it seems like this cannot be used for aggressive hoisting of conditionals by writing "if (comptime foo)", resulting in the the body of the if statement being executed unconditionally or omitted. So it cannot replace my current use of C preprocessor macros in Javascript, though Zig's actual comptime feature could.
I have asked for like the ability to apply a filter that maps real strings from real users to a very small number of character classes, so that I may run some codec on data that is equivalent to user data for the purposes of the codec. I have heard, this is not a good use of my time, and if I really care so much I should instead make various guesses about the production data distribution of user-created strings (that need to be json-escaped) and then we can deploy each of them and keep the best, if we can even tell the difference.
Yes, it is very important to test performance work on data distributions that are representative of the production data distribution. I hope the author has another go at using SIMD that actually makes it into use :)
i replied to a comment by the hacker news user called pyman which claimed incorrectly that distillation was repackaged as "subliminal learning". so if you are asking me, who is talking about subliminal learning, which is unrelated to the topic of the article, the answer is that the hacker news user called pyman is doing that.
In some workloads, you can benefit from paying the latency cost of data dependencies instead of the latency cost of conditional control flow, but there's no general rule here. It's best to try out several options on the actual task and the actual production data distribution.
A while ago I talked to some team that was planning a migration to GraphQL, long after this was generally thought to be a bad idea. The lead seemed really attached to the "composable RPCs" aspect of the thing, and at the time it seemed like nothing else offered this. It would be quite cool if capnproto became a more credible option for this sort of situation. At the time users could read about the rpc composition/promise passing/"negative latency" stuff, but it was not quite implemented.
Really great. I wonder, for the "types encoded as code" approach, is there any benefit to fast paths for data with fields in ascending order? For some json parsers with types encoded as code I have observed some speedup from either hard-coding a known key order or assuming keys in some order and providing a fallback in case an unexpected key is encountered. For users who are stuck with protobuf forever because of various services using it and various data being encoded this way, the historical data could plausibly be canonicalized and written back in large chunks when it is accessed, so that one need not pay the entire cost of canonicalizing it all at once. But of course the icache concerns are still just as bad.
Hello, a friend of mine is a strong systems programmer from Poland but has not graduated from college. Is he basically unable to get a work visa for the US?