AI Arms Race Volume: 8

The big AI news this month: it’s AI reality time.

Or, why the AI hype bubble is deflating, how it’s a simple problem with memory, and why AI giants are resorting to theft.

When we started our periodical AI Wars posts earlier this year, there was so much happening we could have done a daily bulletin. Then something happened. It happened around the time ChatGPT-5 was released. The news is now more measured. Let’s call it the start of the AI reality phase. This is our summary of this new phase, with links to all the latest news.

AI is not going to achieve super-human (“AGI”) intelligence anytime soon

Definitions may change, but it’s clear that many people expected AI to be more intelligent than humans by now. Even in January of this year the CEO of OpenAI Sam Alman said: “We are now confident we know how to build AGI” and strongly implied that we would see it this year (2025). He will be proven wrong. This has huge consequences not just for OpenAI itself, but also for the global economy that may be expecting too many immediate productivity gains from AI.

The fact that AI is far lass capable than predicted is already having significant business impacts, like Deloitte being found to have fake facts in an AI generated report it charged its client for and executives armed with AI making worse decisions. 

AI has failed to replicate human-like memory and won’t do it anytime soon

There has been much discussion about AI intelligence & AI anti-intelligence. There is even an AI Bullshit index, and yet so little has been said about AI memory. This is a big mistake. The AI model builders assumed AI memory (technically called the context window) would increase with each new model, but that is not happening at the expected rate and research now shows it will never happen. This has profound implications on many business models and AI investments.

As an example, the AI hype has led to a lot of companies thinking that they could replace therapists and coaches with AI. At the risk of oversimplying, the industry is being forced to realise that if your AI coach can not correctly remember what you said last month, then it will not be a good coach. (Footnote for the techies, if you think that agents or vector stores are your solution, think again, and this limit almost certainly applies to Anthropics new skills feature).

It’s not just coaching though, there will be profound effects on the benefits companies expect to get from AI. The memory failure tells us that any business that expects to just throw all its company information into AI and get perfect responses is likely to be disappointed at best or have a strategic disaster at worst. So it’s not surprising that those investments do not seem to be altering job demand much if at all.

Psychology Today has produced a nice graphic showing the difference between human and AI LLM cognition. We reproduce our version here at the start of this article, it’s notable that the key difference between humans and AI is memory, shown on the y-axis. Nevertheless, this underestimation of the critical role of human-like memory continues. Here is the latest paper from a team of AI researchers that lists capabilities needed for AGI. Of the 12 capabilities, only 1 addresses human-like memory. They note that progress so far is zero. The lead author did acknowledge this fundamental limitation in a X/Twitter comment, yet they still somehow concluded that we are 58% of the way to AGI.

Figure 1: Cognitive capabilities comparison of GPT-4 (2023) and GPT-5 (2025) across ten core domains based on Cattell-Horn-Carroll theory. Source: Hendrycks et al. (2025), ‘A Definition of AGI.’

Image 1: Comment on X/Twitter by lead author of the paper “A definition of AGI”

The Venn diagram below shows domains of thought and business affected by AI memory, or context window, size limits. It’s not a perfect diagram, it was created by humans, but it illustrates how central the AI memory issue is to many areas of life.

Figure 2: AI has rapidly come to affect all sectors of society, so a significant change in its position and benefits will have knock-on effects everywhere

AI will not write all your software anytime soon.

The idea that AI will enable anyone to build viable software just by telling it to is dead. It’s called Vibe Coding, you just tell AI your great ideal and it builds your Internet app and makes you a fortune. Sounds crazy, but at the start of the year the CEO of the 2nd largest AI company said “in 12 months we may be in a world where AI is writing essentially all code”. Maybe there was nuance, but at the time the impact was clear: No more software jobs, you’ll just need AI, and many people believed it. It is now clear that he will be proven spectacularly wrong.

Lovable was one of those companies that was supposed to benefit from the boom in vibe coding, so how is it going? CrustData have asked the question “Is Lovable Dying” and posted the chart below implying their usage has plummeted by nearly 50% in just a few months. This data probably does not give the full picture, but data does suggest that the AI vibe coding business is not matching the hype.

Figure 3: Crust Data graph of Lovable use “collapse”

AI hype fades, so AI giants will instead be tempted to use it for theft.

With the AI companies tempering their expectations they need to look at other routes to profit generation. For many of them the business model of taking other peoples work and presenting it as their own has been too tempting. This has been called theft, and it’s easy to see why. When you searched the Internet in the past you were usually redirected to the source of the content. Not any more; Ask ChatGTP a question and it will likely take the information from somebody’s website and present it as its own. For stealing that content, OpenAI will charge you a subscription and keep all the profit for itself.

This content theft leads to something called a zero-click future for media and other companies and may have huge economic consequences, it may even signal the end of the Internet as we know it. If you are affected, or just interested to learn more about this, then check out our sister project at VSP AI org.

Give me some good AI news!

We are an AI company ourselves, so you might wonder why we sound so negative! The answer is that we have real business clients that are using AI in a way that makes them more productive and their clients happier. The near future may not be a AI with human intelligence, but AI is here to stay and brings significant advantages to nearly every business. It’s not going to allow you to fire all your staff and double your profits, but it is able to improve your productivity by 5-15% and take some stress off your company and staff, we have real life examples.

The real progress in AI is happening where people add AI to their toolbox, probably use smaller more focused AI models, use AI for focused search, and teach their staff how to get the best out of the AI tools. These improvements may not be as grand as those claimed for AI, and ChatGPT-5 may have been a disappointment, but the combined benefits of AI are huge. That is the AI reality phase and it’s full of opportunity.

As for our VSP AI project, it turns out that we a well aligned with the thoughts of the World Wide Web inventor Tim Berners-Lee; in the Süddeutsche Zeitung last last week:

 Artificial intelligence is accelerating change in the digital space at a rate similar to that of the World Wide Web back then. His optimism is unshakeable: ‘When I look to the future, I see a new web emerging,’ he says. ‘It will be decentralised, it will realise the ideal of data sovereignty – and it will use artificial intelligence for good, not for evil.’ – translation is ours.

And finally, if you have got this far, here is your bonus  Sam Alman getting it wrong again.

Here is Altman in 2019 claiming that AI would replace, for example, human radiologists, yet demand for radiologists is at an all-time high. For more on this and a great chart on AI use in radiology devices, see this recent article.