AI is great at helping to summarise information. But it does a terrible job of thinking for you. Be careful you don’t accidentally ask it to.

Before sharing my last blog post I ran it through ChatGPT, I wanted to know what AI thought of my opinion piece, particularly since it claimed a future direction for ChatGPT’s parent company. The result was unsurprising.

ChatGPT complained that the application of the experience of AI in software development to other industries was too broad. However, the technique of applying the learning from one industry to another is pretty well known. It is often referred to as “cross-industry learning” or “transferable learning”. In industrial research we use the term “transfer innovation”, to apply an innovation in one industry into another. This was a blind spot for ChatGPT.

ChatGPT suggested that having an AI model on silicon would mean that it wouldn’t be a frontier model. This is actually true, and the article points this out. However, the article points out that this does not put the user at a disadvantage, rather it rightsizes the cost of running the model with the anticipated gain it provides.

The AI also argued with me that there was more to creating software with an AI than just the model. This is of course also true - there are agents and tooling. But the costs of agents and tooling is much less than the cost of token consumption

In short the AI missed the point I was making - that the cost pressures and value offering of AI is a balancing act. Having a solution which delivers great value at an affordable price point does not require it to be the latest and greatest model. This opens up the avenue for having model specific chips without the costs associated with the cloud offerings. These were so power efficient that they could run at the edge or in embedded products, and deliver “good-enough” results for real productivity gains.

Why did the AI tool disagree?

This is a good question. I have a hunch. AI agents are trained on what they can find online, often repeated data or facts will take precedence. New views are necessary new, and hence not shared often if at all online. The AI agent will not be aware of the data which underpins the argument. It can only compare the argument against others it has seen before.

The WIRED magazine article “AI Just Isn’t Right” looked at how reliable AI was at fact checking articles, spoiler alert - it wasn’t. In the article the Author, Meghan Herbst explains that most of the information in the world is not online, it is in physical media, books, newspapers, film, photographs, CDs, DVDs, etc. Further more, she pointed out that the really cutting edge information isn’t online, it’s in people’s heads and requires a phone call and a conversation.

Even if an idea has previously been shared online and was discounted at the time, it doesn’t mean that the same idea should be discounted today. Ideas and concepts exist within the current state of the world and if the state of the world changes, then the value of the ideas also changes. A great example of this is AI itself. In the 1980-1990’s AI experienced the “AI winter”, where this topic was not invested in. It was deemed too expensive to explore, that you would need a lot of AI specific accelerators to make Neural Networks function anywhere near a reasonable speed. Look where we are now.

Evaluating my Blog Post

Since my previous blog post was published both OpenAI and Anthropic appear to have plans for their own model specific chips. OpenAI announced their chip in partnership with Broadcom and Anthropic are reported to be in talks with Samsung to release their own. The Chinese model provider DeepSeek are also looking to produce their own chip as well, just this morning there was a news story all about that.

How these chips will be packaged and made available will determine how accurate my prediction was. But so far, I’m kinda happy. I wish I was that good at predicting the current FIFA 2026 World Cup Results.

Remember

Use AI to help summarise information. But be careful it doesn’t try to draw conclusions for you. That’s your job….. not AI’s.

If you want to go deeper into how to peer into the future, I’ve a small, short podcast series all about that, Search for Industrial Research on your favourite player.