Cory Doctorow: What Kind of Bubble Is AI?

I think an overwhelming majority of users prompt it to generate a wedding speech speech in the style of H.P. Lovecraft or whatever, have a giggle and never use it again. I am so sceptical of the claimed figures, especially the often-repeated claim that "ChatGPT [is] the fastest-growing consumer application in history"[0].

[0] https://www.reuters.com/technology/chatgpt-sets-record-faste...

I guess the thousands of rides I've taken so far were just... nothing? I'll tell you what. I remember the world before Uber. It fundamentally improved my life.

If I were one of the shrill victim types, I would call this take ableist against blind people who have objectively gotten value out of Uber. Instead, I'll just call it silly.

I work in AI now and spent 3 years in the crypto space, to say crypto has left nothing of value really is an over simplistic view. To say Uber is a failed company (I worked at a ride share company too), wow. Does he not remember the utter hellscape of getting a taxi in SF 10 years ago?

The executives at these companies (mine too) openly declare a bubble similar to .com in the offing, cash in on it while you can. Maybe you’ll get lucky and get early access to some of these models we’ve got cooking to use for personal projects ;).

Wtf is Doctorow talking about here? Because the voice transcription, text-to-speech, optical character recognition, and language translation applications are already changing the world. I'm open to calling it a "bubble," but the sentence above is absurd. Reminds me of the John McCarthy quote "As soon as it works, no one calls it AI anymore"

This is the part that doesn't seem like an unsustainable bubble. Initial training is very expensive. The cost of running a trained model keeps coming down as we find more and more ways of making it smaller while maintaining a high level of accuracy. It really comes down to if these advancements can keep going and meet costs before investment dries up.

As for other AI technologies, we already have many proven examples in use today generating value. Things like automated text scraping are everywhere. Applications like TTS and ASR are somewhat common-place as well. I think the big difference between AI and previous tech bubbles is that AI has a huge amount of useful research literature demonstrating various kinds of value underlying it. There is a question of whether research value can be translated into business value effectively, but there are very real improvements in AI that underlie the explosion in AI businesses. I'm hard pressed to find the same level of technical innovation in crypto over the same time period. As I understand it most cryptocurrencies run minor variations of one of two algorithms. I also think the dotcom bubble in general had very little of the same research underpinnings. There are a lot of AI problems where we are twice as good as we used to be five years ago. It's not surprising that some start-ups will exist to check if that's good enough to generate business value. I'd be very shocked if this AI Bubble didn't leave a lot of good things behind because I think it's very unlike the other bubbles in the fundamentals that underpin it.

But it strikes me as intellectually dishonest to assume in advance that everything that is currently being built will fail entirely.

Doctorow is so relentlessly cynical about tech in general and Silicon Valley in particular that I find it really hard to read his stuff without rolling my eyes.

Take this section:

> In other words, an AI-supported radiologist should spend exactly the same amount of time considering your X-ray, and then see if the AI agrees with their judgment, and, if not, they should take a closer look. AI should make radiology more expensive, in order to make it more accurate.

Implicit in this statement is the assumption that AI evaluation of X-rays will never get any better than it is right now. Why? Why would we assume that when the long arc of technology bends always towards things getting better/faster/cheaper?

Last week I tried really seriously to use JetBrain's A.I. assistant at work.

It could answer questions about my code sometimes, it could figure out what my database schema was by looking at the stubs JooQ built out of the tables.

I was working on a recursive CTE for postgres

https://www.postgresql.org/docs/current/queries-with.html#QU...

which is a little tricky to express in JooQ because you run into one of those "need to access the definition before it is defined" problems that they invented the Y combinator to deal with... Most people just hand-code the table reference in the join like is done here

https://www.jooq.org/doc/latest/manual/sql-building/sql-stat...

The assistant tried to write the query numerous times, it got the basic idea right but could never quite figure out how to close the loop on the recursion. It exhausting to keep cutting and pasting code from it and then hitting alt+ENTER to bring in imports and finding sometimes I got the wrong imports (java.lang.Record vs jooq.Record), etc.

It would often say things that were totally wrong such as needing JooQ 3.15 to write CTEs because it misread the documentation.

Now my experience is that many ignorant programmers are also belligerent and the assistant is always polite and actually will change it's opinion if you show it evidence to the contrary... But maybe that makes it more dangerous because the better it gets socially, the better it gets at keeping you in a tarpit.

Unlike traditional bubbles we get qualitatively different outcomes at greater scale. A million GPT-2s is nothing like 1 GPT-4. Researchers are continually surprised by new abilities at greater scales.

None of those other bubbles had something truly new at their core. You can say AI image generation is gimmicky, but you can't deny that it's a capability that machines have now and didn't before.

A rush of speculative investment into finding applications for that capability makes sense. Some will turn out to be pets.com and others will become Amazon. But the world has already changed, and all that remains is to figure out who is going to make money from that.

