So far, AI is a money pit that isn't paying off | Hacker News Reader

So far, AI is a money pit that isn't paying off

gizmodo.com

39 pointsmmphosis59 comments

This type naysaying feels like clickbait. The current Tensor-fueled AI wave is easily on-par or exceeding in significance that of the Internet, personal computing, or the smartphone. AI is extremely hyped right now and very expensive, but the landscape is changing quickly. On top of a few immediately valuable applications: text generation, code completion, image generation, etc., there are a million people trying to figure out how to make use of "AI" capabilities that already exist. Even if the technology did not improve at all past GPT4, we would see a decade of innovation as it works through industries. A huge portion of any AI project is collecting and preparing data. Now that we know that datasets like this can be valuable, more effort will be put into collecting: business conversations of all types (customer service, consulting, meetings, etc.), diagrams, blueprints, photos and video from manufacturing, healthcare, and constructions. The value equation is also rapidly changing with dozens of non-Nvidia chip makers furiously trying to capture a piece of the pie, and economically motivated developers (compute time is expensive!) developing more efficient code.

Unfortunately for you, we all heard the same exact promises and grand claims from people pushing previous hype cycles, such as VR, AR, EVs, cryptocurrency, Web 3.0, etc. Some people are addicted to hype.

They hype is definitely similar, but the situation is entirely different. You don't have to take anyone's hype/story about AI on faith. The future state is uncertain, but the current trajectory is clear and verifiable.

In early 2020 I was doing a data science boot camp and one project was feeding in a bunch of scientific paper titles in a given area, then "predicting" a "next" future title, the idea was I could come up with ideas for future projects.

What we have now is exactly the same except with a MUCH bigger training set. I am using GPT4 everyday but still can't believe the nonsense hype around it. ALL of AI would be lucky to generate a few $million a year of actual income for people. At BEST these LLMs are great spell checkers and code template fetching assistants but thats as far as LLMs on their own will ever go.

Microsoft Word and Google Docs will have a built in "AI" spell checker that will dissapera into the background, and google search might get SLIGHTLY more relavent but thats it. No "AI revolution" is coming. I'd love to be wrong but looking at the details its hard to understand where people think this future revenue will come from.

The availability of the auto-text-firehose of LLMs will almost certainly make google search worse as it just gets overwhelmed with spam.

With all these models trained on a corpus of texts from the Internet[1], what's going to happen in 5 years when the internet is full of text generated after January 2022, the current cutoff data for ChatGPT? As an increasing proportion of training data is either pre-2002 or LLM-generated confabulations, when will it all collapse in on itself? Do these companies have any plans for how they are going to get accurate and factual input after their output dominates?

1. Corpora: https://commoncrawl.org/, https://www.english-corpora.org/

> the current trajectory is clear and verifiable.

2016, Geoffrey Hinton, "People should stop training radiologists now. It’s just completely obvious that within five years deep learning is going to do better than radiologists." https://www.youtube.com/watch?v=2HMPRXstSvQ

Good example of the "peak of inflated expectations" from the hype cycle for emerging tech. Whereas crypto and Web 3.0 haven't broadly taken off (and would bet never will), AI is more likely going through a short and shallow "trough of disillusionment" before long steady gains.

> short and shallow "trough of disillusionment"

How many troughs is that now for 'AI'? Three? Four?

Actual profit generation is the metric by which hype-fads should be judged.

VR, AR, EVs (not sure about this one), crypto-mining if you are not just early in the pyramid, etc, have not generated profit yet.

CRUD web apps have generated a LOT of profit.

You're funny. Exceeding in importance the Internet itself? That's an incredible amount of hype. Would you rather have chatGPT dialed by SMS in a world without Internet, or Internet without chatGPT? I know which one I'd choose

It's a fair critique and an interesting thought experiment. Imagine a sci-fi alternative world where a smartphone-like portable device contained a compressed version all the world's knowledge (maybe a few petabytes?) and super-human intelligence (maybe 10x), but the highest level of communications was land-line telephones. It's not a slam-dunk that the world with high-speed Internet would be the superior one. Fortunately for us, we [almost] have both.

It's a little hitchhikers guide

I was considering using that analogy(!), but realized that our near AI future might be something that's closer to a combination of the Hitchhiker's Guide AND Marvin (50,000x smarter than a human).

I can imagine a future in which people are not generally connected to the Internet, but rather to local GPU-powered bubbles that simulate it. Visiting the actual sites that the search engine summarizer references costs extra, but your subscription includes a wide array of websites ranked higher. Most people who are online visit the same few social media / streaming / news sites anyway.

It seems like the high-quality LLMs that drive the hype train are not profitable, and the low-quality LLMs that you can run on consumer hardware and doesn't cost millions to train are not yet good enough to make money on.

If there's been one constant in technology, it's that compute gets cheaper. Imagine how the profitability calculation changes when Nvidia A100-level computation costs 1/10th what it does now.

all new technology is overestimated in the short term, and underestimated in the long term.

