Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AI's $344B 'language model' bet looks fragile

bloomberg.com

Post reply on HN
Print view View original

101–110 of 121 posts

Re: AI's $344B 'language model' bet looks fragile

#101

Post by simianwords »

pmdr wrote: post #94

One can call into question the paucity of AI-critical posts & comments here on HN. Not much is being said about the economics, everybody's living off the hope that "they'll figure it out." And AGI is just a few hundred-billion-dollar loans away, so why quit while we're ahead? If they do "figure it out" (both AGI and a viable business model), a lot of people here will likely be out of a job. If they don't, the whole t…

The answer is so easy to solve.

Do I find value in paying 20 dollars a month for ChatGPT? Yes. Do others? As far as I can see, yes. Most people are happy with the value.

Are AI companies profitable if they stop R&D? Yes.

Where’s the skepticism coming from?

Re: AI's $344B 'language model' bet looks fragile

#102

Post by Fricken »

fellowniusmonk wrote: post #93
Fricken wrote: post #81

Earlier quoted context omitted.

SAE Level 2 automated vehicles are useful now, just not so useful that it's safe to take one's eyes off the road and hands off the wheel. LLMs are in a similar place, and still a long way from doing the whole job reliably by themselves. The current state of AI is one or several unknown unknowns away from real AGI. We're missing something fundamental. Likewise with autonomous vehicles, the mythical SAE level 5 vehicle…

I see you were very precise with what you wrote about SAE 2. I ride in waymo all the time, they are SAE 4. I know many people who take a waymo every day. What percentage of driving represents the delta between 4 and 5? SAE 4 level LLM/Ai (if we can really even make that comparison) would have far less difficulty in deployment and would be far more disruptive in a far shorter period of time than SAE 4 self driving car…

An L4 LLM in my mind, would be one that can perform fully autonomously at some domain specific job, a job being a collection of tasks that have to come together coherently to achieve a desired goal.

Waymo's robotaxis, likewise, are able to do the driving task in geo-fenced areas. Waymo does a lot of hand-built code and testing to deal with particular problems, such as the 5-points in the Cairo district of SF, where it's a completely unique intersection, there's no other intersection quite like it. A ton of effort went into dealing with just that one intersection, and the bespoke effort doesn't generalize to any other intersection.

So if you want to be able to say, have a platform for producing working video games out of prompts, well, I believe that can be done with our current AI, but it will depend on a lot hard work making tools and hand-built code that do not generalize to other domain specific jobs.

Now if you want to make a movie worth watching out of prompts, that could be done too, but it depends on solving a whole different set of bespoke problems that once again need to be solved the hard way using more conventional software.

Re: AI's $344B 'language model' bet looks fragile

#103

Post by boplicity »

chain030 wrote: post #99
boplicity wrote: post #77

Earlier quoted context omitted.

So many things, once you start looking. However, most of the critics seem to focus on what it can't do currently, which seems to turn off their brains to the possibilities. Just look at what Cursor (and similar) have done in terms of the tooling for LLMs. There's still tons of progress to be made there, but similar tooling can happen across a variety of industries and categories. For example, I run a database of info…

"The low hanging fruit is extremely abundant, for those who are able and willing to find it." Ok fella. Its so abundant right. So why not go ahead, start your own firm and profit from this opportunity, that according to you exists? lol.

I'm using LLM tools to do quite a lot of things, and am very happy with the results in my business. So yeah, I am indeed running my own firm and profiting from this opportunity.

Re: AI's $344B 'language model' bet looks fragile

#104

Post by chain030 »

boplicity wrote: post #103
chain030 wrote: post #99

Earlier quoted context omitted.

"The low hanging fruit is extremely abundant, for those who are able and willing to find it." Ok fella. Its so abundant right. So why not go ahead, start your own firm and profit from this opportunity, that according to you exists? lol.

I'm using LLM tools to do quite a lot of things, and am very happy with the results in my business. So yeah, I am indeed running my own firm and profiting from this opportunity.

Link?

Re: AI's $344B 'language model' bet looks fragile

#105

Post by ankit219 »

there is clearly overhype. Given the influx of people building here (and i am not aware if it happened previously too), in a bid to differentiate from other startups building the same thing, many just stripped the nuance out of any technical idea, and made it a simple marketing term. As an exec where ten startups are promising you "training on your data" to provide the best chatbot, it's hard to tell the difference between who would be actually training and who would be just tweaking prompts. This has happened to other concepts too (anecdotally, most famous is how everyone is offering deep research). yes, it helps growth, but comes at the cost of trust. There is this startup which promised "experience based learning" when all they were doing is adding memory to the prompt to get it to perform better. (you can look it up, recenty raised series A).

