What If This is as Good as it Gets?

There’s a phrase I keep hearing in the AI world:

“This is the worst AI will ever be.”

It’s everywhere. Sam Altman said it about GPT-5. You’ll find it on panels, podcasts, Twitter threads. It sounds confident. Reassuring, even. As if progress is a guarantee. As if each new version will naturally be better than the last.

But the more I hear it, the more it bothers me.

Because AI is software. And software isn’t bound to improve forever. Sometimes it hits a wall. Sometimes it actually gets worse. And there’s no cosmic law that says large language models (LLMs) must keep getting smarter.

And what if we’re already there?

The Excel Problem

Let’s take Excel.

Is it better than it was 10 years ago? Probably. Slightly. A few more formulas. Some UI polish. Cloud integration.

But has it transformed? Not really.

That’s not a failure. That’s maturity.

Most widely-used software eventually hits this phase. It becomes refined, stable, and “good enough” for what people need. Further improvements become incremental, not exponential. You don’t upgrade Excel because you need to, you upgrade it because your IT department makes you.

So what if LLMs are approaching that point? What if GPT-5 is already closer to the peak than we think?

The Limits of Human Software

There’s another overlooked factor here: software is made by people.

And people don’t scale exponentially.

Unlike hardware, which can be manufactured and improved through engineering processes, software relies on human creativity, human judgement, and very human mistakes (yes, even AI software). Progress isn’t linear. It's messy. Political. Often emotional.

There’s a reason so many software rewrites fail. Or why new versions of beloved apps sometimes make them worse. Humans get tired. Teams change. Priorities shift. Bad decisions compound.

There’s also the coordination problem: making an LLM doesn’t just require good ideas, it requires aligning hundreds (or thousands?) of researchers, engineers, designers, and execs.

And all of that happens inside companies with deadlines, competitors, and budgets.

So the idea that LLM progress will just keep marching upward? That’s a bet on human systems continuing to function smoothly, creatively, and at scale.

History suggests that they won't.

Running Out of Data

Here’s another constraint: we’re running out of data.

Current LLMs have already consumed most of the high-quality text on the internet. OpenAI, Anthropic, Google have already scraped Wikipedia, Stack Overflow, Reddit, news sites, books, forums, PDFs, codebases, and more.

And that was enough to train GPT-4 and Claude 2 and Gemini.

But what comes next?

There isn’t exponentially more clean, diverse, high-signal training data waiting to be discovered. In fact, we’re already seeing models trained on synthetic data (i.e. the outputs of other models). That can help scale things up, but it also risks introducing noise, drift, or subtle degradation.

Without fundamentally new kinds of training data, or a radically new approach to learning, it’s hard to see how the next model leaps dramatically ahead.

More compute doesn’t magically produce more intelligence. You still need something to feed the machine.

So, What If This Is It?

So let’s imagine, for a moment, that GPT-5 (or its next minor iteration) is close to the ceiling for transformer-based LLMs.

What does that mean?

First: it's not the end of the world. GPT-5 is still an extraordinary tool. It’s already reshaping work, writing, coding, research, education, and more.

But second: it’s a wake-up call for the business models currently betting everything on perpetual, exponential improvement.

Because if you’re investing billions on the assumption that the next version of AI will be smarter, faster, cheaper, and world-changing, then you’d better be right.

The Business Implications

A huge amount of capital is being deployed right now on the premise that AI is still in its early phase. That each new model will unlock entirely new markets, wipe out incumbents, and justify ever-increasing infrastructure costs.

But if LLM progress stalls, or even just slow, the maths changes. Fast.

Open AI is spending money so fast that they are forecast to blow through over $1 trillion in the next ten years!! That is $1,000,000,000,000.

Just so you can get your head around this, if I spent one dollar every second, it would take me 31,688 years to spend $1 trillion.

And they're spending that whilst making bigger losses than any company ever! How about a $78 BILLION loss forecast for 2028?

This only makes sense if we are at the foothills of enormous progress. If we are closer to the peak than anyone cares to admit, then Open AI will be going bust well before then.

Suddenly:

Training a $2B model doesn’t make sense if it only improves outputs by 5%.

Users won’t switch to your app just because you’re using GPT-7 instead of GPT-5.

Monetisation gets harder because you’re no longer selling “smarter”, you’re selling “different.”

And worse: the cost of building LLMs doesn’t plateau. It keeps going up. The infrastructure required to train and serve these models is massive and rising.

If the models don’t keep getting noticeably better, you’re left with an expensive, undifferentiated commodity. That’s not a great business to be in.

The implication? Some of these bets might not just underperform, they might collapse.

But a Plateau Isn’t Failure

None of this is a doomsday prediction, even though it sounds like it. It’s just a different lens.

LLMs don’t need to get exponentially better forever to be useful. They’re already useful. But we may need to shift our expectations.

Think about it:

Search engines got dramatically better from 1998 to 2008. They’ve mostly plateaued since.

Smartphones evolved rapidly in the 2000s. But the iPhone 17 isn’t radically better than the iPhone X.

That’s normal. That’s maturity.

If LLMs are entering that phase, the challenge isn’t to make them “smarter.” It’s to make them useful in more places.

That means:

Better interfaces

Stronger workflows

Real memory

Contextual awareness

Integrated tools

Personalisation that actually works

These are experience upgrades. Not model upgrades. And they might matter more.

What Comes Next?

So, are we at the peak?

It’s hard to say.

We might still see breakthroughs in multi-modal reasoning. Or in memory. Or in new model architectures. But we also might just be tinkering around the edges for a while.

And if that’s the case, the smart move isn’t to wait for GPT-6. It’s to build with what we have now.

Because even if this is the worst it will ever be, it’s still pretty damn good.

Takeaways

Software doesn’t improve forever.

LLMs are software, and they’re made by humans. Human progress is messy and non-linear.

The data ceiling is real.

We’ve already scraped most of the high-quality internet. Future models won’t have exponentially more to learn from.

Business bets are risky.

Companies pouring billions into model training are assuming exponential upside. If that slows or stalls, the economics get shaky.

Plateaus mean opportunity.

This is when the best products get built. Not when models leap forward but when they stabilise.

Shift your mindset.

Don’t just think about what LLMs might be able to do tomorrow. Focus on what they can do well right now.

If you’re building in this space, don’t get distracted by the hype cycle.

Assume improvement will slow. Assume the models won’t save you. Build something useful anyway.

And if I’m wrong, if GPT-6 makes GPT-4 look like a potato? Great. You’ll be ready for that too.

Until next time,

Jon "Doomsday" Kaye

Choosing to live more deliberately

P.S.

We've just launched an app for Do Radio! Not only does it make listening to the show easier, you now have access to all of our "old" shows that you may have missed when they were first broadcast. Download the app here.