Why I Prioritise Questions Over Predictive Claims About AI Capabilities

We’re in an AI frenzy, and amidst the noise, I’ve occasionally fallen into a predictive trap. It’s easy to look at model capabilities today and given previous trends, you extrapolate enthusiastically. Likewise, when existing AI techniques appear to plateau, pessimism takes hold and we undershoot our expectations.

I’ve been guilty on both occasions. When I used GPT-3 for the first time in 2020, my expectations of rapid progress were so high that I imagined we would have significantly more advanced systems by now. In equal measure, I thought it would take longer for video models as good as Veo3 to emerge. Curiously, AI experts aren’t immune to these erroneous forecasts either.

Enthusiastic Extrapolations

“We should stop training radiologists now. It’s just completely obvious that within five years, deep learning is going to be better than radiologists.”

That’s a quote from Geoff Hinton, the “Godfather of AI” and pioneer of the machine learning techniques that power many of the exciting AI products today.

Hinton made that statement in 2016. It’s now almost a decade later. Are we still training radiologists? Absolutely. Demand for radiologists has never been greater and the world will continue to train them for the foreseeable future, as this NY Times article highlights.

Here’s another prediction, this time from a technologist who’s not only running one of the largest AI labs today and co-founded OpenAI before that, but is also the founder of two massively impactful, hard-tech companies: Tesla and SpaceX.

In 2015, Elon Musk predicted that “Tesla will have a car that can do full autonomy in about three years, maybe a bit sooner.” That didn’t happen, and we’re still some way off (though others argue otherwise).

Now I’m sure that self-driving cars will be part of the future. I certainly enjoyed trying out Waymo earlier this year in San Francisco. However, there’s a lot left to figure out. There are technical aspects that are still problematic, as well as social and legislative infrastructure that would pave the way for autonomous vehicles to be deployed widely.

Cynical Outlooks

Notice also that experts don’t just overestimate AI capabilities in their forecasts; they also underestimate them.

Consider the challenge of predicting a protein’s 3D structure from its amino acid sequence. This process had stumped biochemists for 50 years, and the widely held view was that a solution was far off into the future.

However, in 2020, DeepMind’s AlphaFold2 stunned the field by achieving near-laboratory accuracy. This prompted John Moult, a leading biologist and co-founder of the international benchmark for potent structure prediction, to declare that the decades-old challenge had effectively been solved. The timeline for this achievement surprised the world.

LLMs and genAI technologies have surprised, too. For instance, many experts thought that accurate human speech recognition was “A.I.-hard”. In other words, it would take several years and incredibly powerful AI – perhaps artificial general intelligence itself – to solve.

This assumption collapsed in 2022. In that year, OpenAI open-sourced Whisper, a relatively light-weight speech recognition model that could understand over 90 languages. This model, much like humans, can navigate the complexities of everyday speech such as accents, interruptions, ambiguous phrasing, and context-dependent meaning. 

"Possibility Forecasting" Through Questions

AI capabilities are notoriously difficult to predict. And if experts find it challenging, then generalists will find the process equally, if not more, elusive. As someone who works in venture capital as a generalist technology investor (rather than a AI research specialist), my approach here is to lead with questions and inquiry rather than fixed expectations or premature claims about where AI will go.

What are those questions? Here are a few I've pondered. I'm continuously revising what to ask and have taken a preference to open-ended questions rather than precise predictions.

• What new obstacles might hinder progress? The Apollo program is a good case study for exploring such a question. The USA sent humans to the moon 56 years ago. Why aren't humans on Mars by now if the Apollo program achieved its goal in 11 years with a budget of $25bn (an inflation-adjusted amount of $230bn in 2025)?
• What areas of AI hold the potential for significant surprise? We thought speech recognition would require AGI but it’s now commoditised. Are there other hard areas that could surprise us? For example, what if robotics had a ChatGPT moment in the next 2-3 years?
• What dead ends in AI are flying under the radar? Does the transformer architecture at the heart of LLMs have fundamental flaws that no amount of research and commercial funding could ever overcome? What happens when we hit fundamental limits we can't hack our way around? Where might the talent and funding be diverted, assuming AI interest doesn't wane?
• Where is our lack of imagination holding us back in terms of what’s possible with AI? Despite all the rapid progress in genAI, are we under-utilising what we already have, for example with multimodal capabilities?