Sam Altman says he'd look for a mega breakthrough similar to LLMs. What happened to AGI?
If LLMs were the precursor to AGI and we're just a couple of months away from it, then why do we need a new breakthrough?
The house of cards are seeemingly falling. With xAI apparently having a meltdown, Anthropic hiring more traditional developers after expecting to replace them all with AI by the end of last year, we’re noticing the AI hype slowly coming to terms with the reality. All you have to do is read between the lies and you’ll find small snippets of admission from CEOs.
In a recent interview of Sam Altman on Tree Hacks, he was asked the question -
“What AI subfield today feels like Open AI in 2016?”
“I bet there’s a new architecture to find that is as big as a gain as transformers were. I’d look for a mega-breakthrough and use the [current] models to help me”
On its own, this looks like a harmless prediction- technology has always had breakthroughs with previous learnings rocketing us into the future. However, this falls short of some of the claims of Sam Altman and others that we will achieve by end of the year or the next. But this doesn’t make sense. If LLM is the breakthrough, and we can achieve AGI, then why do we now need another technological breakthrough?
The Scaling Wall Nobody Wants to Talk About
For the past several years, the dominant religion of AI research has been scaling. The gospel was simple: more data, more compute, bigger models, and intelligence would emerge. This wasn’t entirely without merit. GPT-2 to GPT-3 to GPT-4 showed remarkable qualitative leaps, and the research community pointed to smooth, predictable scaling laws as proof that the path to AGI was just a matter of building bigger.
That narrative is quietly unraveling.
The gains from scaling are demonstrably flattening. OpenAI's own GPT-4 to GPT-4o improvements were largely about efficiency and multimodality, not raw reasoning leaps. The jump from GPT-4 to o1, and now to o3, required an entirely different paradigm - test-time compute, where the model is given more time to "think" through chain-of-thought reasoning before answering. That's a meaningful architectural shift, not a scaling increment. And crucially, it comes with enormous compute costs that scale poorly in production environments.
If brute-force scaling were sufficient, we wouldn't need that pivot.
Did Altmas finally admit to the limitations of current LLMs?
When Altman says he believes there’s “a new architecture to find as big as transformers,” he’s not making an optimistic prediction. He’s diagnosing a limitation.
The transformer architecture, for all its success, carries fundamental constraints. It processes context through attention mechanisms that scale quadratically with sequence length. Workable but not elegant for the kind of continuous, unbounded reasoning AGI would require. More critically, transformers are stateless between inference calls. They have no persistent memory, no ability to learn from new experiences without retraining, no model of the world that updates in real time. Every conversation starts from scratch.
What CEOs Say vs. What They Fund
Perhaps the most telling signal isn’t what AI leaders say in headlines, its where they quietly direct resources. Like I said above, the hiring of traditional developers by Anthropic, investing heavily on interpretability research are not the investments of an organization that believes it's twelve months from AGI.
The gap between the press release and the org chart is where the truth lives.
None of this means AI isn't consequential, transformative, or worth taking seriously. It clearly is. But there's a meaningful difference between a powerful tool that is reshaping industries and an artificial general intelligence that matches or exceeds human cognition across all domains. I think its time we all hit the brakes on the hype train and see AI for what it is.
If you found this interesting, I write about tech, AI, and the gap between hype and reality. Connect with me on LinkedIn/akshayramabhat.