This does not mean ideas are not working. I personally think pretraining has done its job. We did not know what the job previously was, but now we do given the way RL works. Pretraining and test time compute enables models to develop a generalized prior they can use to solve any given problem (much like how humans solve such problems). Sometimes priors are lacking so you need to train more using RLVR, and still early days, but directionally I think we have another scaling curve here.

Re: AI's $344B 'language model' bet looks fragile

#106

Post by rurp »

simianwords wrote: post #96
zippyman55 wrote: post #14

Earlier quoted context omitted.

There are no error bars, no confidence intervals. Just a one trick pony that pastes tokens together to give you something that may look good to many people. Sure there are many good use cases, but there are still enough non-patchable and indeterminable in size holes in the watering pail to limit its effectiveness.

Is there any error bar or confidence intervals in stackoverflow answers?

Absolutely. Votes, comments, competing answers, posted/edited dates, and the context of being on stack overflow all provide useful signals about how reliable a bit of information is.

Re: AI's $344B 'language model' bet looks fragile

#107

Post by boplicity »

rsynnott wrote: post #86
redwood wrote: post #4

This technology demos incredibly well and you can just see how everyone gets giddy with excitement around using it. I watch my colleagues and executives proud to show what they could do or make endless jokes about it. It reminds me of when people first got their phones and couldn't stop showing everyone how cool they were. This leads to an over rotation in the perceived value.. the value is significant just as the mo…

One thing I find fascinating; go to any forum/subreddit/whatever for any LLM thing, and it will be full of people complaining that it's not as good as it used to be, and that OpenAI/Anthropic/Google/whoever is intentionally degrading it, because they are so evil and want their products to be worse. Then a new model or tool comes out, all is wonderful for a bit, then repeat (except for GPT-5, oddly; that one seems to…

When GPT 5 came out, I was using GPT 4 mini for a regular automated task. This had been working quite well for some time. It stopped working right when GPT 5 came out. I switched to GPT 5 mini, and it started working exactly the same as it did previously. So yeah, I'm pretty sure they nuked GPT 4 when they launched GPT 5.

Re: AI's $344B 'language model' bet looks fragile

#108

Post by mvdtnz »

rsynnott wrote: post #86
redwood wrote: post #4

This technology demos incredibly well and you can just see how everyone gets giddy with excitement around using it. I watch my colleagues and executives proud to show what they could do or make endless jokes about it. It reminds me of when people first got their phones and couldn't stop showing everyone how cool they were. This leads to an over rotation in the perceived value.. the value is significant just as the mo…

One thing I find fascinating; go to any forum/subreddit/whatever for any LLM thing, and it will be full of people complaining that it's not as good as it used to be, and that OpenAI/Anthropic/Google/whoever is intentionally degrading it, because they are so evil and want their products to be worse. Then a new model or tool comes out, all is wonderful for a bit, then repeat (except for GPT-5, oddly; that one seems to…

I can't speak for all models but I can tell you for an absolute fact that Claude has degraded recently. Yes I've noticed it subjectively but Anthropic have also straight up come out and said "lol oops we 'accidentally' made it way worse, now that you've all noticed we'll roll that back".

https://www.reddit.com/r/ClaudeAI/comments/1nc4mem/update_on...

Re: AI's $344B 'language model' bet looks fragile

#109

Post by superblas »

34679 wrote: post #64
brookst wrote: post #25

Earlier quoted context omitted.

$100k/year is literally nothing. Think of it as maybe $10k/employee, figuring a conservative 10% boost in productivity against a lowball $100k/year fully burdened salary+benefits. For a company with 10,000 employees that’s $100m/year.

That's literally not how the word "literally" works.

That’s literally how the English language works. It literally evolves

Re: AI's $344B 'language model' bet looks fragile

#110

Post by pmdr »

simianwords wrote: post #101
pmdr wrote: post #94

One can call into question the paucity of AI-critical posts & comments here on HN. Not much is being said about the economics, everybody's living off the hope that "they'll figure it out." And AGI is just a few hundred-billion-dollar loans away, so why quit while we're ahead? If they do "figure it out" (both AGI and a viable business model), a lot of people here will likely be out of a job. If they don't, the whole t…

The answer is so easy to solve. Do I find value in paying 20 dollars a month for ChatGPT? Yes. Do others? As far as I can see, yes. Most people are happy with the value. Are AI companies profitable if they stop R&D? Yes. Where’s the skepticism coming from?

I'd actually be content knowing $20 a month covers inference and makes them a decent profit, but I have my doubts. If there's proof out there that it does, I haven't found it yet.

> Are AI companies profitable if they stop R&D? Yes.

By what metrics? Where's the data?

Post reply on HN