A thread currently exploding on Hacker News is about something many in the AI community have been whispering about for a while, but few have said out loud: the euphoric trajectory we have witnessed in the LLM world over the past couple of years appears to be leveling off.

The linked article — published on wheresyoured.at — argues that the major capability breakthroughs in large language models are becoming rarer and more expensive to achieve. This is no longer a fringe opinion. It is a sentiment bubbling up from people working close to the field.

When 736 comments on HN agree that something smells off, it is worth paying attention.

What does "slowing down" actually mean in concrete terms? Is it that benchmark numbers no longer impress in the same way? That compute costs are escalating faster than performance gains? Commenters in the thread point to both — and add that the very "scaling law" mantra that has driven the investment wave is beginning to lose credibility in practice, even if no major player will admit it publicly.

This is closely tied to something else being hotly debated in other corners of the AI underground: we have built ourselves into an architectural dead end. Large autoregressive transformer models are still useful, but they are energy hogs, expensive to train, and carry structural limitations that do not disappear no matter how much data you pour in. The alternatives — neuromorphic networks, neuro-symbolic systems, Mixture-of-Experts, specialized small models like Phi-3 — are starting to look more interesting, not just academically, but commercially.

AI growth is slowing down — and the tech underground is feeling it - Bilde 1

It is not that the AI field is grinding to a halt. It is more that the definition of "progress" is in the process of changing. Training an even larger GPT clone is no longer automatically synonymous with a breakthrough. Investors and product teams are beginning to ask different questions: what is the actual utility per dollar, and are there narrower, faster, and cheaper architectures that do the job better for specific use cases?

These are still early signals from a community, not established consensus. But when an article about AI stagnation tops HN with hundreds of quality comments on a Tuesday morning in June 2026, it is a sign that the mood is beginning to shift — and that mainstream tech media will likely pick up this narrative within a matter of weeks.

Watch how the major labs communicate around their next model releases. The rhetoric there will say a great deal about whether they too can feel the brakes.