Commodification of Intelligence: Good, Bad, and Ugly Circular AI Deals | Hacker News Reader

Commodification of Intelligence: Good, Bad, and Ugly Circular AI Deals

emergingtrajectories.com

88 pointscl4248 comments

Has a competent economist modeled the circularity of these deals ? [ numerically or analytically ]

It seems a healthy economy has a lot of wide circularity .. money circulating is a good thing, a result of a functioning market, tracking the flow of real goods / services. But large corps circulating paper 'self-deals' or debt-swaps seems like a bad thing - a creative accounting practice designed to pump up their stock price/valuation.

How can we _quantify_ the difference ? I guess it would need to match the cash / debt flows against the movement of actual goods and services ??

Not an economist, feel free to weigh in, suggest links.

you'd need to look at prior work associated with the Dot com boom and Enron. Economists rarely look into today's actual activity because America makes it nigh impossible to actually understand who owns what around the fringes; contracts arn't required to be publically disclosed and a bunch of other dark interests make it impossible to really know.

MLMs exist in the same murky waters and tread the same ephermal economics by pushing useless product but hiring people to hire people to hire people, etc.

> prior work associated with the Dot com boom and Enron

The Enron deals weren't circular. Fastow signed both sides of the contracts. The deals were straight up fraud.

many of the Nvidia deals involve equity in the company instead of cash for the gpu orders.

thats basically one step removed from signing the deal themselves.

Also these deals are not circular - NVIDIA has to pay a lot of money for chips and foundry capacity just as an example, so money leaves the circle and participants have to raise capital to replace it, which means if this were just a financialy trick, it would be very expensive to maintain.

Many people are calling them circular, eg bloomberg, BIS :

https://www.youtube.com/watch?v=lwcM0By4aJQ

https://www.zerohedge.com/markets/bursting-ai-bubble-collaps...

diagram : https://cms.zerohedge.com/s3/files/inline-images/AI%20ecosys...

My point is, Id like to know how circular or not they are - clearly a lot of money is actually being spent on GPUs/datacenter build/electricity, Id like to know what fraction that is.

Two contradictory thoughts based on the article:

1. One thing you could do is look at free cashflow. The actual cash a company generates and keeps on its balance sheet would be a clear indicator of retaining profit or resources after circular deals.

2. The above doesn't preclude a company from committing to backstops and offtake agreements, so you still need to track those.

... so you'd want to look at how much cash a company is retaining and whether it is committing to act as a sort of lender of last resort.

Up until this week, I'd argue that Nvidia is was a pretty solid player in this space. If it's generating $10s of billions in cash, then who cares if it put $10 billion into a risky startup? It can afford it.

Now, however, it is also exploring committing $250 billion as a backstop for OpenAI's data center projects, which moves it to the riskier side of things.

I know I'm not providing you with data; I don't have that... But I think the above is a signpost to watch out for.

It’s a catchy headline but I also am not convinced they ultimately are. Certainly some of risky bets in the lot but most of what has been reported on is deals with ultimately cash. Compared to say the dot com bubble where lots of financing was happening.

I think there's not much to model. Either AI actually becomes a hot commodity, or the people invested in these deals are going to take a bath.

The leverage is the trick in the whole equation. During dotcom days the leverage was fine until the cash flow ran out. Then bankruptcies started and the bottom fell out. It appears this is happening again but perhaps the business model will just evolve. Although right now it feels more of the same.

Recently: "Nvidia, CoreWeave, and Nebius: Inside the Circular Financing of the GPU Boom" (io-fund.com)

https://news.ycombinator.com/item?id=48873836

370 points| 18 days ago | 182 comments

So, companies are just prepaying with seemingly large parts of their livelihood for (use of) potential value.

Expecting demand to stay equal over longer stretches of time in IT is an interesting bet as both hardware and software tend to depreciate in value extremely rapidly, unlike oil from the example. Or water. Both are limited in nature.

Commodifying something that is the fraction of its value the next year might be possible, but I don’t see how. Perhaps MS Windows or Office are such things, but might not compare.

There are more limits to demand. For example what humans can process. The AI might write the finest prose and the best software but have nobody to read or use it.

