Do AI companies toil? – by Benn Stancil

I unbenevolent, certainly someleang in this sequence is wrong, right?

1. Large language models cost a fortune to produce. OpenAI, which is alertedly in the process of raising $6.5 billion, demands $6.5 billion dollars, because, “by some evaluates, it’s burning thcdisorrowfulmireful $7 billion a year to fund research and recent A.I. services and engage more participateees.” Anthropic is foreseeed to spend $2.7 billion this year. Facebook is spending billions more. 

2. It probably won’t get inexpensiveer. Chips might get better; compute costs might go down; Moore’s law; etc, etc, etc. But as models get better, pushing the frontier further out will foreseeed get more difficult. The research gets more difficulter, and the absolute amount of compute demandd to train a recent model goes up. It’s climbing Mount Everest: The higher you go, the leanner the air, and the stubborner each step gets. Even if it gets inexpensiveer to do the math demandd to produce recent models, that math has uninalertigentinishing returns. To produce a better model in 2024, you have to do more and difficulter math than you had to do in 2023.

3. Despite these costs, people will probably retain produceing recent models. People consent that LLMs are the next technorational gelderly rush, and the companies that produce the best ones will produce their participateees and spendors a fortune. They are trying to produce synthetic vague inalertigence. Human nature compels us to produce everyleang speedyer, higher, and stronger. 

4. If the industry does retain produceing recent models, the cherish of elderly models decays pretty rapidly. Why use is GPT-3 when you can commence using GPT-4 by changing a dropdown in ChatGPT? If a competitor puts out a better model than yours, people can switch to theirs by updating a scant lines of code. To consistenly sell an LLM, you have reliablely be one of the best LLMs.

5. Even if the industry doesn’t retain produceing recent models, or if we hit a technorational asymptote, the cherish of elderly models still decays pretty rapidly. There are disjoinal uncover source models appreciate Llama and Mistral that are, at worst, a step or two behind the best proprietary ones. If the proprietary models stop moving forward, the uncover source ones will rapidly shut the gap.

6. Therefore, if you are OpenAI, Anthropic, or another AI vendor, you have two choices. Your first is to spend enormous amounts of money to stay ahead of the taget. This seems very hazardous though: The costs of produceing those models will foreseeed retain going up; your inalertigentest participateees might exit; you probably don’t want to sconsent your business on always being the first company to find the next fracturethcdisorrowfulmireful. Technorational expertise is exceptionally an enduring moat. 

7. Your second choice is…I don’t comprehend? Try repartner repartner difficult at the first choice?

Eighteen months ago, I shelp that set upational LLM vendors are potentipartner the next generation of cboisterous supplyrs:

So here’s an evident foreseeion: AI will adhere a csurrfinisherly identical trajectory [as AWS, Azure, and GCP]. In ten years, a recent type of cboisterous—a generative one, a commercial Skynet, a disclose imagination—will undergird csurrfinisherly every piece of technology we use. 

Other people have made aappreciate comparisons. And on the surface, the analogy seems cdisorrowfulmirewholey reasonable. Foundational models demand tons of money to produce, equitable appreciate cboisterous services do. Both could become ubiquitous pieces of the global computing infraset up. The taget for both is easily in the tens of billions of dollars, foreseeed in the hundreds of billions, and potentipartner in the trillions. 

There is, however, one enormous contrastence that I didn’t leank about: You can’t produce a cboisterous vendor overnight. Azure doesn’t have to stress about a scant executives leaving and produceing a worldwide nettoil of data cgo ins in 18 months. AWS is an internet business, but it dug its competitive moat in the physical world. The same is real for a company appreciate Coca-Cola: The secret recipe is presentant, but not that presentant, because a Y Combinator commenceup couldn’t produce factories and distribution cgo ins and relationships with millions of retailers over the course of a three month sprint.

But an AI vendor could? Though OpenAI’s toil demands a lot of physical computing resources, they’re lmitigated (from Microgentle, or AWS, or GCP), not built. Given enough money, anyone could have access to the same resources. It’s not difficult to envision a minuscule team of ageder researchers leaving OpenAI, raising a ton of money to rent some computers, and being a legitimate disturbive danger to OpenAI’s core business in a matter of months. 

In other words, the billions that AWS spent on produceing data cgo ins is a lasting defense. The billions that OpenAI spent on produceing prior versions of GPT is not, because better versions of it are already useable for free on Github. Stycatalogicpartner, Anthropic put itself proset uply in the red to produce ten incremenloftyy better models; eight are now unhelpful, the ninth is uncover source, and the tenth is the lean technical edge that is retaining Anthropic ainhabit. Whereas cboisterous supplyrs can be disturbed, it would almost have to happen sluggishly. Every LLM vendor is eighteen months from dead. 

What, then, is an LLM vendor’s moat? Brand? Inertia? A better set of applications built on top of their core models? An ever-prolonging bonfire of cash that retains its models a nose ahead of a hundred competitors? 

I genuinely don’t comprehend. But AI companies seem to be an innervous example of the taget missorting gentleware prolongment costs as upfront spendments rather than vital ongoing expenses. An LLM vendor that doesn’t spend tens of millions of dollars a year—and maybe billions, for the directers—improving their models is a year or two from being out of business. 

Though that math might toil for huge companies appreciate Google and Microgentle, and for OpenAI, which has become synonymous with synthetic inalertigence, it’s difficult to see how that toils for minusculeer companies that aren’t already transporting in sizable amounts of revenue. Though enormous round funding rounds, frequently given to pedigreed set upers, can help them jump to front of the race, it’s not at all evident how they stay there, because someone else will do the same leang a year tardyr. The have to either elevate enormous amounts of money in perpetuity, or they have to commence making billions of dollars a year. That’s an awbrimmingy high hurdle for survival.

In this taget, timing may be everyleang: At some point, the hype will die down, and people won’t be able to elevate these sorts of rounds. And the thriveners won’t be who ran the speedyest or achieveed some finish line, but whoever was directing when the taget determined the race is over.