Thoughts

Yesterday I watched Ed Zitron make the case that generative AI is a con.

I don’t think I’d go that far.

In fact, I disagree with him on something fairly fundamental: I think this stuff is incredibly useful.

I use AI every day. It has materially changed the way I develop software, research problems, explore ideas and get from an idea to something tangible. I’ve seen enough genuinely useful applications of it that I find it difficult to accept the argument that the underlying technology is somehow fraudulent.

But I agreed with an uncomfortable amount of everything else he said.

In particular, there’s a question I’ve probably avoided while enjoying extraordinarily capable AI for a relatively small monthly subscription:

Who is actually paying for all of this?

The product and the business are different things

It’s easy to collapse two arguments into one.

Is generative AI useful?

And:

Is generative AI currently a sustainable business?

They’re not the same question.

Something can be an extraordinary technology and a terrible business.

Zitron’s argument focuses heavily on the economics: enormous infrastructure investment, expensive compute and AI companies spending extraordinary amounts of money while trying to establish a market for their products.

I don’t have enough information to confidently predict where those economics ultimately settle.

But as a user, there’s something obviously strange happening.

I pay a relatively ordinary software subscription and receive access to an extraordinary amount of computation.

I can write code, analyse documents, generate images, research subjects and have long conversations with increasingly capable models.

And I can do it all day.

The obvious question is whether what I’m paying bears any resemblance to what I’m consuming.

We’re being subsidised

This was probably the most interesting part of the interview for me.

The suggestion is that heavy AI users can consume substantially more compute than their subscription price could economically support.

Whatever the exact number is for an individual user — and estimates here should be treated cautiously — the broader point is important.

AI companies are subsidising adoption.

That’s not particularly unusual.

Technology companies have been doing versions of this forever. Venture capital subsidised taxis, food delivery, streaming, cloud storage and social networks while companies raced to establish themselves.

Make the product cheap.

Acquire users.

Establish the habit.

Work out how to make money later.

The difference with AI is the scale of the infrastructure required to keep doing it.

Every time I ask an AI model to do something, something actually has to happen.

Compute is consumed.

Electricity is consumed.

Infrastructure has to exist to service the request.

The marginal cost isn’t zero.

And there are an awful lot of us asking an awful lot of questions.

We’ve heard this before

There is something familiar about all of this.

For years people questioned how social networks could possibly be sustainable while billions of people used them for free.

Turns out they found a way.

We paid with advertising, attention and data.

So I’m wary of assuming that because the economics of AI look strange today they can never work.

Markets change.

Hardware gets better. Models get more efficient. Competition changes prices. New business models emerge. Businesses find valuable applications they’re willing to pay considerably more for than consumers are.

Today’s cost structure doesn’t necessarily tell us what AI looks like in ten years.

But it also doesn’t guarantee that everything works out.

The social networks found their business model.

We might reasonably argue that we didn’t particularly like what it turned them into.

What does AI cost when AI costs what it costs?

This is where it becomes interesting.

What would happen if the subsidy disappeared tomorrow?

Imagine your £20 AI subscription suddenly cost £100.

Or £300.

Or usage was metered aggressively and every conversation had a visible price attached to it.

Would you still use it in the same way?

I probably wouldn’t.

I’d still use AI.

But I’d think much harder before using it.

And that potentially changes the product.

Part of what makes current AI so powerful is that experimentation is effectively free at the point of use.

I can try something.

Then try it another way.

Then ask a follow-up question.

Then abandon the entire thing.

That freedom is incredibly important to how I use it.

Start putting a meaningful price against every interaction and behaviour changes.

Suddenly we’re not just asking whether AI is useful.

We’re asking whether this particular AI interaction is worth what it costs.

That’s a much higher bar.

Especially when it’s wrong

There’s another problem with charging the true cost.

AI is still frequently wrong.

Not useless.

Not slightly imperfect.

Wrong.

It can misunderstand a requirement, invent a fact, produce broken code or confidently lead you down completely the wrong path.

I’m comfortable with that because I understand the bargain.

I use AI as an extraordinarily capable assistant whose output I still need to evaluate.

At current consumer pricing, that’s an incredible proposition.

But imagine being charged substantial amounts for individual outputs.

The relationship changes.

If I pay pennies for an answer and it’s wrong, I shrug and try again.

If I pay £5 for an answer and it’s wrong, I’m annoyed.

If an organisation spends thousands running an automated AI process and then needs humans to check everything it produces, the economics become considerably harder.

Traditional software has bugs, of course.

Humans make mistakes.

But generative AI introduces something unusual: unreliability is an inherent characteristic of the system rather than simply a defect waiting to be fixed.

That matters when we’re trying to work out what the service is actually worth.

Cheap AI makes mediocre AI useful

This might be the bit we underestimate.

AI doesn’t always have to be amazing to be useful when it’s cheap.

If it gets me 80% of the way to something in 30 seconds, that’s valuable.

If three attempts are rubbish and the fourth works, that’s still potentially valuable.

If it saves me ten minutes on something mundane, brilliant.

But all of those calculations depend on the cost being sufficiently low.

Increase the price dramatically and suddenly reliability matters much more.

The economics and the capability aren’t separate.

They define the value together.

Enterprise might be the answer

There’s an obvious counterargument.

Maybe consumers aren’t where the money is.

If an AI system saves a company £500,000 a year, nobody cares whether the underlying API costs £20,000.

That’s a perfectly good business.

And there are clearly applications where AI can create that kind of value.

Software development is already one for me. There are tasks that would previously have taken hours that I can now complete substantially faster.

Multiply that across a development team and there is real economic value.

But that doesn’t mean AI creates that value everywhere.

Putting an AI assistant into every application isn’t automatically a business case.

Generating thousands of pieces of content nobody particularly wanted isn’t productivity.

Replacing a cheap deterministic process with an expensive probabilistic one isn’t innovation.

The technology still has to earn its place.

Maybe the bubble and the technology can coexist

This is where I probably differ most from the argument that AI is a con.

I think two things can be true simultaneously.

Generative AI can be genuinely transformative.

And:

The current AI investment boom can be completely irrational.

The dot-com bubble didn’t prove the internet was pointless.

It proved investors could be right about a technological transformation and wildly wrong about how quickly value would appear, which companies would capture it and how much those companies were worth.

The internet survived the bubble.

Most of the companies didn’t.

I wonder whether AI might follow a similar path.

So what happens next?

I don’t know.

And I’m increasingly suspicious of anyone who says they do.

Maybe inference costs collapse.

Maybe specialised smaller models dramatically reduce the compute required.

Maybe enterprise AI becomes sufficiently valuable to subsidise consumer access.

Maybe subscriptions become much more expensive.

Maybe advertising arrives.

Maybe today’s enormous infrastructure investment turns out to have been entirely justified.

Or perhaps there really is a painful correction coming.

What I am fairly confident about is that AI isn’t going away.

I’ve used it too much and seen it solve too many real problems to believe that.

But that doesn’t mean the current economics make sense.

And perhaps that’s the distinction we should be making more often.

The interesting question isn’t:

Is AI a con?

It’s:

What is AI actually worth when somebody eventually has to pay the full bill?