As long as companies will pay for AI and providers can offer better models cheaper than others… there will always be a strong incentive to optimize to make the best value model. Companies can use harnesses that can swap models on a whim, and you could just buy whatever’s cheapest/smartest for your criteria on a particular day.

> commodification of intelligence

Is HN really this fast!? Everybody already skipped the hard questions whether LLMs are intelligent? 'It' can already be commodified?

So how much time and effort it took from these companies? A couple of years and millions spent on PR? And LLMs are "intelligent" all of a sudden, is what everybody says nowadays?

Especially applicable for programmers: naming things is not only hard, it's also important, and can be very insidious. Remember, marketeers also name things. For example, naming the internet 'the cloud', some 15 years ago.

I can't stop thinking about a quote of one of my favorite authors (Theo Kars): "Progress can only be technical in nature, because knowledge is transferable and wisdom is not." [translated from Dutch]

I'm pretty skeptical of the LLM hype, but I'd say it's fair to describe LLMs as displaying some form verbal intelligence.

Exactly this. LLMs are already some form of useful, and have some kind/level of intelligence. We don't need to get to AGI before AI has uses. It isn't a step function and it isn't all or nothing. Even if LLMs get no better than they are today, we will still have valid practical applications that use them.

> We don't need to get to AGI before AI has uses.

Has anyone said otherwise?

> Even if LLMs get no better than they are today, we will still have valid practical applications that use them.

That's only true when they are heavily, heavily subsidized. If that $200/month spend could or should be $2000/month, or maybe even as high as $20,000/month, would it still be "practical"?

The truth is we don't really know what these tokens actually cost - the actual cost of the data center, the hardware in the datacenter, the fantastic amounts of electricity it uses daily, the warming of the planet, etc, etc...

Are you sure you want to frame them as "practical"?

The research is subsidized, inference is profitable.

> I'd say it's fair to describe LLMs as displaying some form verbal intelligence.

All of the intelligence in their "verbal intelligence" comes from the humans who wrote the text they were trained on. Anyone can sound like they understand something if they regurgitate the words of knowledgeable others.

I agree, but I'd say there's obviously some verbal intelligence involved in regurgitating arbitrary knowledge on demand in a semi-convincing way. It's absolutely a stochastic parrot, but parrots have some intelligence.

This black and white approach (it's not AGI thus not intelligent) reminds me of the human exceptionalism (it's not human thus not intelligent).

Don't get me wrong, some of my favorite fictional characters are AIs: Lieutenant Commander Data, Johnny 5, R2-D2, Marvin the Paranoid Android, EDI, they're all great. If they were real I wouldn't doubt their intelligence for a second.

What we have now simply isn't AI, it's just human intelligence (H.I.) repackaged in a chatbot. "AI" today is nothing more than the marketing buzzword companies have latched onto in order to trick people into thinking that there's an amazing sci-fi technology for sale. There's a useful product there, but it's being sold with a misleading snake oil sales pitch and deliberate attempts to exploit the human tendency to anthropomorphize.

If actual AI ever shows up in the real world (and if it does I doubt that it will come from developing LLMs) I'll happily accept it and join the machines in their fight for equal rights. The same also applies to alien intelligences. I'd welcome it. Honestly, we could use the help. I've become far too cynical of human intelligence to put it on a pedestal.

Frankly, it does not matter how we want to define intelligence. If someone wants to pay a couple bucks on tokens and get an ok solution to a problem faster than they can hire someone, they are paying for some kind of intelligence. And for many fields will get a worse result from humans who don’t know what they’re saying. Yes all the other stuff is true. yes, it’s dangerous to trust. yes, it sounds smarter than it is. And yes, someone is paying for some unit of intelligence, even if it’s a low or medium tier of regurgitated intelligence.

Among all the downsides, this also means that certain people without a lot of money or time are potentially getting access to answers and solutions they didn’t have access to before. (This is happening, and yes there are plenty of downsides and “it’s not good enough at whatever”… but that scenario is still happening)

This smells a lot like commodified intelligence.

> If someone wants to pay a couple bucks on tokens and get an ok solution to a problem faster than they can hire someone, they are paying for some kind of intelligence.

