We are at the tail end of the first wave of LLM-based AI, which performs better than the average human in some tasks but is not enough and is expensive to train
On a recent cross-country trip—I (Paul) drove from California to Illinois and back again. On the drive, I saw a sign …
Irregular Ideas
Context & Ripple Effects
This is an early diagnosis of the capability-and-cost gap in LLM deployment: models can clear a meaningful performance bar on selected tasks without yet being broadly sufficient or cheap to build.
LLM developers and prospective buyers must distinguish demonstrations of above-average task performance from workloads where reliability and training cost justify deployment.
Near-term adoption concentrates around bounded, repeatable tasks rather than treating a general-purpose model as a complete substitute for human judgment.
Companies integrating LLMs are pushed toward workflow design, oversight, and task selection as much as raw model capability, making useful-task cost a key competitive measure.
Third-order effects
If capability gains and costs remain uneven, AI competition will reward firms that turn models into dependable products and workflows rather than those with scale alone.
The longer-run market may segment between high-cost frontier training and lower-risk, task-specific deployment; the pace of that split depends on whether reliability improves faster than operating costs.
The trend: This is one early marker of AI industrialization: value shifts from general model promise toward economically viable, narrowly reliable work.
Piercing the AI hype, Paul with an interesting take that says: “We are already at the tail end of the current wave of AI. We are bumping against many of its limits...”
Some papers mentioned: • Population Aging and Economic Growth https://www.nber.org/... • Automation and New Tasks: How Technology Displaces and Reinstates Labor https://www.aeaweb.org/... • Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks https://arxiv.org/...
As a postscript, while we say and believe this, we are also actively investing in our venture partnership around this thesis. We can be accused of talking our book, but we think we are mission-driven with thoughtful ideas about technology and the potential for human flourishing.
The piece by @defrag and I for our SK Ventures is up. It's called “AI isn't Good Enough”. It is on structural imbalances in the US workforce, demography, the risks of so-so automation, and the potential for human flourishing. You can read it here https://skventures.substack.com/ …
We go into some current developments in AI, like the rise of RAG (retrieval augmented generation), which is timely in the context of OpenAI's latest moves. And, for econogeeks, dive into some automation models from Acemoglu and Restrepo, before making a Led Zeppelin Zoso claim.
We argue in the piece that AI needs to be much better or, weirdly, much worse, to avoid the so-so automation trap, where we have massive economic dislocation without compensating productivity gains. We also map out our views of what the next decade could look like.