Wikipedia's definition:

>An economic bubble (also called a speculative bubble or a financial bubble) is a period when current asset prices greatly exceed their intrinsic valuation, being the valuation that the underlying long-term fundamentals justify.

I'm not sure the AI stuff even qualifies as a bubble under that. If you assume better than human AI comes, which is quite likely, it will be able to do all the things and more so you could hazzard a value of the order of world GDP which is approx $100 trillion.

I agree. A product needs at least an LLM under the hood, or Bayesian inference, or machine learning, or neural nets to be deemed 'AI'. Otherwise it's just a bunch of if...else statements.

We'll just run self-hosted models, they're already becoming reasonable.

Microsoft owns 49% and we’re already seeing its usefulness in the world of enterprise organisations where Microsoft is frankly becoming irreplaceable on the office side of things. We’re only in the early stages of using Copilot, bing agent, 365 or whatever it’s called, and it’s rapidly changing our “boring” business processes. It’s things like having it listen in on Teams meetings and write summaries. Making non-ugly sleep inducing presentations in a financial company focused on green energy… seriously… It’s replaced all our external spending on corporate designs, it’s not on par with what you’d see on top of the line communication, but it’s good enough to send out to our investors. It’s also good enough to replace any form of icon libraries and stock pictures. We use it to enhance excel (not sure how exactly, and I’m pretty sure I don’t want to know considering what lives in our shared excel sheets isn’t supposed to live there). It’s helping more digitally inclined PMs and business people automate a lot of their processes in ways no other no-code or RPA tool has ever done before them. It’s helping BI write shitty software, which is perfectly fine since we can always re-write the stuff that needs to be efficient and maintainable. And so on.

But it’s also just part of the Microsoft package. If you or I buy into this, or even smaller companies do, you’re going to pay the prices listed by Microsoft. That’s not how enterprise pays though, we go through these 3rd party dealers with all sorts of fancy certifications and whatever that they pay to have, and then we buy packaged subscriptions on things we need. Of course if you compared, we probably pay more, but it doesn’t look like that on the budgets, and that’s what matters when you roll these things out. So for us AI is similar to how Microsoft Teams became our only communications tool. Not because it was better (at least not at the time) but because it was “free”. AI is also “free” for us, and because it is, it’s going to lock us even harder into Microsoft as a vendor. It’s not like there was really an alternative anyway, but now there really isn’t. And that’s not a bubble, it’ll only strengthen Microsoft unless EU or US regulation does something about it, but how can they when they’re using these tools themselves?

If not, we don’t have a bubble, yet.

That part, I actually agree with. So much lately feels like people trying to inflate a garbage bag, make it look real pretty, and then run away with the money before you notice. And not really care about the fallout afterward, cause they got their island, and in Hollywood parlance, their "** U money". Never understood that in Hollywood (your job's like, the most, desirable).

Tech bros, I can kind of understand wanting to get away. Specially if you never cared about tech, and just wanted quick cash. Probably a huge number of humans who get drawn toward the FAANG realm cause of money, and never even slightly cared about tech. Just see $'s.

Man this guy really knows nothing about energy...

Enron happened at the start of the shale and natural gas revolution that reshaped American energy and global politics. Enron's energy traders alone (including John Arnold) helped structure global markets for natural gas. Their physical pipeline assets are owned by Berkshire Hathaway. The fraud of Enron was on top of a real set of businesses, with massive physical assets!

I expect better from Doctorow, who has long advocated for people taking personal control of technology.

The cool thing about this generation of AI is that you can run it on home, or even on your phone. A $3000 MacBook (yes, expensive) can run a useful voice controlled AI that can do a lot of cool stuff. Image generation models also run locally, and there has been an explosion of tools, running on people's PCs, that allow for entire new creative workflows.

Is the training subsidized? Yes, it is. However finetuning is crowd sourced, and again we find ourselves in a new field where anyone who wants to can quickly get up to speed and make serious real contributions.

He later degradates the "small" models, ignoring that Mistral's latest release is powerful, people can run models that approach GPT 3.5 (which is incredibly useful!) Powerful voice recognition, the type only dreamed about a decade ago, is now available for everyone. The image generation models that are available for personal use are incredibly powerful.

Research papers are coming out left and right about how to do video generation. Models are getting cheaper to run all the time, and we are exactly 1 (hard fought...) manufacturing node improvement away from this stuff being usable on affordable home machines.

> But with AIs’ tendency to “hallucinate” and confabulate, there’s an increasing recognition that these AI judgments require a “human in the loop” to carefully review their judgments.

FFS people hallucinate things that aren't there. Every time a human accesses a memory in their brain the memory gets altered a little bit. Fact checking exists for a reason.

> AI companies are implicitly betting that their customers will buy AI for highly consequential automation

How about 10xing current employees?

