AI Investments Are Booming, but Venture-Firm Profits Are at a Historic Low (msn.com) 31
Silicon Valley's venture-capital firms are having an easy time finding promising startups to back. The hard part is cashing out. From a report: Last year, U.S. venture firms returned $26 billion worth of shares back to their investors, the lowest amount since 2011, according to the data provider PitchBook. Startup investors say 2024 has continued the trend, with high levels of investment and few acquisition deals or initial public offerings. "We've raised a lot of money, and we've given very little back," Thomas Laffont, co-founder of investment firm Coatue Management, said at a recent conference. "We are bleeding cash as an industry."
Last year, U.S. venture firms invested $60 billion more than they collected, the highest such deficit in PitchBook's 26 years of data. As a result, the investors that back VC firms, such as university endowments and pension funds, aren't seeing the type of profits the industry has long delivered. The decline is particularly notable because the past three years have been the highest three on record for total VC firm investments since 1998 -- as far as back as the PitchBook data goes. Much of that money has recently gone to artificial-intelligence startups -- a white-hot space in which valuations are rising fast and companies quickly burn through cash to develop new technology.
Last year, U.S. venture firms invested $60 billion more than they collected, the highest such deficit in PitchBook's 26 years of data. As a result, the investors that back VC firms, such as university endowments and pension funds, aren't seeing the type of profits the industry has long delivered. The decline is particularly notable because the past three years have been the highest three on record for total VC firm investments since 1998 -- as far as back as the PitchBook data goes. Much of that money has recently gone to artificial-intelligence startups -- a white-hot space in which valuations are rising fast and companies quickly burn through cash to develop new technology.
Re:Bubble AI (Score:5, Interesting)
Waiting for it to burst. IronNet AI scam, DarkTrace. Let's see some more VC billions going to the void.
Perhaps so. But tell me, how DO you define a valid R&D period for AI? Would that not also require a metric fuckton of VC investment?
Would it not look exactly like we have today? If not, why?
Re:Bubble AI (Score:4, Interesting)
There is R&D and there is a "...but using LLM", just like with the first wave of eCommerce when it was "its a pet store but on the web"
We are in the phase right now where there is a lot of "AI" being used to do everything from recommend movies to setting your thermostat. It costs money to build and implement stuff even if its bolting an off the shelf LLM product to a database for RAG and gluing some web apis onto it. It then costs money for the compute resources to run it. You can and people are doing quite a bit of this, without evidence the results are any 'better' then simply counting up some keywords and doing basic statistics to make picks. -Oh but it can spit out a nice paragraph about why you'll love 'Martian Madmen II - Revenge of the Red Planet' based on your prior viewing of 'Martian Madmen - Just Add Water' while avoiding asking if that is any real value to the consumer or anything they'll pay for at least.
That is a different proposition than I want to hire a bunch of math wiz kids and some computer engineers to see if we can integrate this improved of matrix multiplication algorithm into silicon so I can make COTS AI tools to sell to people that want to do stuff like make terrible movie recommendations based how similar the script is.
Re: (Score:2)
Token processing / clock cycles needed.
Right now, LLM processing is so expensive that I have my doubts about ROI on using it. Worse than that, the stubborn hallucination problem means even if you can afford the compute time, you only have about a 60% chance of the answer being correct.
This isn't your grandpa's database.
It's going to take at least 5 to 10 years to sort out those issues. Until then, VCs will be losing money in this sector.
Re: (Score:1)
There is reason to believe that hallucinations cannot be "sorted out".
Re: (Score:2)
A more binary definition of truth would be most helpful. One of the most useful AIs I know is a religious AI, taken from a philosophy that has a very well defined definition of truth, and that gives you results that include bibliographies to research.
And I'd rate it as being politically biased and only maybe 65% accurate, with a huge Italian Argentinian bias.
Re: Bubble AI (Score:1)
Take this statement of what an LLM is:
"Given a sequence of words, what word is likely to come next?"
