FieryTransition

That's very awesome, I feel like it would be fun to make a feedback loop with the LLM outputting syntactically valid programs which also log the status of various invariants and pre/post conditions as the program runs to validate correctness and maybe even train a LLM this way.

Would be interesting to see how far it would get writing programs, and wrt. the problem of stopping too early on a token, it could resume with some extra context from where it stopped.

Maybe construct program such that it is made by building blocks which can fit into context, such that each block/function has a pre/post condition and invariants are inside it, which the LLM will try to test against each time it runs. I think I just found my next side-project, I know there are many similar tools, but I haven't seen anything trying to couple the compiler, running the program, creating pre/post and invariant checking against emitted code yet.

Would be interesting to test this hypothesis to see if the llm can actually build a program this way. I think about it like humans have a short-term memory as well, and for each sub-structure of a program, the llm would have to work in a similar way. Then reserve a bit of context to control the long-term memory for general goals, or create something like an AST of thoughts which it would recurse through, while reasoning about it.

Does that mean you could make any model at least always produce syntactically correct code as output?

I live in one of the myriad of countries in the eu, so I'm confused about what you mean when you say protectionist taxes and which country you are referring to or just something else?

Best thing I did for privacy was to use things like nextdns and pihole, they block alot of the stuff with minimal effort.

I live en Europe too (northern countries) and we almost never handle coins anymore, but nickel itself is in so many things as an alloy. So I can imagine it must be very annoying to buy something and suddenly realise you can't use it due to a rash, that does indeed suck.

Mixed interval training for phone holding, I sincerely hope that was a joke, but judging by the last paragraph, I'm not sure :-D

You might be allergic, I would get that tested, but it's nothing to worry about, just nice to know. AFAIK most people shouldn't react badly to prolonged skin contact with aluminum in the range that is holding a phone.

yeah, I had some weirdness with the UI and disconnects as well. My takeaway from trying it was that I wouldn't want to use it for something if I need peace of mind for my data.

Used it for a while, recently tried to restore some things and it failed, taking a really long time to restore some snapshots compared to other things I've tried. Switched to restic instead. Really like what kopia is but I'll wait a few more years before considering it for something, but right now I'm happy with restic.

Too bad no one besides kopia does ecc, which is the reason I switched, but when I checked out why restic didn't do it, it was because they saw what other people did and apparently it's way too easy to make a bad implementation.

Does unpoly try to solve the same problem and how does their approach differ compared to htmx?

Curious to hear thoughts, I haven't done web in a while, so understanding your thoughts might help understanding the whole thing better.

One way to do it is to use cosine similarity[0], the reason to do this is to get around the context window limitation, and hope that whatever text chunks you get which via the similarity function is the correct information to answer your question.

How do you know that Claude doesn't do this? If you have multiple books, you end up with more than 100k context, and running the model with full context takes more time so it is more expensive as well.

[0] https://en.wikipedia.org/wiki/Cosine_similarity

Yeah, it's quite interesting how energy works, AFAIK we still don't know if task fatigue is really a thing or not. But if it isn't, doing more in the morning would not leave you more tired than usual, assuming you get regular good sleep.

The body can be affected by so many things, and it's different for people what affects them. What you eat in the morning, sunlight exposure and consistent bedtimes is what worked best for me. I get an afternoon slump no matter what I do, but that is usually remedied by some physical exercise or breathing exercises I've found, and the longer I go without taking naps, the more my natural rhythm gets used to not sleeping at that point. But I've had periods with extra stress where I just need to take naps, but that is just something I've accepted and enjoy now when it happens.

On the whole, I agree with you, but sometimes there are actually issues and that's why we get sick, when the body response is out of wack or lacking. Sometimes it's less about being tired and more about something which causes you to be tired, like being stuck in an unwanted situation.

But yes, our work environment are full of dogmatic rules and preconceptions which are hard to break.

I often wish I could sort Hacker News into two categories. Actual software/tech/STEM and everything else. I think both are interesting, but often, the niche tech stuff gets drowned out fast. So this is great for that :-)

I wish that the regulations would exist in America to begin with, and the American people would push more for it. Instead we have to adhere to whatever their laws are and try to play catch up with regulations. And when we threaten with lockout or gating, they threaten back to exit the market completely, because they don't have any ramifications in their own country, and are not super interested in the issues which the regulations try to address.

But is a Hilbert system easier to run in a computer since there are many axioms which clarifies the cases where contradictions or some inference rules are used which might be hard to program? For example, on the top of my head, in predicate logic, how would you encode an inference rule like exists elimination? It's tricky enough for me to think about, but to be able to encode a computer with it, is harder, and I can't remember if something like this falls under the undecidable problems category or not.

I think sim cards have been able to run java for quite a while[0], and who knows what else they can do. So I guess encrypted post cards it is.

[0] https://www.extremetech.com/defense/161870-the-humble-sim-ca...