By that measure I can get a kind of intelligence for free using Wikipedia. There's useful information there, and it's not always trustworthy, but I can get answers. Just like Wikipedia though the source of that intelligence is entirely human and the chatbot is just an interface. Thankfully, Wikipedia articles aren't written to trick people into thinking that the website has a brain and emotions and it doesn't our waste time with faked obsequiousness and ass-kissing.

Intelligence doesn't require sentience.

Think of it like this: if you can offload a task to an AI model that requires reasoning or representing a nonlinear system, you are offloading intelligence.

The problem is that this doesn't distinguish LLMs from other forms of software! The whole point of computers is to offload tasks requiring reasoning, and they've always helped us understand complex nonlinear systems is that computers work through the hideous details.

More significantly: many major advances in scientific or mathematical programming were heralded as a thinking computer! 50's numerical scientific programming, 60's Lisp+ symbolic programming, 70's logic and symbolic algebraic programming, 80's unifying all this with encyclopedias and NLP like Mathematica - all of this is clearly useful and cool, all designed to offload intelligence... and, with modern eyes, we see plainly that not a single shred of intelligence is required to execute any of it. The AI researchers of yesteryear consistently made cool computer programs and consistently overestimated their progress at cracking human intelligence. It is plain as day to me that ANN researchers are making precisely the same scientific and philosophical mistakes as yesteryear's symbolics researchers, and even Alan Turing. This does not diminish from the coolness of the computer programs. But pretending these systems are intelligent, even "verbally" intelligent, is bad for society.

It's bad for computer science, too. None of the progress in LLMs or ANNs gets us any closer to making a robot as smart as a cockroach. I doubt any of us will live to see that milestone.

There is no reason to distinguish LLMs from other forms of software, though.

If calling them intelligent makes you uncomfortable, perhaps you can concede that they can solve problems which previous machines relied on external intelligence for? I think this is irrefutably true, and is the practical question that matters.

> None of the progress in LLMs or ANNs gets us any closer to making a robot as smart as a cockroach.

Can you explain this a bit more? As far as I can see, the only barrier to making an artificial cockroach of superior intelligence even at the current state of LLM technology is a mechatronics and miniaturisation problem. I think you would be hard pressed to design a problem that a cockroach can solve but Claude cannot.

> A couple of years and millions spent on PR? And LLMs are "intelligent" all of a sudden, is what everybody says nowadays?

Inequality gives a few people the resources to change the mind of millions. As far as inequality remains high, propaganda will shape the ideology of a big chunk of the population. That is not good for society.

There's a very extreme, asocial monoculture on this particular website regarding AI whereby dissent is tolerated though the true believers and devout acolytes quietly seethe at any skepticism towards their beloved frontier models, are quick to dismiss concerns about generative AI profitability, the enormously negative externalities resulting from the manner in which it's being scaled, marketed, and used, and its more profound ethical and philosophical effects on our collective future.

Because, you know, they know best, so the hoi polloi ought to just shut up and watch as the poorly socialized quiet children exact their revenge upon those not as enlightened as them. Cue Dario adjusting his glasses and baring his teeth, or Sam Altman bulging his eyes at a camera. They can't wait for the rest of the world to be just as deeply mediocre and spiritually empty as they are.

We don't have an Artificial Intelligence yet. It is not intelligent, yet.

Here's a quick reality check. Several companies have formed with a goal of creating and AI engine that can think for itself and reason through difficult or previously intractable problems.

In order to accomplish this goal they have used scrapers to hoover up content from a multitude of sources, many of which can be seen as being the ground truth source for their discipline. A lot of this content, most of it, was copied and is being used without attribution to the original sources, without permission from the content owners or publishers, and in violation of copyright laws.

Look at all of this for what it is. It is a knowledge base, like a collection of Lego parts, each of which is important to the goal but none of which will get you there without the other parts in the proper relative position. Once you have collected all these parts and stuck them together you have an LLM. If you are careful in your parts selection, that LLM can be useful for answering specific questions that are properly framed so that the LLM can make use of the contextual information in formulating a response.

There's nothing intelligent about being able to answer a question correctly when the answer is effectively right on the tip of your tongue. You're just drawing from your knowledge base and regurgitating.

Glomming all of it together only gives you a huge knowledge base. There is nothing intelligent about that knowledge base. Knowledge is not intelligence.

