Weird time to do it IMO, right on the brink of agents potentially becoming the main entity closing purchases for many. If agents succeed at becoming e-commerce middlemen for the masses, I think these decisions will age poorly and there will be some walking back.

If you plan to stay in tech.. but it's a great time to make alternative plans if you aren't committed to adapting to the changes the industry is going through. The day-to-day of tech work could be unrecognizable in 5-10 years

I don't think it's a binary. Even among people who are pro-AI, there are uses of AI that almost everyone is against. I doubt most people have a problem with a lightweight model that runs on a laptop and does grammar checks, but models that are wildly energy-inefficient or used for war + crime + scams are a different story.

And I'm a fatalist regarding AI. I put aside whether I'm for or against. AI just is and we have to deal with it.

Below the fatalism though, I've been excited for this day ever since I saw "2001" in the matinee theater as a kid (and also as a kid: played against a Tic Tac Toe computer at the Museum of Science and Industry in Chicago).

And that I can run some of these models locally? Heady times for my inner kid.

I've tried (and struggled) to understand how they work as well so that I am not ignorant on the subject—so that I can have an opinion a little better than a layman's.

> I've been excited for this day ever since I saw "2001" in the matinee theater as a kid

You've been excited for the arrival of AI ever since you saw a movie about a murderous AI who tries (and mostly succeeds) to wipe out the entire human crew of its ship? Dr. Strangelove also got you psyched up for a nuclear war, I guess?

Ha ha, yeah, well, I understood the film to be fiction.

Yeah, to be against ai on some level is to be against computing. An abacus is ai, for some looser definitions of ai. But we can all accept our mortality while we give first priority, breath by breath, to staying alive.

I agree. I am extremely worried about the development of frontier AI models. I would like to see an immediate stop of further developments of such models and a permanent, global, ban on superintelligence.

On the contrary, I think that less capable models could be used for several, largely positive, applications.

I also think that narrow AI models developed for specific scientific applications (e.g., protein folding, weather predictions...) are in general beneficial for humanity.

sorry but are we really gonna be calling it superintelligence now ? haha

I totally agree with the benefits of the smaller models, but it's the large frontier models, potentially or accidently removing the security barriers from those can cause things far worse than the "hugging face" incident..

> we really gonna be calling it superintelligence now

I will not be, we engineered it, so it is an artificial type of intelligence, and given the dumb things they do with consistency, I find it quite the stretch to call it super or intelligent

I think the current generation more closely resembles a spaghetti throwing machine at the first point of difficulty, this is not how intelligent beings like ourselves approach these situations.

Very well said..

But why would they stop when they have a clear path up? They don't care about the people they just care about growing their wallets. I don't see them ever pressing the stop button aslong as results are there, also I do think that results will probably deteriorate, as I would like to believe transformers will hit a limit or rather a bottleneck, perhaps several bottlenecks. Where progress will massively slow down.

Regulate it globally

Of course they won't stop by themselves

Yeah but how would you realisticaly enforce this ? global regulation?

Let's pretend US and EU agreed to slow down development, what would stop China or other countries outside "west" from continuing?

I think that international coordination would be very difficult when AI is a national competition.

It's not something you can easily hide. You need the energy usage, GPUs, huge data centers

> too capable and too dangerous

I see negligence as I look past the fear mongering from our dysfunctional media ecosystem

That's fair. A lightweight self-hosted model can't be so harmful(yet).

Perhaps more about the huge frontier models being developed at openai or antrophic.

Or perhaps the controversial stuff being associated with Palantir.

People are starting to ask questions, I mean a lot are just worried, I had so many people tell me they were pro-ai a few years back, but now they just have that worried face whenever you mention it.

This continues to be an absolute no-go for competitive real-time games due to input latency. I honestly think there's a chance that game clients might go the way of the recording industry, and they just might not be worth that much.

As doubtful as I am about the viability of cloud streaming, this IS going to cause a bigger push into live service functionality where maybe you have a light client but the server is authoritative over more and more aspects of the game. Even then, AI can write its own server, so it needs to be full of FOMO if you're not on the official server.

I don't even know if a smaller company like Nintendo can pull something like that off. They have been behind so much in the tech space because they spend modestly and their primary differentiating factor is their IP. Their lawyers are going to be working non stop until they figure something out. I guess have a strategic partner?

I'm not sure I'd totally agree with this characterization. They've certainly been behind in tech that's like "stuff the most FPS and fanciest graphics into this small box as possible", but I'd argue they've been pretty much the only one playing in the "make a good experience for most users without configuration with relatively cheap local hardware" up until relatively recently (SteamOS certainly has made some inroads, but I've yet to see anything particularly impressive in terms of an Xbox handheld, and Sony has basically given up it since the Vita). I don't think either of these broad technical areas are particularly aligned with making cloud streaming viable for gaming, but if I had to pick one side, I'm not sure I'd pick the side that's invested way more into expensive hardware on the local side rather than figuring out how to make the constraints of lower power go further.