OpenAI is doing a terrible job of capturing the value it provides. Heck most people are doing a terrible job of utilizing AI in their day to day, versus what they could be doing.

> All the big, exciting uses for AI are either low-dollar (helping kids cheat on their homework, generating stock art for bottom-feeding publications) or high-stakes and fault-intolerant (self-driving cars, radiology, hiring, etc.).

"Take these 10 PowerPoint slide decks, extract the graphs we used in last quarter's summary report, and bring all the chart styles into alignment with our corporate branding".

"Go through these last 5 years of corporate financial filings and highlight any areas of business that you think the company is trying to cover up and obfuscate".

"Add error checking and retry logic to this shell script, as well as output logging when an error occurs"

"Write a mock for this object"

"The explanation by teacher gave of how to integrate a sine wave didn't make sense, can you provide an alternate explanation?"

I think the term bubble is getting mistreated here. Poor little thing.

If enthusiasm and investment is a measure of “bubble” then every new investable technology wave that comes along will be one.

If “bubble” means “getting more resources thrown at it than its potential justifies” then we have an entirely different conversation going on.

So, which is AI? I’d argue the potential for AI is larger than the entire Internet thus far, though whether LLMs specifically are going to pan out is another thing entirely.

What we have is the tendrils of capitalism reaching out to try to fund the next general purpose technology (gpt!) wave, which when you think about it makes a lot of sense. They aren’t bubbles, they are experimentation phases while the mechanisms of our society attempt to determine the breadth and depth and general utility of the new innovations. Call that a bubble if you must, but it’s really something else.

Rather than use the bubble terminology, I prefer to think of things as waves as it’s entirely more accurate.

Here’s a well written summary by Nicholas Bergman so I don’t have to write so much:

Charles A. Beard warned almost one hundred years ago: “Technology marches in seven-league boots from one ruthless, revolutionary conquest to another, tearing down old factories and industries, flinging up new processes with terrifying rapidity.” This image of technology is bleak at best, but there is a lot of truth tucked away inside this Technological Determinist statement.

Throughout the history of mankind, significant changes in how society, business and culture function has followed key developments in our technology. Across thousands of years, humanity has ridden five distinct technological waves – waves that have all uniquely impacted how we experience the world, and waves that have all eventually taken backstage and given way to a new on one.

The five waves that have crashed into our society so far are as follows: “Early Mechanisation” (1770s to the 1830s), “Steam Power and Railways” (1830s to 1880s), “Electrical and Heavy Engineering” (1880s to 1930s), “Fordist Mass Production” (1930s to 1970s) and “Information and Communication” (1970s to 2010s). We are now beginning to see the surge of mankind’s sixth wave: “Miniaturization”. How do we know? Because all prior waves have been charged by the same forms of technologies: General Purpose Technologies (GPTs) to be exact. GPTs are paramount in connecting the past with the present, and, in the most reductionist sense, a GPT is a single generic technology that develops over time but can still be described as the same technology over its lifetime – the printing press, the wheel and the computer are three quintessential examples of GPTs. Initially, these technologies have immense space for growth and are eventually widely adopted – permeating different sectors and industries over time. In order for a new technological wave to grow, there must be at least one, but often multiple GPTs that drive it. These technologies themselves must go through a three-step process in order to reach widespread adoption: experimentation, expansion and transformation.

Taking cars as an example, at first, no-one really understood the applications of this new technology, it wasn’t viewed as a great method of transportation and was just a novelty way of getting from A to B – this was its experimental phase. Following this, the expansionist phase, where we begin to see interesting innovative use cases as businesses incorporate the technology into their scopes, and a democratization of the technology due to increased affordability and reliability. Finally, we see the transformational phase, characterized here by the emergence of suburbs, drive-in cinemas and shopping malls. Essentially the technology began to transform how society worked, how society was now experienced was determined by this new technology.

So, calling it “a bubble” cheapens what it is. It’s phase 1: experimentation. And in many ways Cory is following the same template when he talks about whether it will leave anything behind or not, but it’s not a pessimistic process like the term bubble is implying. It’s a deeply healthy and recurring process of humans exploring new technologies in order to determine whether they are a general purpose technology and whether it can be, and how, widely deployed.

This really puts some perspective on the crypto wave as well. If you think of it as a giant beta to see what worked and what didn’t, with eventual deployment of solid working solutions for digital exchange finding foothold, it’s hard to see it as a “failure” as that nomenclature is again a diminishment of what is simply the healthy and natural process of exploring new technology waves for broad purpose and suitability.

Whether you believe that “miniaturization” is the wave is not the point though. We may not know exactly what form the wave will take until it’s quite obvious, or every investor would be right 100 percent of the time. But the pattern is the point.

If you want to read more, this is where the quotes are from:

https://www.maize.io/cultural-factory/technological-waves/