AI: Not Magic, Not a Bubble Pop, Just… a Lot of Compute (and Some Useful Tools)

I’ve been writing software for a little over twenty years, which is long enough to have seen the same story arc repeat itself in different costumes:
new tech arrives → people declare it changes everything → vendors slap it on a slide deck → everyone argues about whether it’s a bubble → eventually it becomes… normal.

AI is currently doing all of that at once, just louder and with better branding.

I watched an interview with Dario Amodei (Anthropic CEO) recently, and it landed in an interesting place for me. I’m not an “AI will replace all humans by Tuesday” type, but I’m also not in the “it’s just autocorrect” camp. What he said felt like the most sensible version of the hype: the technology is real, the business risks are real, and the social knock-on effects are going to be… spicy.

This is my attempt to explain it from the perspective of a working software bloke who’s mildly allergic to buzzwords, but perfectly happy to use a tool that helps me get stuff done — including writing blog posts like this one.

The bit where I agree with the hype (annoyingly)

The most convincing part of Amodei’s argument isn’t “AGI” or “superintelligence” or whatever acronym we’re meant to be frightened by this week. It’s simpler:

Keep pouring compute and data into these models and, with a few tweaks, they keep getting better at a wide range of tasks.

If you’ve built products for any length of time, that should ring a bell. General tools that keep improving tend to get adopted. You don’t need sci-fi. You just need enough usefulness per pound spent.

The “usefulness” is showing up most clearly in boring, high-value work:

• coding assistance
• summarising long documents
• drafting content
• pulling structure out of chaos
• turning a vague idea into a first pass you can actually edit

That last one matters more than people admit. A surprising amount of professional life is… starting.

The bit where I stay sceptical

Where Amodei gets properly interesting is when he says: tech confidence doesn’t automatically mean the economics make sense.

That’s the part hype merchants always skip. They talk like:

• capability growth = guaranteed profit
  when the truth is:
• capability growth + bad timing = you can still blow your legs off

He uses this idea of a “cone of uncertainty”: you have to commit to massive compute spend now (because data centres take ages to build), while you’re still guessing what demand and revenue will look like in two years’ time.

If you’re a software person, this is painfully familiar. It’s the same energy as:

• “We’ve got to pick a database and an architecture now”
• “We don’t yet know whether we’ll have 1,000 users or 10 million”
• “Also, changing it later will be expensive and embarrassing”

Except the numbers here are… slightly larger than my AWS bill.

So yes: I’m sceptical of the “this will obviously pay off” certainty. Not because the tech isn’t impressive, but because markets punish bad forecasting and everybody’s forecasting in fog right now.

“Bubble” is the wrong argument (but the right worry)

When people ask “is this an AI bubble?” they often mean one of two things:

1. Is AI fake? (No. It’s very much real. A bit too real in places.)
2. Are we overbuilding and overspending? (Possibly. And that’s the more adult question.)

Amodei’s point isn’t “AI will fail.” His point is:

• AI can succeed
  and
• some companies can still overextend themselves trying to win

That’s not a contradiction — it’s Silicon Valley’s favourite pastime.

Chips, “vendor financing”, and why everyone’s suddenly a utilities company

There’s also a segment on what he calls circular or vendor-financed deals: chip makers and cloud giants invest in AI companies… who then spend a lot of money on chips and cloud.

If you’ve been around enterprise software long enough, you’ll recognise it: this is not a new trick, it’s just wearing a hoodie.

The charitable version is: it’s a practical way to fund huge infrastructure commitments when no one has £50bn sitting in a drawer.

The uncharitable version is: it can hide risk if everyone assumes tomorrow’s revenue is guaranteed.

Reality is probably… both, depending on the deal and the optimism levels involved.

Model wars vs actual work

One bit I appreciated: he basically shrugs at the consumer “code red” drama and says Anthropic is more focused on enterprise use cases.

As someone who builds software for businesses, that tracks. In enterprise, the questions are rarely “is this the most viral model?” They’re usually:

• can it help my team ship faster?
• can we trust it with our data?
• can we integrate it cleanly?
• will switching away later be painful?

And yes, switching often is painful — not because models are identical commodities, but because how you use them becomes embedded in your workflows (prompts, tools, guardrails, downstream behaviour). Once it’s in production, it’s not a toy.

The jobs bit: I don’t love it, but I can’t ignore it

Amodei’s “jobs” section is the one that made me put down my tea for a second.

He’s not doing the usual apocalypse routine. He’s saying:

• let’s acknowledge the landmine
• so we can walk around it

He suggests three “levels” of response:

1. Companies should aim not just for cost-cutting automation, but also for new value creation where humans become higher-leverage.
2. Government will likely need to get involved (retraining, policy, possibly tax/distribution) because the transition could be too fast for the market to absorb cleanly.
3. Society might eventually need to rethink the role of work itself — less “work or die” and more “work for meaning”, which is a huge cultural shift.

I’m sceptical of big, top-down redesigns of society (I’ve seen what happens when someone tries to redesign a menu navigation). But I’m also realistic: if productivity jumps hard and fast, the distribution argument becomes unavoidable.

So where do I land? Use the tool, ignore the incense

My current stance is basically:

• AI is not magic
• AI is not harmless
• AI is not optional anymore for a lot of knowledge work
• AI is, however, extremely useful when you treat it like what it is: a tool for accelerating first drafts and reducing friction

I use it the same way I use any decent dev tool:

• not to replace thinking
• to remove busywork
• to get from “blank page” to “something I can improve”

A good model won’t write your blog post for you.
But it will happily produce an awkward first draft, a structure, a set of bullet points, or a tidy summary of a messy transcript — and that’s often the hardest part done.

Which, ironically, is exactly what I’m doing here.

Practical takeaways (without the motivational posters)

If you’re a software person or you manage software people, here’s the grounded version:

• Invest in workflows, not vibes. The ROI comes from repeatable processes, not demos.
• Treat cost and latency as product features. Compute is a real constraint, not an abstract cloud line item.
• Assume model switching has hidden costs. Build abstraction layers and keep prompts/tooling maintainable.
• Don’t wait for perfect certainty. But also don’t bet the company on a straight-line forecast.
• Use AI to start faster, not to think less. It’s best as a multiplier for competent humans.

The future is probably not “AI replaces everyone.”
It’s more like: people who can use AI well replace people who can’t — and that’s a much more mundane, and therefore much more plausible, outcome.

Now, if you’ll excuse me, I’m off to do what every seasoned software developer does after publishing something:

Ship it… then immediately spot five things I want to rewrite.