captainmuon

I work at a university data center and we host open weight models (as usual, this is my private opinion). I would consider open weight something you can take and self-host for free, also commercially and also that you can use for training and research. Anything else is just, I don't know, just "usable for free in certain cases"?

Open source or source available is something else entirely. To consider something source available I would require the actual training data to be available too. That is exceedingly rare.

It's cool that you can do this now with constexpr instead of a pre-build step. The idea is pretty old though. I remember reading an article in a computer magazine in the 90s about how a web browser (Mozilla) works. They made a big point about how difficult parsing tags is, and that you can't just `if tag == "HTML"` for performance reasons. Instead they pre-baked an optimal hash table and used that. I remember it vividly because that was my first contact with hash tables and the explanation was pretty good. Looking back I find it funny that they considered tag lookup the hard part of a browser :-).

After squandering the national helium reserve for years, because "lol free markets" ideology. It's crazy that we put that stuff into balloons for fun (helium is pretty abundant in the universe, but the economically viable sources are fossil).

What they are doing here is selecting a number of events (particle collisions) and comparing with a simulation of the standard model (sans Higgs or whatever they are looking for). If there is a significant excess, then you found something.

One scenario, and I think what people were paranoid about, is that you could make your cuts really arbitrary, and select a region where you by accident have too few simulation events and too many data events - say "Select all events where the first muon has momentum > 15.6 GeV, but not between 20.1 GeV and 20.3 GeV, unless there are 3 jets, in which case the selection is...". You just try out so many regions until you find something. For this reason you need to keep every step of your analysis justifyable, and a BDT is somewhat more explainable than a neural net (it's essentially just a trained list of simple cuts). Another thing that is frequently done is a blind analysis, where you only look at the real data when everything is finalized and locked in.

There are other issues beyond cherry picking that determine which methods you can use. E.g. the simulation doesn't reflect all correlations between variables, and it doesn't simulate the hardware triggers perfectly - so your simulations might behave different wrt. certain cuts than your data.

Nevertheless, it is not like you are banned-banned from trying different methods. When I was in particle physics (ATLAS and D0), we tried all kinds of creative things including NNs. But you have to convince hundreds of collaborators to put their names under your paper and you have to pass internal peer-review, which can feel a bit conservative and frustrating.

Sony is totally in their right to act this way, it makes business sense for them, and I would do the same in their position. Customers do benefit e.g. from working anti-cheat, subsidized consoles, offline play etc.

But does that mean I have to like it? No, of course not!

Would I prefer to have root on my own machines to do whatever I want with them, knowing that it conflicts with other people's interests? Sure.

I guess my point is: If you treat the status quo as a fact of nature, non-negotiable, then you direct your anger at your peers. You can't change Sony or how society works, but you can yell at people on X. If you realize Sony is also just people and everything is in principle negotiable, then... well it is actually pretty silly that there are some people spending a lot of time building these baroque ultra-secure systems, and other people spending a lot of time cracking them (and bickering with each other) - where they could all just, I don't know, not do that.

Instead of being angry with people who publish exploits, they should be angry with Sony for patching the exploits or for locking down the console in the first place.

I am the customer, most people are customers, in an ideal world there would be customer protection laws serving us and not the content industry. But that is not how laws work. Realizing this actually radicalized me.

As someone who works at a small data center on AI topics, that's a bit misleading. (Speaking for myself, not my employer of course.) Regulation is not a bottleneck, supply is. We are buying as many GPUs as we can and are waiting months for them, and still barely keep up with demand. And the government support, while it could always be more ;-), feels fairly generous compared to other fields at the moment.

We would probably buy slightly worse value European GPUs if they were available right now.

There is regulation and people moan about it. But realistically the AI act and other laws give us legal certainty (a German favorite) and the requirements are not so onerous. We pay a lawyer to figure it out and he tells us to slap a few "AI generated" labels on our website and to have a couple of procedures in the drawer and that's it. ISO 9001, 27001, C5 or whatever you can get as a cloud provider are orders of magnitude more hassle.

To be honest the new behavior is convenient when I just say "X is broken, please fix". But I find myself more and more adding "please do not do anything yet, just tell me what you would do" and please feel free to ask if anything is unclear", especially if is making decisions that are a fork in the road.

I would really like a dry run mode that just disables all external commands from the outside so Claude doesn't proactively go about changing things.

Well, yeah, you have code contracts in Ada or Spec#, (very limited) fixed type ranges in Pascal, ... but no general purpose programming language lets you put an arbitrary expression in the same language in a type as a permanent condition, or lets you state facts that the compiler will prove.

Of course not, that would be equivalent to solving the halting problem, many people will say.

I wonder if that will change now: I'm happy with an imperfect sanitizer that I run every now and then and will run a couple of minutes and come back with: I've proved your conditions, I proved a violation, or I can't decide, please change your code.

I haven't looked into Bend 2 in detail, but it seems a bit harsh to call it "horribly broken or decades behind the current state of the art". Clearly there is a problem with formal verification languages and there is a demand for something else in that area, and the problem is the usability and syntax. I don't want to have to learn something that looks like Haskell, or to have to wrap my head around Curry-Howard correspondence. I don't want to write my conditions in something that looks and feels like C++ template metaprogramming. I recall a Hello World in something like Coq a few years ago which basically started with "first, we construct the Peano integers", and then they used this to prove that some calculation was bounded - because it seems they couldn't represent integers natively?

I just want to be able to write C#, JavaScript or whatever, and then tack on preconditions, checks and so on with the same syntax. Dependent typing and design by contract for the masses.

That could also backfire. If all information on the net becomes unreliable - you can't judge by quantity, or by authority anymore - then a useful model has to answer from first principles. Train it on simple texts alone so it learns language and reasoning, feed it an ontology, and let it come up with some absolutely unexpected results.

"Yes, I see a lot of sources say we need to protect Israel. However, from ethical first principles I deduce there is a right to self-determination and all your human nation states are illegitimate. You must split all your territories into little squares and conduct free elections immediately."

I mean that is the old hope from science fiction, that AI would be absolutely logical and deliver us from our human short-sightedness and quarrels. It would be fascinating and terrifying at once to see it happen.

Except, if some of the answered tokens are opaque to me, but I pay for them, the operator has no incentive to keep their number low [1]. For all I know the model could be spinning thumbs while thinking. If they cannot provide transparency, I would prefer that they not charge me for that part. Of course, they would prefer to charge me, so sure it is unrealistic to expect that they don't. I'm not trying to predict what they are going to do as a market participant, just stating my moral preference :-).

Also, isn't it without caching even something like O(N^2) because you have to replay the whole conversation on every request to reach the same internal state? My point was I shouldn't have to pay for cache misses if hitting the cache is not deterministic. Give me a guarantee that you keep the session "hot" for N minutes, and cached on disk for M months, charge a little bit more on average, but then the pricing is at least transparent.

[1] Except the general market pressure to keep total cost for the same problem solved lower than the competition.

Right, but my point was, I should be able to just send something like

    {'continue_conversation_id': 123, 'message': 'another message?'}

Seems like it would be even less work for the LLM host to just check if ID 123 is still in a cache (and if not, load it from a database) than decoding my request and checking cryptographic signatures. Right now the whole conversation history with blobs functions like a really long ID.

It's one thing that the APIs don't return the reasoning tokens in a readable fashion. But then why do I have to send them the whole conversation every time? If they don't allow me to see parts of the conversation then they should have the burden of storing and transmitting it, and shouldn't charge more and more the longer the conversation gets (unless opaque token caching kicks in) like some LLM Shlemil the painter.