Hallucinations are an inherent feature of LLMs, expressly because next token prediction is what they do.
There is a whole path, using NP as a lens that can provide insight of this if not being exact. As choosing the next word can be thought as a decision problem.
NP is equal to SO-E, the second-order queries where the second-order quantifiers are only existantials.
coNP is equal to SO-A, the second-order queries
Re: (Score:2)
> you only have about a 60% chance of the answer being correct
Serious question, Is that better than a Magic 8 Ball?
"At the end of the day, there is no definitive answer as to whether or not the answers generated by a Magic 8 Ball are accurate or reliable. It all depends on how much stock you personally put into them and what type of questions you choose to ask. While they may spark interesting conversations and can provide some insight into difficult situations, they should never be relied upon as being
Re: (Score:2)
It's going to take at least 5 to 10 years to sort out those issues
That's optimistic. so-called 'hallucinations' aren't mistakes, they're exactly the kind of output we should expect, given how these sorts of models work. Larger models and incremental improvements are never going 'sort out those issues'. For that, we'll need something radically different.
Re: (Score:2)
I'm not so sure. It seems to me that by requiring the Gen AI to produce a bibliography of its answer, another AI scanning the bibliography could come up with a truth vs fiction score.
But first we have to label all the training data with source URL and "Fact/Fiction" bits.
Re: (Score:2)
another AI scanning the bibliography could come up with a truth vs fiction score.
Sure, but it won't be any more accurate than the usual nonsense it generates.
But first we have to label all the training data with source URL and "Fact/Fiction" bits.
That's not how this kind of model works.
Re: Venture is broken (Score:3, Interesting)
Re: Venture is broken (Score:1)
Re: Venture is broken (Score:2)
Re:Venture is broken (Score:4, Insightful)
Venture capital also thrived when there was basically infinite free money from the fed and any firm that promised eventual massive profits on a vague basis got shit-tons of investment. Investment that was unjustifiable given the underlying sustainability of their business model. But since venture capital got in first, they could come out ahead purely on capital gains.
LLM as AI (Score:1)
LLM as AI - is going to be dot-commedy 2.0.
Is there a lot of useful technology there sure, and will it likely be present and change the way we do things for decades to come, I bet so.
However I think it is also clear at this point that bolting an LLM to something and calling it new does not make it so. I also fully expect when this shakes out the ration of Pets.com(s) to Amazon.com(s) will be worse than last time.
Promising? (Score:5, Insightful)
> Silicon Valley's venture-capital firms are having an easy time finding promising startups to back
For different values of $promising.
There are so many companies trying to make it big, and seemingly so many rich people looking to lose their money. AI is almost the best ever sales pitch you could hope for - far better than crypto, or big data, or any of the others. AI can be made to offer "limitless" possibilities, if only someone had the nerve to just keep on investing, keep on scaling and keep on trying - honestly, utopia is any day now...
No surprise. (Score:2)
Compute and now some resemblance of intelligence has been commoditized. Beyond this point there is very little value-add with experts building services, since total n00bs can automate just about anything they want with a few clicks and prompts. The digital space has fully entered a post-scarcity economy within the last year. I'm curious when that will happen with other jobs and services and what the results will be.
I fundamentally don't even get why F@ceb00k, 1nst@gr@m, Sl@ck and others even exist as a busi
poor VCs (Score:4, Funny)
Those poor VCs. Someone should start a GoFundMe for them.
This is exactly how supply and demand works (Score:2)
When something is in high demand, buyers pay more for it. Paying more for things, tends to reduce profits.
Add to that, the product these companies are selling isn't in equally great demand (because the technology isn't yet proven for most applications), and you get lower prices when selling. Also contributes to lower profits.
Buy low, sell high, right? These VC firms are buying high, selling low.
That's called a pyramid scheme. (Score:1)
Not surprising (Score:1)
That's what happens when you skip doing the hard work of finding solid things to invest in and just chase the shiny.