GPT-5 and the Motte-and-Bailey Problem

GPT-5 was by all accounts a rocky launch and left many wanting more. In the two weeks since release, after some heavy use, conversations with peers (investors, operators, developers, founders, writers) we’ve seen little consensus on how it changes daily work or life.

Looking online, there’s arguably even less consensus. Opinions range from ‘this is the best thing ever and will replace every developer,’ to ‘this is worse than GPT-4o, about as useless as a toaster’ and of course, the now-infamous chart mistake during the official release video. Whatever your opinion of GPT-5, it opens up an interesting dialogue about managing expectations of new technology and the pace at which AI is improving.

The infamous GPT-5 launch chart is misleading, showing 52.8% taller than 69.1%

Now, one wouldn’t blame you for thinking GPT-5 would be revolutionary. Sam Altman did say it would be ‘smarter than any human’ and even called it ‘the Manhattan project of AI.’ High praise, of course, that echoes his earlier declaration nearly a year ago that artificial ‘superintelligence’ was ‘just around the corner.’ Yet after GPT-5’s release, Altman shifted gears: he has not backtracked on AGI outright, but now says the term ‘AGI’ is ‘not a super useful term.’

Critics frame this as a classic motte-and-bailey fallacy: promote a bold, controversial claim (the ‘bailey’) to inspire excitement and investment, then retreat to a safer, more defensible position (the ‘motte’) when pressed. In this case, the bailey was that GPT-5 would mark a historic leap toward human-level intelligence. The motte became a vague point, that ‘AGI’ is just an unhelpful label. Whether you see that as strategic framing or rhetorical backpedaling, it is notable because the real progress of GPT-5 may not be in intelligence, but in efficiency, bringing down costs and making AI more usable and commercialized.

When the GPT-5 launch did not deliver on this revolutionary (and admittedly, unrealistic) promise, the argument retreated to the "motte." Altman's new, safer position is that AGI is "not a super useful term," shifting the focus away from a singular, human-level intelligence and toward a more incremental, practical definition of AI's value. This new position, that AI is simply an evolving, useful tool, is far harder to dispute and serves to temper expectations without undermining the company's long-term vision. This tactic allows the company to benefit from the initial hype of a dramatic breakthrough while having a safe, non-controversial position to fall back on when faced with scrutiny.

In reality, the greater discussion here isn’t whether or not OpenAI overpromised, under delivered, or fell somewhere in between but the wider discourse is where do we stand with LLM advancement? Is AGI still the north star? Or is this all just a big overreaction to what was, by all accounts, a rocky launch of a still great product.

To draw a parallel, there was a time when self-driving cars shifted from ‘they’ll be here soon’ to ‘why aren’t they here yet?’ It feels like we are approaching the same plateau with generative AI models. It raises the question: did people once think the printing press was a failure because there was not a vast library in every 15th-century home? The real transformation unfolded slowly, over decades. We may be seeing the same with LLMs. The hype cycle grabs headlines, and writers who pride themselves on being at the forefront of tech and startups are just as guilty of feeding into it. Yet the true impact of these changes often comes gradually, in ways that are less dramatic but more enduring.

Now the rate of change is interesting because there are very real bottlenecks to growth. At this point, the story with LLMs may be less about whether GPT-5 underwhelmed and more about the hard constraints that shape its trajectory. Compute is the most obvious one: there are simply not enough bare-metal GPUs to meet global demand, and even the most well-capitalized labs are rationing resources across training runs and inference. Without more and better GPUs, models will hit performance ceilings not because of a lack of ideas but because of infrastructure scarcity.

But compute is the first domino. A topic we write often about, energy demand is scaling just as fast, with AI data centers drawing as much power as small cities and forcing utilities to rethink grid capacity. Supply chains are another chokepoint, with advanced chip manufacturing concentrated in a handful of foundries, geopolitical fragility that no amount of software cleverness can offset. And even when the hardware and power exist, adoption itself becomes a bottleneck: regulation, cost, and workflow integration all move more slowly than model releases.

Framed this way, AI’s limiting factors look less like questions of intelligence and more like questions of scale. The frontier isn’t just smarter models, it’s the ability to overcome the structural constraints of compute, energy, supply chains, and adoption. Until those bottlenecks are resolved, every new model launch will feel a bit like déjà vu: impressive, incremental, but still defined by the hard edges of reality.

What often gets lost in the debate over intelligence versus scale is the reality of economic adoption. Most enterprises aren’t asking “is this AGI?” but instead “does this save us money or generate revenue today?” For many, the unit economics don’t yet work, inference costs remain stubbornly high, and thin SaaS margins for wrapper apps prove hard to defend. The startups that survive will likely be the ones that are AI native and deep: embedding AI into workflows, integrating with proprietary data, and solving industry-specific problems where switching costs are high. That is why many investors see more promise in infrastructure, tooling, or vertical plays where defensibility is stronger. The bottleneck isn’t imagination, it’s proving ROI at scale.

Beyond compute and capital, there’s also a talent bottleneck. The world’s best researchers and engineers are locked inside a handful of frontier labs, bound by NDAs, lawsuits, and limited mobility. Academic labs can’t compete: they don’t have access to the GPUs or the closed-weights models that power breakthrough research. That concentration of talent slows diffusion, creating a dynamic where innovation is gated not just by chips but by people. Historically, open ecosystems, think the early internet, drove compounding innovation. In AI, the walls are higher, and the brain drain into a few companies risks narrowing the field of ideas.

Then there’s regulation and geopolitics, which layer on new friction. The EU is moving forward with the AI Act, the U.S. is leaning on voluntary pledges, and China has imposed algorithmic oversight, creating a fragmented global landscape. At the same time, compute itself has become geopolitical. The U.S. is restricting GPU exports to China, Europe is scrambling for “sovereign compute,” and chip manufacturing remains concentrated in Taiwan, a single geopolitical flashpoint. The result is a race dynamic: nations see AI as strategic infrastructure, and companies have to navigate both compliance drag and great-power competition. Bottlenecks aren’t just technical, they’re national.

If this feels familiar, it should. History shows us that transformational technologies rarely follow a straight line. Electricity didn’t change the world overnight, factories had to be rewired, cities had to modernize their grids, and adoption played out over decades. The internet’s early years were filled with hype, false starts, and a crash before the real durable businesses emerged. AI may be replaying the same script. Progress looks messy and incremental in the moment, but over time, these bottlenecks resolve, infrastructure catches up, and what once felt underwhelming becomes woven into everything.

Which is why the real question isn’t whether GPT-5 wowed or disappointed, but whether we’re willing to zoom out. AGI may or may not be the north star. The frontier may not be model size but infrastructure scale. And the true transformation may not be one big leap, but the slow, compounding integration of AI into every layer of the economy.