[sajithdilshan]

This is quite a short sighted analysis. I do think the valuations are quite and they would need to meet the reality, but don’t think there’s gonna be a crash or we’d ever go back to pre-AI era. It would more or less would be a correction to valuations.

The future of AI would be on-device models which are as powerful as current frontier models and also I can imagine companies have their own deployments of inference of open weighted models for most of the use cases and use the frontier models for extremely niche or higher intelligence tasks.

As an example I use Claude code heavily for every day development and Opus 4.8 was already good enough for my use cases and never used Fable. Also note that I use AI as a tool to help with my work and I do not offload everything I have to do to AI in a single prompt

[skippyfish]

> I do think the valuations are quite and they would need to meet the reality, but don’t think there’s gonna be a crash or we’d ever go back to pre-AI era. It would more or less would be a correction to valuations.

Something to consider: would your description also apply to the dot-com boom of the late 1990s? The internet was real, the ideas for internet business were real, and we were not going to the previous reality. But the valuations weren't quite right and a "correction" happened at some point.

When people talk about AI crash, that's what they mean. Not that AI is a hoax, but that the correction could be quite violent and have effects on the broader economy.

[sajithdilshan]

But there’s a difference between fearing mongering with a crash and a long term correction. The correction would make all the startups/companies that created a wrapper around frontier models go away, but the ones that truly solved a problem would remain.

Also I remember reading that Anthropic is on its way to be profitable in 2028

[skippyfish]

With respect, the way you talk about corrections makes me suspect you weren't working in tech in the US at the turn of the century.

A "correction" in this model isn't a reasoned and gradual re-evaluation of the market cap of every company. It's a broad pullback where people panic and no one wants to be left holding the bag. Money shifts into other assets for years and tech employment, incomes, and the availability of funding takes a big hit.

As to your comment about profitability... every unprofitable company is on a path to be profitable. Some even get there.

Plus, look at it this way: Kellogg's is a profitable company and a part of almost every person's life. Does this make them worth trillions of dollars? No, they just provide boring, commodity products, their valuation is basically a low multiple of the assets they hold and the revenues they bring. There's a future where OpenAI or Anthropic are more powerful than all the world's governments combined, but also a future where they're Kellog's.

[satvikpendem]

That's exactly what a crash is, there's not much difference. For example that's exactly what happened in the dot com era.

[dgellow]

> but don’t think there’s gonna be a crash […]. It would more or less would be a correction to valuations. [...] The future of AI would be on-device models which are as powerful as current frontier models […]

That’s the crash… that’s pretty much exactly Ed Zitron’s thesis

[sajithdilshan]

On device models would work for most of the use cases we have, but we would have new use cases in the future (e.g. with AR/VR glasses, humanoid robots) that would need more advanced models that cannot run on device and frontier models can provide that

[mncharity]

> [local, but for] extremely niche or higher intelligence tasks

Access to a centralized copy of the web/literature/media, scraped and indexed/integrated, seems another non-local center of gravity. Versus a local model's last minute reaching out to "manually" search and browse.

Also mass parallelism for large ensembles. Perhaps unless/until we get those local models as 10k+ tok/s chips. Local can follow frontier because frontier is still sort of "expensive rack local". If STOA becomes massive burst-parallel ensembles, that following may get harder.

[cortesoft]

> As an example I use Claude code heavily for every day development and Opus 4.8 was already good enough for my use cases and never used Fable.

I think the flaw in this logic is thinking about how AI is currently used only. Yes, Opus is good enough for the task you are asking it to do, but that doesn't mean that is all you will ever need.

As AI gets better and better, it will open up new use cases that require the better performance.

[Silamoth]

You don’t “need” any LLMs to do software development. Plenty of us were productively writing code for years before these models were released, and many of us continue to do so. The “AI” companies want to sell you hype. But you don’t “need” them - not if you’re a legit, professional programmer.

[dan_gee]

Thank you for your comment.

The narrative that LLMs are essential to the future of the trade often feels like an assault on my professional expertise.

Whatever your experience is, mine is that LLMs don't work well for software development. I wish people who use AI would be more willing take that perspective seriously.

[satvikpendem]

You should read this thread and the linked article for a discussion about your last sentence: https://news.ycombinator.com/item?id=49052023

[dan_gee]

Could you leave a comment expressing your perspective?

[satvikpendem]

I did actually reply to a few of the threads there. I think you'd find more perspective in reading the article and the author is also there talking about it.

