Predictions for 2025

It’s been two years since I tried my hand at this, and looking back, I don’t think I was too off the mark, so I thought… why not try my hand at it again?

I expect 2025 to herald the start of a second AI winter, as both hype and somewhat misguided discussions about scaling laws converge into pointlessness.

The way I see it, there are two inescapable truths about AI:

• There is no more low-hanging fruit to be had either in terms of training data or model architecture.
• The only way to improve our current form of AI is to make both training and inference vastly more efficient, which is a much harder problem than simply throwing more data and compute at it–and paradigm shifts like transformers don’t come along every day.

These two things are going to come to a head in 2025, as a third factor comes into play:

• After over two years, we still haven’t found a killer application for AI that isn’t just a glorified recommendation or information retrieval system.

Or, rather, an application that consumers want. There are plenty of advertising, customer segmentation and profiling applications in the advertising and finance industries, but these are not things that consumers want. They are things that companies want, because they enable them to make more money either by increasing sales or by reducing costs.

(Which, incidentally, is where I see a lot of image generation being used – but I am much more worried about LLMs)

Alas, not all businesses are able to sit down and reason through exactly how they can improve their business through AI.

And no, we are not going to get AGI soon, even if o3 has gotten an impressive (and expensive) score. At least not something we can all agree is AGI, even though I suspect OpenAI will make a few bold claims to the contrary (o3 may be able to do multi-step reasoning, but it is not AGI).

We might eventually get something that is a lot cheaper and more flexible than what we have now, but it is not going to be AGI. Not unless there are fundamental changes in both training and inference (o3 is starting to explore that space, but it is still just a stepwise improvement).

And don’t get me started on “agentic” AI. The difficulty in coordinating context and goals across multiple models should be plain to anyone who’s ever had to manage humans, and yet all of a sudden putting several unfathomable chains of reasoning in little boxes and chaining them together is touted as “the future”.

I’m not even going to go there other than to say the hype-to-reality gap will, eventually, result in a cooling period for AI. My money’s on this year, but considering how long crypto hung around, the bubble might only deflate (with a whimper rather than a bang) in 2026.

Then there’s Ethical AI, which is going to become a much bigger deal in 2025 because we are now at a point where many companies are starting to use AI in ways that are not just creepy, but also potentially harmful–like employee monitoring, performance reviews, and hiring decisions.

I previously hinted at this and compared it to weapons dealing (“we just supplied a better gun, the customer pulled the trigger”), but the bottom line is that technology companies are toeing the line in all interesting sorts of legal liability that they are not really prepared for.

I fully expect 2025 to be the year where I refuse to work on a project due to ethical concerns, and I suspect I will not be the only one.

AI model providers are going to have to start taking responsibility for the models they deploy, and the companies that use them are going to have to start taking responsibility for the decisions they make based on those models.

I’ll start off with an easy one:

• Apple Intelligence, despite Apple’s clever marketing, will remain plainly distinguishable from useful AI.

But it will, unsurprisingly, improve – after all, it can’t get any worse, and I suspect Siri will progress through schizophrenia (“here’s what I found on the web” vs. “here’s what ChatGPT has to say”) to a state where it might actually understand context and chained intents – but I’m not holding my breath, just as I am not holding my breath for Apple to give us an iPad mini Pro or a touchscreen on a Mac.

• We will get a HomePod with a screen (and, of course, an unremovable power cable), but… who cares about that when there is so much to be fixed in macOS and iOS?
• There will be a Mac Studio update with a beefier M-series processor–but I don’t think the timings are quite right for it to be an M4 Ultra, and TSMC fabbing (while it remains in Taiwan) needs to be prioritized for best-selling, mass-market products.

But there is one more Apple prediction I can confidently put forward:

• Apple will still not allow you to run your own applications on iOS devices for more than a few days without the Developer Program or its kill switches.

• AMD will continue to gain market and mindshare in the PC market even as mini-PCs continue to rise to be the dominant form of desktop computing. What is left of the custom/gaming PC market will continue to shrink.
• Intel’s attempts at fighting back with discrete GPUs and will be off the mark as most people find out that mid-range machines using their increasingly confusing Core Ultra CPUs (or slightly beefier Ryzen equivalents) are just fine.
• NVIDIA will have another year of uncontested reign over the datacenter landscape, but they will start facing competition from ASICs even as the AI bubble begins to leak.
• The low-end maker/hobbyist market will continue to fragment (like it did during 2023 and 2024) with a few more ARM devices, but even more low-end Intel-based ones.
• RISC-V will continue to be irrelevant, although it will continue to compete with ARM for low-end, non-power or performance-critical scenarios.

• More and more businesses will publicly announce they are using/not using/not intending to use/revising their stance on (pick the one you want) AI, but there will be at least one major scandal involving AI ethics that will make the news.
• Every single form of customer support you can imagine will become shittier.
• The telco market (which is now undergoing fast compression down to the “dumb, cheap, unreliable pipe” point) will bottom out, with 5G confirming its status as an expensive licensing mess. Internet outages will become more common, and not just due to sabotage.
• The technology job market will continue to be a mess as companies continue to believe they can cut down on skilled IT workers by using AI, ignoring the fact like with Big Data and analytics, AI will require more knowledgeable people to maintain new solutions.
• Continuing my prediction from 2022 (which completely panned out, by the way), RTO mandates and the bias against remote work will continue to escalate.
• I’m still bullish on 4-day workweeks, although I’m less certain they will actually happen since the economy is going to go down the drain even further.

We’re going to have another Trump administration. That means another year of insane, unpredictable and (let’s face it) unbecoming announcements from the West’s foremost democratic state even as it teeters on the edge of remaining so.

• X will completely implode as political discourse alienates the last few thick-skinned users, and BlueSky will continue to grow simply because Meta dropped the ball on Threads in the EU.
• I expect a Taiwan incident of some sort to happen, although I hope it will be a minor one.
• Either that incident or the continuing hostilities in Ukraine (which are likely to last another year at least) will cause the economy to tank again, screwing up the markets to a fair degree.
• Given the way our civilization handled COVID and the rapid expansion of the new bird flu, I’m pretty sure it’s solely up to the virus whether we’ll have another pandemic or not.

And yeah, I was wrong about crypto. Let’s hope I’m wrong about most of these too.