I remember when voice recognition was the big thing. Think era of "dragon naturally speaking" and similar. It was hyped, but practically speaking it was tedious and made lots of mistakes.

And then over time it found niches like phone trees, and it gradually became better until it was accepted.

Now AI is super hyped, but cars are driving themselves, closed captions are automated, people are getting assistance with art and writing, and chatbots are livening things up.

They’re waves. Going up the front it’ll change the world and going down the back it’s doom and gloom. Everyone is shouting all those things all the time, it’s just the balance shifting that we call a wave.

But, as long every crest and trough are a little higher, that’s what we call progress, no matter who you’re listening to along the way.

AI feels like "internet" in the 90's, which caused a huge crash in 2002 before taking over the world.

The "internet" itself was growing wildly through the dot com crash, though. That was an investment cycle: more money in aggregate was flinging around to "make network stuff" than there was revenue to be had selling it in the short term. There was no doubt in anyone's mind that everyone wanted and was rapidly adopting a global network and creating new markets.

AI right now is way behind that spot on the adoption curve. We have a lot of very impressive stunts that don't yet seem like "products" or "markets". And we have a lot of ideas awaiting implementation. But it needs to cross that bridge.

I think it's reasonable to assume there's at least one bust/boom cycle to come in the AI world before we get to a steady state.

Yep. The money crashed. But the technology never stopped

This seems like a silly point to make. It’s still an emerging technology, and will ultimately be the end-all-be-all technology if humanity can solve AGI. Even if we don’t any time soon specialized models will certainly provide massive productivity gains.

I think people are being unreasonably pessimistic about technology that is almost sci-fi-esque in its current capabilities, let alone its potential.

The end-all-be-all technology, until the power and internet goes out...

To be honest, if the power and internet go out, not much would be left, including meat-based technologies such as elbow grease. Can't power the amount of bloodsacks we have with enough biofuels once the electricity goes down and supply chains are disrupted.

Massive scaling down would be required, and we would need to "liquidate" a lot of bloodsacks if the loss of power and internet is permanent.

That's a truly bizarre claim, since plows and shovels work just fine without power or electricity...

At any rate, everything else can survive the lack of power and connectivity for a few hours or a few days. AI is only usable when the customer, and the AI provider, have both.

A link to gizmodo is the internet equivalent of those direct mailers that go directly to the recycle bin.

For the original wsj article: https://archive.ph/jMM1k

To take issue with the lede,

it's completely clear that the bet will pay off, just not for whom, when.

There is zero ambiguity about that fact that everything is going to change, for a pretty sizable interpretation of "everything."

That this hasn't been reflected in the quarterly numbers betrays either a droll sardonicism in service of generating clicks, or, a comical loss of perspective.

> it's completely clear that the bet will pay off

I don't think that's completely clear. I mean, it probably will in some form, but it doesn't appear certain. That's why it's a "bet".

"So Far, X Is a Money Pit That Isn't Paying Off"

This can be said about most new technologies at some point in their development cycles.

They said the same thing about the Internet is 2001.

Industry is still reacting. Hollywood in particular will be completely upended. There are those in entertainment that haven't grasped the impact of what's available.

The Chinese graphic design industry changed overnight and made millions of their designers redundant. In short: it's coming.

And yet GitHub's copilot just crossed $100M ARR https://twitter.com/swyx/status/1711792178031460618

But is it saving money on the user side. I’m less certain there’s real value in daily development. It’s something people are trialing still. Not sure who all will continue being subscribed in a year.

Have you used it? I’m never ever canceling my GitHub Copilot subscription. It’s revolutionary.

I used it for a while, but didn't find that it was beneficial enough to incorporate into my daily work. Perhaps I'm using it wrong, though.

It's best used for boilerplate, utility functions that are really simple but annoying to rewrite over and over. Or test harnesses, or things like that. Basically, things it's seen before, rather than application-specific stuff.

Basically: It can't think for you, but it can copy and paste for you. As long as you audit the output, which should be easy for this stuff.

Yeah, that's what I heard so that's one of the ways I tried using it. It's not compellingly useful in that role for me, though. I found it no quicker or easier than just writing (or copy/pasting from other code I wrote) it myself in the first place.

But I'll confess that I may be unusual -- I also don't find the old-fashioned code-completion stuff useful. That's for a different reason, though: it actively gets in my way.

I don't use it or find it useful either.

I dont find it useless but if it wasn’t for work giving if for free i wouldn't pay for it.

I’ve been using it a bit to learn rust. It often generates code that is correct or seem correct but is horribly inefficient and sometimes just wrong. I don’t feel I can trust it.

When I use it for programming languages I am experienced with, it often isn't giving super useful suggestions or just generates very ugly code.

The first I tried it, not impressed, but on some projects, it literally does 75% of the typing for me. I think it has to do with whether there is a similar open source project out there. The more unique your work, the less likely it is to find a good recommendation

Yes. I don’t get it. I don’t write tons of lines of code but when I do it has to be secure, readable, and well understood. Some of the answers from copilot are just wrong and others are less than useful.