[satvikpendem]

You could say this for any new technology though, it's a vacuous statement. No one needs a car, they can ride a horse. But step changes in technology do occur and raise the standard for everyone in a society, hence why we don't see horses for transportation anymore.

[sajithdilshan]

You don’t need it, but using it would make you more productive. As an example, I have worked in a code base which is pretty much opinionated. I have many Claude skills to build the boiler plate starting code and tests for the feature I have to build following the existing conventions and patterns in the code base and that saves a lot of manual coding time for me.

This is a very personalised use case, but I think everyone can find these kind of use cases that saves a lot of time for you.

[what]

> to build the boiler plate starting code

Rails had that 20 years ago, no LLM needed.

[cortesoft]

I have been programming for 35 years and a professional programmer for 25 years. Of course I don't NEED AI to program, because I did it before.

This is completely unrelated to the how useful AI is for programming or how much more efficient you can be with it.

I never said you would need AI to do things you can do today. When I said it will not be 'all you ever need', my point is that NEW uses will come that will need more powerful AI. By definition, you can't NEED a new technology to do something that is already being done, because the fact that it is being done already proves you can do it without the new tech. However, that doesn't mean the tech can't do something new that does need the tech.

For example, no one needed an airplane before they were invented. However, you do need an airplane if you want to get somewhere 6000 miles away in less than a day.

There hasn't been a single bit of technology that is 'needed' if you use your strict definition of the word, because obviously humans existed and survived before the technology existed. Being absolutely necessary is not what makes a technology persist or spread, that is an artificial bar to reach.

Some people have such a natural aversion to LLMs that they will make incoherent arguments as to why they should go away. There are plenty of legitimate concerns and critiques of LLMs, you don't need to articulate arguments that you would never make about any other piece of technology to argue against their usage.

[sajithdilshan]

you sounds so angry, like an old man throwing a temper tantrum. Nothing you said makes sense

[cortesoft]

I definitely did not have any anger when I typed that, although I do admit I have a lot of confusion now. I have no idea how my message could come across as angry, even re-reading it now a day later. I also don't think you seem like some sort of anti-AI zealot, so it confuses me even more that you would say nothing I said makes any sense.

If you have interest, I would love to understand why what I said didn't make any sense, or why you think I sound angry.

[sroussey]

The money being poured into AI infrastructure means there is a market for new ways of doing things that take 1/1000 of the power or are 1000x faster, or both.

And there are many such moonshot startups.

AI on GPUs is an efficient as gaming on CPUs.

All that math where perfect precision is not required means that you can’t tell do things in different ways.

[d4ng]

Why is on-device AI the future? What is your reasoning behind this? Look at the proportion of things we compute on someone else’s computer relative to what we compute on our own device. Why would this change for LLMs?

[adolph]

I think the reasoning is similar to why Amazon notification emails are very generic: user data is used for various purposes that a growing number of users and 3rd parties have concerns about.

In addition to on-device, Apple is making efforts to secure computations that need to occur off-device, see: https://security.apple.com/documentation/private-cloud-compu...

[gorgolo]

I don’t think users (as in the general population) have a good track record of being judicious with how they share data. In the past with social media etc it was consistently the opposite, if the price is lower or free people tick any box and hand of all their personal data. Might turn out the same here too no?

Not arguing against on-device LLMs, I also wish it were so.

[DANmode]

The validity and permanence of a technology has literally nothing to do with how irresponsible people have been while placing speculative bets on it.

In this case, it’s really irresponsible.

[s0ss]

How's it "quite short sighted"? Are you saying the math _does_ make sense? if so, how?

[orangecat]

Back of a very tiny envelope: suppose there are 100 million jobs that benefit from spending $10k/year in tokens. That's a trillion in annual revenue; at a 30% margin that's $300 billion in profits, which can easily support $6 trillion in valuation.

[s0ss]

numbers are fun.

[KaiserPro]

Thats great but this is like when 3g came out. Sure it was the future, but it was half baked, impersonal, expensive, unreliable and required a culture shift to be adopted.

Onboard decent LLM performance thats _power efficient_ is at least two/three hardware generations away. (assuming linear performance)

but, the valuations, with debt trade and private credit obscuring exposure is a recipe for disaster.

[AussieWog93]

Seriously, you should try Fable. It picks up on subtleties that Opus misses.

[sajithdilshan]

I have heard good things about it, but I truly don’t have a use case for it to justify the inference cost. It’s double the price of Opus