The charismatic prototype

Like with any prior technological breakthrough, there’s a downpour of AI applications at the early stage. Not surprisingly: it’s a magic technology bottled in an API that makes it so easy to tinker with. I believe the most interesting aspect of AI is that it unlocks new markets and use-cases for software; many others intuitively sense that too — hence the flood.

Most of AI applications and the teams behind them face the challenge of the charismatic prototype - a phrase Shyam Sankar (Palantir’s CTO) tossed in his recent interview:

One of the contradictions of the LLM revolution is it is so easy to build a charismatic prototype, yet it is so hard to get from a brittle prototype to something that’s valuable, and rock steady in production.

In other words, it’s so easy to see the promise of the differentiated, magical business value, yet delivering on that promise proves to be really difficult. The unique challenge with AI is that it appears to have human-like performance, but to offload real work to AI, you need predictability and failure-resilience—qualities we have come to expect from regular software. The human-like nature of AI greatly confuses our minds, similarly to how some charismatic people can dazzle us initially. Only a rigorous, upfront-prepared interview process (e.g., for a job) can reveal all the gaps that a person’s magnetism might cover up.

Similarly, the AI prototype needs to be scrutinized, tested, and coached so it graduates into a dependable product. In practice, this process requires a new software toolchain to be built for a particular problem domain.

I’ve been looking for the right metaphor for this gap between the appearance of a working product and an actual, value-generating system. I think Sankar’s “charismatic prototype” neatly nails it.

He has another great line in that interview:

Europe is looking kind of like a deer in the headlights. They’re trying to think their way through the AI problem. And if there’s anything about AI, it’s fundamentally experiential. You cannot think your way through it.

I’ll skip the Europe bashing of its meager performance in creating ambitious software companies and just focus on this deeper point he makes: AI is fundamentally experiential. I think this nails yet another non-obvious European cultural handicap: we tend to over-intellectualize our processes in Europe. This manifests in the continent having the AI Act before spawning any AI products (from which you could learn about real-world regulatory issues), calls for an “AI strategy”, endless panels discussing initiatives, conferences, and nothing of importance actually being done. When all strategy astronauts go home, they click a buy button on an American product that American companies built and adopted with the experiential mindset¹.

The “charismatic prototype” and “AI is experiential” are two great lines I will borrow from Shyam from now on.