You can fine tune a open source model for your task and achieve better results, at least, instead of just using them directly. But they are still not close to the openai models in generality. Huggingface is the place for exploring models, recently went through a lot of them for my use case, and they are simply not good enough, yet.

AAA devs target the consoles level of performance, and its been like that for a while. A lot of "extreme" settings in games are usually just poorly optimized and sometimes even looks ugly or almost does nothing artistically or experience wise.

Wouldn't that be too big? A typical failure point of all my wired headphones/buds has been the connector or whatever the connector plugs into. A bigger connector creates more leverage, which in turn creates more force, which has been a problem. So for mobile usage, wireless audio seems to just eliminate a whole range of problems and failure points, so it's not all good to use physical connectors?

My hunch is that they combine it with some kind of web crawl and index for source reference, I've used both gpt4 and phind, and find phind better for tasks like api exploration and questions for typical sample programs and error messages for that reason.

You are right I think, it's about impulse, force over time, surface area and how much energy can be transfered through the surface area through the underlying tissue and create a ripple basically. I suspect that at some point, the 'transfer function' of energy just maxes out and mass of the vehicle itself is pointless over that point, as it's all lethal.

Make.md is trying to solve this, you could take a look, I tried it, but it's still not stable for my use case, but apparently they have people who use it successfully.

I guess you could argue, that most places suffer from that, as everyone has a boundary, at which they don't understand things. But the effect itself, is probably more relevant in an organization, as it was used to describe management.

I would worry that if the infrastructure to access these things would be taken down, either by law, war or some kind of great depression, people would value other things than a picture of a monkey and how to keep IPFS running.

I've known a dude who taught at university on the side, besides working on logistics of all things, he said he had a mild form of dyslexia, so he taught himself to read books, by reading each page, on each side at the same time, as some form of compensation. I don't know if I could believe it, but there's document cases of this, and his retention and speed was crazy fast. Could be that he just taught himself to glance and chunk structures really fast, which is what I do, if I try to read fast without regard for safety or pleasure.

I totally understand what you are saying, in terms of probabilities, but what I'm arguing against, is thinking that safety, robustness etc. for a program is just a given, just because you happen to use Rust or whatever else is claiming X feature. In it's essence, it's like saying something is made of military grade materials, but then the materials are used all wrong in the actual design or not of importance at all. The fact that someone used Rust, does not really tell me much, other than certain types of errors are eliminated, and that is assuming they used Rust correctly, for the type system to work properly.

If you ask people explicitly about whether Rust just makes better products, they might say that they have doubts, or give a vague answer. But asking like this is just a leading question, and therefore the process itself of asking the question, will probably cast doubt over some automatic assumptions about what the type safetly of Rust really does for the end product.

But the reality is, "Made with Rust" is starting to look like a certificate/claim of validity, and its definitely not, just like any other language isn't either.

No language will save you from: - A lack of knowledge about the domain, this includes the technical framework in which your program exists, behaves and interacts - Code errors which result in valid code, but does the wrong thing, Rust isn't magic - Bad choice of datastructures and algorithms - Incomprihensible code documentation and bad development practices.

While I get that languages can eliminate a type of problem, they don't eliminate badly written software. Most of the time, the software development process itself, is more important, if you really want to examine what it means to build secure/safe software.

Having a team dedicated to developing software safely and using a language with stronger type safety guarantees vs. having a person just build something with the same language, does not magically imbue the end product with good qualities.

Take your example, if you don't have a good foundation in distributed systems, and try to build a distributed system in Erlang, you will only get so far, before you hit your head against the wall so hard, you start studying distributed systems instead.

Just reread what you meant, yeah, I see that self-signing wouldn't neccesarilly solve the issue. As you say, it might also be that some apps use root status. After they updated our digital signing platform, a colleague who had a phone from china which wasn't even that old, stopped working, and I've had friends which had the same thing happen. They weren't rooted. So I don't know if there is some hardware component to it, like TPM that you mentioned? It's worth saying, you cannot use banking apps here, without a valid digital signature, which proofs your identity. It's not just the financial app itself, it's the legal requirements we have here, which make the financial apps use this digital identity verification.

So if self signing cannot guarantee those assumptions that are being made, there's no way around it.

true, it gets tricky when there's an online component to it, since they can just keep their key secret, if the service I'm using is also online. Would there be no way to spoof it? Like,

Client (validate)-> server, requires a valid signature which I cannot sign unless I have access to their private key

Client <-(SpoofedAuthSuccess) SpoofServer, is also impossible if the client requires data from a server to work properly going forward. The only thing you could attain, is to unlock the client locally if you reverse engineered it, but any data not stored locally, is impossible to get. So wrt. games, since you mentioned DRM, it might be possible to unlock the content if it is local, given a clever reverse engineering solution? Even if they encrypted the data on disk, at some point, they have to decrypt it in-memory locally.

But for server, where all validation and data is gated behind an 'authoritative' server, I guess it's just game over for unlocking anything yourself with a certificate. Even if you manage to magically solve it, they will just issue a new certificate, and quickly invalidate the old one I guess.