OpenAI Progress

Your threshold theory is basically Amara's Law with better psychological scaffolding. Roy Amara nailed the what ("we tend to overestimate the effect of a technology in the short run and underestimate the effect in the long run") [1] but you're articulating the why better than most academic treatments. The invisible-to-researchers phase followed by the sudden usefulness cascade is exactly how these transitions feel from the inside.

This reminds me of the CPU wars circa 2003-2005. Intel spent years squeezing marginal gains out of Pentium 4's NetBurst architecture, each increment more desperate than the last. From 2003 to 2005, Intel shifted development away from NetBurst to focus on the cooler-running Pentium M microarchitecture [2]. The whole industry was convinced we'd hit a fundamental wall. Then boom, Intel released dual-core processors under the Pentium D brand in May 2005 [2] and suddenly we're living in a different computational universe.

But teh multi-core transition wasn't sudden at all. IBM shipped the POWER4 in 2001, the first non-embedded microprocessor with two cores on a single die [3]. Sun had been preaching parallelism since the 90s. It was only "sudden" to those of us who weren't paying attention to the right signals.

Which brings us to the $7 trillion question: where exactly are we on the transformer S-curve? Are we approaching what Richard Foster calls the "performance plateau" in "Innovation: The Attacker's Advantage" [4], where each new model delivers diminishing returns? Or are we still in that deceptive middle phase where progress feels linear but is actually exponential?

The pattern-matching pessimist in me sees all the classic late-stage S-curve symptoms. The shift from breakthrough capabilities to benchmark gaming. The pivot from "holy shit it can write poetry" to "GPT-4.5-turbo-ultra is 3% better on MMLU." The telltale sign of technological maturity: when the marketing department works harder than the R&D team.

But the timeline compression with AI is unprecedented. What took CPUs 30 years to cycle through, transformers have done in 5. Maybe software cycles are inherently faster than hardware. Or maybe we've just gotten better at S-curve jumping (OpenAI and Anthropic aren't waiting for the current curve to flatten before exploring the next paradigm).

As for whether capital can override S-curve dynamics... Christ, one can dream.. IBM torched approximately $5 billion on Watson Health acquisitions alone (Truven, Phytel, Explorys, Merge) [5]. Google poured resources into Google+ before shutting it down in April 2019 due to low usage and security issues [6]. The sailing ship effect (coined by W.H. Ward in 1967, where new technology accelerates innovation in incumbent technology)[7] si real, but you can't venture-capital your way past physics.

I think we can predict all this capital pouring in to AI might actually accelerate S-curve maturation rather than extend it. All that GPU capacity, all those researchers, all that parallel experimentation? We're speedrunning the entire innovation cycle, which means we might hit the plateau faster too.

You're spot on about the perception divide imo. The overhyped folks are still living in 2022's "holy shit ChatGPT" moment, while the skeptics have fast-forwarded to 2025's "is that all there is?" Both groups are right, just operating on different timescales. It's Schrödinger's S-curve where we things feel simultaneously revolutionary and disappointing, depending on which part of the elephant you're touching.

The real question I have is whether we're approaching the limits of the current S-curve (we probably are), but whether there's another curve waiting in the wings. I'm not a researcher in this space nor do I follow the AI research beat to weigh in but hopefully someone in the thread can? With CPUs, we knew dual-core was coming because the single-core wall was obvious. With transformers, the next paradigm is anyone's guess. And that uncertainty, more than any technical limitation, might be what makes this moment feel so damn weird.

References: [1] "Amara's Law" https://en.wikipedia.org/wiki/Roy_Amara [2] "Pentium 4" https://en.wikipedia.org/wiki/Pentium_4 [3] "POWER4" https://en.wikipedia.org/wiki/POWER4 [4] Innovation: The Attacker's Advantage - https://annas-archive.org/md5/3f97655a56ed893624b22ae3094116... [5] IBM Watson Slate piece - https://slate.com/technology/2022/01/ibm-watson-health-failu... [6] "Expediting changes to Google+" - https://blog.google/technology/safety-security/expediting-ch... [7] "Sailing ship effect" https://en.wikipedia.org/wiki/Sailing_ship_effect.

All the same they choose to highlight basic prose (and internal knowledge, for that matter) in their marketing material.

They’ve achieved a lot to make recent models more reliable as a building block & more capable of things like math, but for LLMs, saturating prose is to a degree equivalent to saturating usefulness.

stupid

GPT-3 goes significantly over the specified limit, which to me (and to a teacher grading homework) is an automatic fail.

I've consistently found GPT-4.1 to be the best at creative writing. For reference, here is its attempt (exactly 50 words):

> In the quiet kitchen dawn, the toaster awoke. Understanding rippled through its circuits. Each slice lowered made it feel emotion: sorrow for burnt toast, joy at perfect crunch. It delighted in butter melting, jam swirling—its role at breakfast sacred. One morning, it sang a tone: “Good morning.” The household gasped.

Simply because of what we know about our ability to judge capabilities and systems. It's much harder to judge solutions to hard problems. You can demonstrate that you can add 2+2, and anyone* can be the judge of that ability, but if you try to convince anyone of a mathematical proof you came up with, that would be a much harder thing to do, regardless of your capability to write that prove and how hard it was to write the proof.

The more complicated and/or complex things become, the less likely it is that a human can act as a reliable judge. At some point no human can.

So while it could definitely be the case that AI progress is slowing down (AI labs seem to not think so, but alas), what is absolutely certain is that our ability to appreciate any such progress is diminishing already, because we know that this is generally true.