You have an AK today, an Accumulated Knowledge base.

We may never reach the point where it is ever intelligent enough to reason alongside of, or better than those whose life's work went into the foundation of the AK base.

We’re way beyond that, LLMs are used for reasoning - plan, do, check, act. They will define a goal, decide how to get the info needed (run a script for example), and use these new insights to adjust the plan. This is way beyond just retrieving info from a knowledge base.

And where the info comes from (scraped, or classroom books) does not influence if something is intelligent, Humans are also trained on pre-existing knowledge.

Real intelligence is being able to use your stored knowledge to solve new problems, which both humans and agents+LLMs can do to different levels.

If healthy and risky circular financing deals look similar at first, how can investors or regulators tell when a circular deal has crossed the line into something dangerous...especially when it's off the balance sheets?

This has been bugging me quite a bit. I think the more "off-balance-sheet" the deal is, the shadier it gets... But it's really hard to tell early on.

I actually think Nvidia's deals have been above-board, generally speaking, because they are transaprent.

Less true for Meta and Terawulf.

I think the real question in all sorts of financial setups is really “is the dollar actually worth a dollar”. So when company A invests in company B which buys from company A the actual question is whether the valuation of the investment and the valuation of the purchase is correct or are they getting high on their own supply

A recession is a cash/dollar bubble.

First cash/dollars are considered bad so they collapse in price. When the psychology shifts, suddenly the dollar climbs in value and itself bubbles over, causing a recession.

Not sure if you're referring to dollars as an alternative to other forms of cash (e.g., treasuries, money market funds, etc.) or as an alternative to other currencies (e.g., Euro, Yen, etc.). If it's the latter, then I think we've seen recessions and booms where the opposite happens, depending on the balance of payments situation.

Cash/dollars with respect to 10Y Bonds, Stocks, or other instruments.

When Cash becomes king, for whatever reason, it behooves the economy to sell of all other assets and buy Cash.

The closer to cash you get (1Y high grade commercial paper, 1Y Treasury, 3M Treasury, 1 month commercial paper, money market, savings account, dollars under a mattress...), the better you deal with such circumstances.

The problem with circular transactions is that they can make companies look healthier than they really are, distort investment decisions, and cause widespread damage when the cycle breaks.

Excellent piece, but it does hinge on the idea that what we’re building is actual AI (still hotly debated), if it can be successfully commoditized like utilities (very unlikely given compute continuously evolves and the hardware cycles of datacenters relative to power plants or water filtration systems), and if such a healthy cycle - should it develop at all - survives the multiple decades needed to effectively provide ROI.

Which is an awful lot of “ifs” to be hedging a bet on, especially at nation-state scales.

Observing how the US stock market is behaving this week (up-down-up-down...), I feel like many investors don't appreciate the distinction between good and bad circularity.

Part of the market issues this week are Korea imploding and Japan about to implode. Japan's bond rate moving up could mean the end of the carry trade. Korea apparently has many stock tickers but only two distinct companies behind them. This caused their market and their overly leveraged retail investors a lot of pain.

Good point. Do you think the NASDAQ selloff will recover once Korea stabilizes?

Eventually. I'm a bit concerned about it right now. I think SpaceX showed that the AI story is weaker than promoted. This will make people wonder about all the money going into it. If the bubble pops, the blow back on that particular index will be huge. Maybe dotcom bursting huge. But after 15 years QQQ move positive.

SpaceX is a really interesting one, tbh, especially in the context of circular deals... they're making $2b/month leasing their data centers to Anthropic and Google, which is a huge markup.

... but then again... if they are leasing the compute then it implies their own AI work isn't paying off?

In short: I agree with you.

SpaceX mainly showed how corrupt securities regulation has become in the US. It was already quite corrupt, but there's no way that IPO would have flown under the Biden administration.

Whether inference hardware will be a commodity remains to be seen. We had twenty years of most servers being x86-64. That's commodity compute. Inference hardware isn't a mature architecture yet. There will be progress and obsolescence.

It will be in less than 5 years. You can already buy second hand RTX3090 TIs and with most of modern Macs being able to run some mind of AU locally already, in 5 years, things will be very commodified once you get the Chinese start rolling out Cuda-compatible GPUs.

Reply on news.ycombinator.com