Hacker News Reader

Comment by praveen9920 | Hacker News Reader

> in an age of low-size disk drives and expensive software, looking up data over a dedicated protocol seems like a nifty2 idea. Then disk size exploded, databases became cheap, and search engines made it easy to look up words.

I love this particular part of history about How protocols and applications got build based on restrictions and got evolved after improvements. Similar examples exists everywhere in computer history. Projecting the same with LLMs, we will have AIs running locally on mobile devices or perhaps AIs replacing OS of mobile devices and router protocols and servers.

In future HN people looking at the code and feeling nostalgic about writing code

View on HN

I've been working with other programmers for over 30 years, and the impact of this scarcity mindset has very real implications in behavior today.

In much the same way that someone who grew up with food insecurity views food now, even if the food is now plentiful.

For example, memory and disk space were expensive. So every database field, every variable, was scrutinized for size. Do you need 30 chars for a name? Would 25 do?

In C especially all strings are malloced at run time, predefined strings with max length are supported but not common.

Arguments (today) about the size of the primary key field and so on. Endless angst about "bloat".

I understand that there are cases where size matters. But we've gone from a world where it mattered to everything to a world where it matters in a few edge cases.

Given that all these optimizations come with their own problems it can be hard to break old habits.

I wish some of those old timers who care about bloat could work their magic on the web development ecosystem

I think they just have the good sense to run away, because even if they fixed things, they would be hated and mocked relentlessly for it. See for example every contemporary blog rediscovering “wow, makefiles just work, and this tech is 50 years old” and the number of contrarian reactions to it in discussion threads. For every dev that values ubiquitous, simple, predictable, and stable solutions.. there’s 3 who want to use something that’s much more complex, and has a shelf life of just a few years before it’s replaced with the new hotness.

There's only so much one can do. Convincing other people that these things matter is hard. Even when you have good evidence (which isn't always easy to come by), people often won't put in the work to keep things small.

It is, until it isn't. This stuff very much matters for games.

There's also a general argument about resource usage, but I think the AI and crypto people have largely won the argument that it's OK to use as much electricity as you want as long as you're making money somehow.

Gaming and Porn are the main reason for lot of advancements in tech actually. They are always on cutting edge and always need more resources

On the basis of GANs or diffusion models turning default skeletons into whatever style you want including photorealism, and that those skeletons can be animated by other AI, and that plots in most* games and porn are worse than the extremely basic stuff the original ChatGPT could provide, I think that's nearly at an end.

When they run in 240 fps on your phone, that's the end. I've seen 5 fps on unknown hardware, so give it a decade at most.

* Exceptions exist, but there aren't enough hours in a year to support even just 23 games like Bauldur's Gate 3 even if you're unemployed, and there's more coming out each year than that.

I think because of levels of abstraction of software/pltforms, top layers are bound to have « bloat ». Unless someone( something) changes radically across the stack, I think bloat will just increase for each layer

But on the other hand, for some applications, disk requirements exploded as well and require dedicated protocols and servers for it; for example Google's monorepo, or the latest Flight Simulator, the 2024 version will take up about 23 GB as the base install and stream everything else - including the world, planes, landmarks, live real-world ship and plane locations, etc - off the internet. Because the whole world just won't fit on any regular hard drive.

>Because the whole world just won't fit on any regular hard drive

Except it did. FS2020 had a base level of terrain quality that was installed such that you could play it even if you never connected to the internet. It wasn't Bing Earth quality sure, but it was way better than what you got in FSX thirteen years earlier. It was 200gb.

Which is conveniently about the same size as a single copy of whatever Call of Duty game is currently in vogue.

> Projecting the same with LLMs, we will have AIs running locally on mobile devices

That’s not much of a projection. That’s been announced for months as coming to iPhones. Sure, they’re not the biggest models, but no one doubts more will be possible.

> or perhaps AIs replacing OS of mobile devices and router protocols and servers.

Holy shit, please no. There’s no sane reason for that to happen. Why would you replace stable nimble systems which depend on being predictable and low power with a system that’s statistical and consumes tons of resources? That’s unfettered hype-chasing. Let’s please not turn our brains off just yet.

What is an AI? Just an LLM? What do you call Siri, Cortana, and Google Assistant? I've already received a Gemini app on Android, and they're promoting the premium version too. Runs locally, yes?

30 years ago, my supervisor wrote, from scratch, an "AI" running on the company web server (HP/UX PA-RISC; 32-64MB RAM), that would heuristically detect and block suspected credit-card fraud. Remember that "AI" is a perennial buzzword with fluid definitions, both an achievable goal right now, and a holy grail. ¡Viva Eliza!

> I've already received a Gemini app on Android, and they're promoting the premium version too. Runs locally, yes?

Are you asking me? I don’t know nor do I care, I don’t use Android or Gemini in any capacity.

> Remember that "AI" is a perennial buzzword with fluid definitions

You don’t need to tell me that. I didn’t use the term “AI”, I just quoted the other post and responded in what I understood to be their terms. I don’t think LLMs are intelligent, thus not AI. You’re nitpicking the wrong person.

You are going with assumption that llms will remain same forever. There are attempts everywhere to make them smaller, efficient and more focused.

Would you have called « transferring mails online « in the 90s is hype-chasing because our postal system was working great? Probably not a great analogy but you get the point

> You are going with assumption that llms will remain same forever.

I definitely am not and that’s stated directly in my post. I specifically said “but no one doubts more will be possible”.

> There are attempts everywhere to make them smaller, efficient and more focused.

That doesn’t matter when the issue is inherit. An LLM, by definition, needs large amounts of data and acts on them probabilistically. If you change that, it’s no longer an LLM. Any system that requires predictability needs to be programmed with rules we can understand, test, and reproduce reliably. LLMs ain’t it, and won’t ever be. Something else by a different name, maybe.

> Would you have called « transferring mails online « in the 90s is hype-chasing because our postal system was working great? Probably not a great analogy but you get the point

I understand analogies are never perfect, but that’s a particularly bad one. At least stay within the same realm of physicality. You made a bad comparison then ascribed a bad argument to be against it. That’s a straw man. I don’t even believe our current digital system is “working great”, so the analogy fails on multiple levels.

There are bad programmers and bad technology everywhere, the world is barely held by proverbial spit and bubblegum. Yet that doesn’t mean any crap that’s invented afterwards, be it “web3” or LLMs are immediately the solution.

You are missing the whole point of comment. Its a prediction/ speculation. I agree that it may seem bad with current state of things but no one can say with certainty it WILL be bad. There might be hybrid models which can do both probabilistic AND heuristic work within the scope, you never know. That’s my whole point.

If you see my comment, I said AI, not LLMs.

Imagine if dict://internet was renamed to agent://source, then agentic calls to model sources could interconnect with ease. With HTTP/3, one stream could be the streaming browser and the other streams could be multi-agent sessions.

I was thinking on the same lines but with improvements towards running local models is right around the corner. Unless someone has specific usecase or proprietary models, this may not make sense. But in current situation, that would be a great thing to standardise communication between various llms and clients

Given that most current AI generated code is dogshit, I would say we are well off from that.

Not right now but we are seeing the accelerated trends of generated code. How long would it take for regular use cases to be completely AI generated on the go. Edge cases exists everywhere.

It seems grossly inefficient for coding to go from NL prompt -> LLM -> formal language -> compile/execute. I'd envision AIs that would construct apps directly from the prompts, without the fragile middleware.

In fact, if it is optimal coding to interface existing libraries and frameworks, with minimal novel code, then just go full LEGO and pull in those dependencies, minimizing the error-prone originality.

Reply on news.ycombinator.com