Size Doesn't Always Matter—What Does Strawberry, the latest AI development, signify?

With the release of Open AI's o1—Strawberry—where is AI currently headed?

Making models bigger is not the only option

Open AI’s newly released AI, Strawberry, reflects the idea advanced by some in some contexts that ‘size does not matter’. The last couple of years have seen an explosion in accessible AI, largely based on building bigger models. However, some commentators have been saying that there may be diminishing returns from just producing larger and larger models. While the jury is still on that point, what we see with Strawberry AI is that companies are increasingly exploring other avenues to advance AI beyond just making them bigger.

Regardless of all its obvious abilities, recent AI has been criticized for sometimes making mistakes. This is obviously a problem if you want to use AI in real-world situations where you can’t afford to have it providing unreliable answers. The approach Open AI has taken with Strawberry is not just to create bigger AI but to basically tell the AI we already have it to slow down and work step by step. This builds on approaches others have been recently taking to improve AI’s performance.

‘Suzie just needs to slow down and concentrate more ’

It’s a bit like going to a parent-teachers evening and the teacher saying: ‘Suzie is certainly great at doing sums, but she needs to slow down a bit and focus on accuracy by taking each problem step by step’. OpenAI claims that using a step by step process, Strawberry ‘performs similarly to PhD students on challenging benchmark tasks in physics, chemistry, and biology’. It will be interesting to see if it lives up to this claim as others widely use and test it.

We are watching AI progressively ‘upskill’

The development of Strawberry highlights a point I make in my new book Surfing AI. We should look at AI’s development as a matter of it systematically ‘upskilling’ over time. Just as we would with a human as they mature and proceed through their education. Strawberry reflects another step in AI’s journey along the same type of route.

All of this points to the fact that AI is continuing to progressively get smarter over time in various ways. It’s cool to watch Strawberry tell you the steps it is taking at it goes about answering a question you have given it. One way of understanding what has been going on with AI so far is that it has just been blurting out an answer without having had the chance, as a human does, to formulate its response and critique its own thinking in a step-by-step manner. Strawberry now uses this approach which is something others in AI have been working on for some time now. The end result is a more coherent anwser.

I am finding Strawberry impressive

I have been playing around a bit with Strawberry and I am impressed. This does not mean that it will not make any mistakes, just that it is just another step in the process of AI developing. But it also means that it is a mistake to think that progress in AI only depends on making models larger and larger—something that is not currently yielding as much progress as it was in the past. I got it to do various tasks that I usually use to test new AI models and I feel it performed well. It is also doing well in regard to some of the benchmarks that companies test new AI models on.

The importance of being Ernest

One of my favorite tests is to get a newly released AI to write in the style of some of the authors I have read. Here is Strawberry giving a rendition of Hemingway talking about going hiking. I have tried this same test on versions of ChatGPT and other AI chatbots. I think that this is getting pretty good at capturing Hemiingway’s voice.

We left the city behind, the hum of its streets fading with each step. The trail to the Orongorongo Valley was narrow, winding through dense bush. The air was crisp, and the scent of damp earth rose around us. Birds called from the treetops, their songs sharp against the silence.

The river appeared beside us, a ribbon of clear water over smooth stones. We followed it deeper into the valley. The mountains pressed in, their slopes heavy with ancient trees. Sunlight filtered through the canopy, casting patterns on the forest floor. We walked on, the weight of our packs settling into our shoulders.

At dusk, we made camp by the riverbank. The fire crackled, sending sparks into the cooling air. We sat quiet, watching the light fade from the sky. The stars came out, bright and indifferent. The valley held its peace, and we found ourselves a small part of its enduring silence.

AI is moving fast through the hype cycle

Some are saying that there is an AI bubble. While it may be the case that there has been over-investment in AI by some looking for short-term profits, from a technical point of view, AI is still progressively advancing and I don’t see this stopping any time soon. There is a hype cycle in technology with overblown expectations at the start and then crashing expectations as it does not deliver as fast as people thought, followed by it actually delivering a lot of value in the longer term. What I think is happening with AI is that it is just moving through the hype cycle really fast rather than taking a decade or so.

As I said at the start of this piece, if any one pathway toward AI’s advancement starts to slow down, AI researchers will simply pursue other pathways. The advances being made by making models bigger may be slowing down—but we are not yet certain that this has run its course. Regardless, researchers are now turning to another strategy—getting AI to work in a systematic step-by-step manner—in order to improve its performance.

People should be trying AI out

We have only just started to scratch the surface in terms of the number of people using AI and the ways in which it can potentially be used. I would suggest that this is a great moment for people who have not yet started to try out AI to jump in the pool and have a go. It’s important to realize that, like anything, it can take a while to work out how to use AI in the most productive way. It has been said that it takes people about ten to fifteen hours of using systems such as ChatGPT to get to understand how they can be used effectively. It’s important that lots of people try out AI at the moment, not only because they may be able to use it in their work or home life. It is also important so that we all can get a better idea of its potential and how society should manage the introduction of AI to advance prosocial AI use.”

Dr Paul Duignan is a psychologist and AI strategist who has written a new book on AI strategy and its psychological, social, business, government, and civil society impacts—Surfing AI: 30 Fresh Terms and Smarter Ways of Talking About Artificial Intelligence.

Get in touch if you want a strategy workshop or talk on where AI is taking us all.

Some relevant links for this article

https://www.wired.com/story/openai-o1-strawberry-problem-reasoning/

https://futurism.com/openai-released-strawberry-o1-preview-model

https://openai.com/index/introducing-openai-o1-preview/

Image credit: Wirestock. https://www.freepik.com/free-photo/closeup-shot-fresh-ripe-strawberries-isolated-white-surface_20949014.htm