Edit: I don’t write a lot of code, but I read a ton. I think that lines up with most mature code bases and senior staff.

> The problem is that users pay $10 a month subscription fee for Copilot but, according to a source interviewed by the Journal, Microsoft lost an average of $20 per user during the first few months of this year. Some users cost the company an average loss of over $80 per month, the source told the paper.

$90 of inference time a month is... a lot. Are people just generating code in vscode to cause calls to copilot or something?

Copilot sends huge prompts continuously since it includes adjacent files for context. That's why it's so good at "figuring out" your own esoteric code.

I'm not sure why they don't charge per usage.

Because presumably the goal of copilot is not to make substantial amounts of money at this time, but instead to produce tons of training data for the tools that come after (and are significantly more profitable, because they reduce the amount of money firms spend on software developers.)

Yeah, so far this looks like home assistants which are all money-losing. There isn't yet a killer app, like google ads/search, which is sticky and pervasive and where an effective monopoly emerges.

Home assistants have no per-month cost to the customer, so introducing one would be a deal breaker for many users. Copilot is already a paid service, so it's easier to increase the prices if needed. The ongoing compute costs will also drop in a Moore's Law-like way

FWIW Nat Friedman says this is just completely false.

https://twitter.com/natfriedman/status/1712140497127342404

Is that true for the newer versions they’ve launched, since he left GitHub?

this is not a strong counter argument. Costs will come down. When there is insanely high demand for a product (like there is here) and the thing makes people more productive, costs always come down due to pure incentives to make it cheaper. This happened with electricity, the car, air travel, solar power, etc etc

In the last decades I read that about:

- web ads

- web stores

- streaming websites

- chat apps

— Last-mile Delivery

— Blockchain

— Metaverse

— Scooters

— Anything strapped to your head

TBH, Metaverse still hasn't taken off yet. But Blockchain already has quite a place in cryptocurrency. Actual, useful applications... like buying drugs on the dark web.

Not satire, done it myself.

Love that I can grab Nvidia A100 and V100 GPUs on the Akash Network anytime - no wait! No KYC!

The environmental impacts of AI training are also abhorrent.

Well ya, it doesn't cost a billion dollars to teach a child to read.

In many ways, AI has been rolled out just exactly the opposite way that I might have done it. Just like I would have done almost everything the opposite way of what's been happening since the Dot Bomb around 2001.

A few examples: we should have had CPUs with 100+ cores and local memories standard by now instead of SIMD video cards and various abstraction layers more abstruse than OpenGL. We should have had free distributed general computation runtimes like SETI@home but for any language. We should have had rich media documents with progressive enhancement like htmx instead of single-page applications and walled gardens that can't even be scraped or crawled by search engines. We should have had an open data interchange format like SOAP/JSONAPI/GraphQL/Open Graph/etc but not terrible, for supporting the semantic web. We should have had an approachable functional language like Erlang with the readability of Go and the vector handling of MATLAB, but we never got it, because the geeks who could make that stuff spent the last 20+ years as workaholics or struggling to survive on shoestring budgets in the open source community. We should have pushed for decommodification and right to repair instead of draconian IP laws like the DMCA. We should have had open funding like prosper.com and kickerstarter.com that actually worked (for all projects, not just creative ones) instead of VC firms and billionaires. We should have had Pragmatism instead of Move Fast and Break Things. We should be automating work to liberate us from forced labor, not automating art to strip away our dreams. We should be working towards UBI for every single person on the planet, not starting proxy wars between empire and the people who do the actual working and living and dying to support the 1%. I could go on forever, and the eventualities of these life choices have grown to capture our entire collective psyche.

If we had all that stuff, I truly believe that AGI (or at least the liberation it represents) would have already arrived.

But in the direction it's headed now, AI will almost certainly adopt the form of corporations (which arguably already are AIs with inhuman/amoral agendas) and deliver us into a dystopian future where the choice is lock-in under the yoke or survival in the wasteland.

I still hold onto some small hope that AI can liberate us from the short-term scarcity mindset and bring about the New Age under some form of democratic socialism like on Star Trek. But it's so easy for the powers that be to sabotage all forward progress, even going as far as shutting down governments to get their way, that I worry it may already be too late. The safe bet is that we'll get the libertarian paradise of Star Wars instead. Present-day technologists will be tomorrow's rebels.

>We should have had an approachable functional language like Erlang with the readability of Go and the vector handling of MATLAB, but we never got it

No love for Julia?

I want to heart this <3 and there's also some interesting stuff in Clojure, F#, even Immutable.js with higher-order methods.

What I mainly object to is functional languages like Haskell that have steep learning curves but also impure functionality that puts the onus on users to not shoot themselves in the foot. I feel that they miss the point in some fundamental way that's hard to quantify. The simplest example of that might be that a spreadsheet is equivalent to functional programming, but one is readable and one is not. A bridge between them hasn't been successfully built yet, but I'd sure love to try.

Reply on news.ycombinator.com