Here’s my (slightly tardy) review of artificial intelligence at the end of 2025.

The Economic Bubble

In 2025 AI was in the midst of an immense economic bubble. That is so obvious a statement that it barely deserves to be said. The year began with the launch of DeepSeek, which subsequently triggered a selloff of NVIDIA stock leading to a fast 17% drop in price. Why exactly did the launch of a new LLM cause such a selloff? Because the DeepSeek model didn’t use enough GPU compute given its performance, and investors became scared that other companies might not need to buy as many of NVIDIA’s chips. Despite this absurd blip, NVIDIA’s stock price did quickly recover. I admit that despite my low-risk financial temperament, I saw the opportunity and did ‘buy the dip’ on NVIDIA stock.

And by end of year, NVIDIA had achieved a market value of 5 Trillion dollars (yes, that’s a ‘T’). That value is greater than the GDP of every single nation on earth except China and the United States.

I’m no economist, and neither is the journalist Ed Zitron, but he’s done amazing work on this subject and I highly recommend it. He’s been saying the quiet part out loud on behalf of the entire tech industry.

Large Language Models

I recently re-watched an episode of the Linus Tech Tips WAN Show, from December 2, 2022. This was the first episode after the initial launch of Chat GPT. The entire two and a half hour episode was dominated by conversation of this new chat bot from Open AI and its capabilities. Linus and Luke were floored, and could not stop gushing about all the amazing and human-like things it could do. They were able to go on text adventures with it, they could write articles, emails, create code for complete and non-trivial programs.

My takeaway from listening to this episode again (I remember listening to the original airing three years ago) was that LLM’s have not improved as much as I was assuming. Quite frankly, I don’t see LLM’s as having achieved any fundamentally new capabilities since the original launch of ChatGPT. Everything that anyone would care to do with an one was all there in 2022. Obviously as the models scaled their performance in these categories improved, but that’s not what the marketing material has been saying, and that’s certainly less than what you would expect given how much money is being spent on this technology.

There are some who predict that LLM-based AI is a stepping stone to AGI or ASI. I was never wholly convinced either way, but as of now I am nearly certain that it won’t happen. I don’t believe any amount of scaling could achieve that, but even if it were possible the physical limitations of this planet would prevent it. It is not possible to continue growing these models by yet more Orders of Magnitude when they are already consuming a significant faction of humanities wealth, compute, and energy.

Image Generation

My opinion is that progress in image generating AI has been more pronounced than text. The generated images of Will Smith eating spaghetti have demonstrated this point well. In years past it would have been possible to detect a generated image by seeing defects, such as too many or too few fingers rendered on hands. By the end of 2025 nearly all of these defects have been resolved entirely, or have become so subtle as to be nearly impossible to notice.

Some people are becoming increasingly concerned with the negative impact that this technology will have, myself included. It is now a foregone conclusion that this will replace clip art wherever it is used. It will be used heavily in advertising and marketing, likely to the point that human-made content becomes noteworthy or even impressive.

I think the interesting question is what cultural norms will develop around this technology. Will people be largely accepting? Will people fill the theatres to enjoy generated movies? Will generated music replace the standard rotation of pop songs? I think about this a lot, and ultimately I feel it is too soon to know for sure what place this technology will have in our culture.

Unfortunately, I do believe there are reasons to be pessimistic. There’s evidence now that generated music can do well on platforms like Spotify. I’m seeing the generated videos on Youtube, with seemingly high view counts. I see the generated avatars on social media profiles. This year for the first time I even saw what looked like generated art used in the physical Christmas decorations in my town.

But I also see some push back. I’ve seen it in gaming. It is, for now at least, considered controversial to have used AI in the game development process. It’s examples like this that give me some hope that generated content won’t become normalised.

What I think

If it isn’t already obvious, I personally don’t like ai-generated content. I don’t like the way it looks, and I’m deeply concerned about the consequences of its widespread use. Predictions of the future used to say that all physical labour would eventual be automated, and that humanity would be free to pursue creative expression to an extent never before possible. I do believe that human physical labour will eventually become obsolete, and I feel it would be both tragic and ironic if we chose to allow human creativity to become obsolete as well.