You make a very valid point but it really sounds like you are describing whether a company provides immense value to their customers whereas I am describing: are they up to date technically?

Nintendo is excellent at providing that value using low tech. They managed to make a compelling product using cardboard and sell it for 80$. They are the masters at providing good experiences with relatively cheap local hardware.

[1]: https://en.wikipedia.org/wiki/Nintendo_Labo

Cloud game streaming has gotten to be extremely impressive, at least on a wired connection, to the point where it’s almost hard to tell you’re streaming. It’s been hard for me to consider upgrading my old gaming PC considering how good GeForce Now is for the $20 a month that I spend.

I subscribe to XBox Game Pass and used to subscribe to Geforce Now too. They're both really great for me as more of a casual gamer, but at the competitive level in the fighting game genre, it's totally insane and people will notice a sub-100ms lag in their inputs because they're sort of pushing the limits of human reaction time.

> at least on a wired connection

There goes 95% of your market. If not more.

I didn’t say only on a wired connection. All of my experiences streaming wirelessly have also been great. I’m just not willing to make the same level of claim when I’ve probably spent less than ten hours streaming on WiFi and hundreds on a wired connection.

If I as a person who only plays on local hardware; went to streaming tomorrow, do you think I would notice the difference?

Or is it something you get used to like playing at 60hz, then going back to a PS2/etc at 30hz for awhile?

I can't tell if this is what they're talking about, but I frequently stream from my desktop to a SteamOS handheld on my home network, and on pretty much any game that I can max out on my desktop's specs (Radeon 6900 XT, so not exactly the newest GPU, but it's able to do quite a bit), I can get 120 FPS on the handheld without much noticeable input lag.

I don't have any reason to believe this would translate into a good cloud gaming experience though, because it mostly seems to have become viable from local wifi just getting a lot better in recent years. I would expect the latency to still be pretty high for a remote server, and it's possible that the streaming server I use is just able to "correct" things more easily for a local connection in a way that a fully remote connection might not be able to.

I’m talking about cloud streaming though I used to do local streaming through steam and remote streaming through parsec. GeForce now in my experience has worked better than either of those options running on my own hardware regardless of the hardware being local occasionally.

I did preface my comment with "I'm not sure if this is what they're talking about", because I didn't know if it was what you were talking about. I guess it's not.

I know that you did and didn’t mean anything by my comment if it came across a certain way. I just thought it would be worth clarifying and sharing my experience with having done local and self hosted streaming previously versus the cloud streaming I’ve been doing in place of it.

Fair enough, apologies for misunderstanding. It is a genuinely interesting data point!

As someone who used to upgrade their GPU every year to the newest top card I find it to be indistinguishable from local hardware 95%+ of the time. It’s fairly mind blowing how well it works. Also just to clarify again I pay for the ultimate or whatever the highest tier of GeForce now is for like $20/month and use it primarily on a wired connection so that’s where my experience comes from. With their service you still need to pay for and own the games (though it also connects to Xbox game pass) so you’re only paying for the streaming. It’s also not something that depends on developers allowing it to be on the service so not every game is on there but the catalogue is quite large. I play on a 3440x1440p ultrawide at 100hz and I believe the ultimate sub supports up to 4k@120hz

I think this is only true if you assume the classic server model where the remote rendering streaming is overhead and not colocated with the game server itself.

Surely no one playing competitive real-time games are doing it on a PS5?

Basically all fighting game competitions are done on consoles

Ahh .. yep. Completely forgot about that genre!

There's plenty of reason to still be skeptical about the future of VR, or Meta's practices, but shoving all of that aside it does look like a big step toward the level of comfort that's going to be needed for VR to succeed. Not at this price point, but it's incremental progress.

Yeah I use a custom fork of llama.cpp that has an abliteration feature that basically does this. It's sloppily vibe coded and I don't have time to coordinate on a way to do this cleanly upstream, but it's absolutely possible and saves a lot of time and bandwidth from being wasted

The HN audience is overwhelmingly aware of the issues around right to repair and data collection, so there isn't much reacting to do there - I assume there's already near-unanimous agreement that it's a good thing to educate people on, but we will have opinions on how to do it (or not do it) effectively.

Seems like a more organized way to do the equivalent of a folder full of md files + instructing the LLM to ls that folder and read the ones it needs

If so it would be most welcome since LLMs doesn't always consistently follow the folder full of MD files to the same depth and consistency.

what makes it more likely that claude would read these .md files then?

Skills is hopefully put through a deterministic process that is guaranteed to occur, instead of a non-deterministic one that can only ever be guaranteed to happen most of the time (the way it is now).

It is literally just injecting context into the prompt.

It includes both the file names and a configurable description string. That’s where you put the TLDR of when to use each skill.

This improves it a great deal but at a certain point, maybe 60-80% of the way it can start fading.

trained to