So everyone and their dog is trying to predict what AI will bring in 2025. Based on what you are working with you will likely say that AI in 2025 is bound to be done on quantum computers, be all about speech, become agentic and solve complex business problems and much much more as evident from this fastcompany article.
Asking AI about AI
Now what happens if we actually as AI? I asked chatGPT, claude.ai and gemini 2.0 flash the following:
“What are your predictions for 2025, what are the top 3 industries that will be changed by AI“
chatGPT (GPT-4-turbo)
Gave me 2 different options to choose from that did not make identical predictions. Hedging some bets there?. Anyway the predictions spanned: Healthcare and biotechnology, manufacturing and supply chain, retail and e-commerce, and education.
What stood out in the predictions were the evergreen promise of precision medicine, AI powered robotics to drive supply chain, hyper-personalized shopping, and tailored tutoring.
Claude.ai (3.5 Sonnet)
Made somewhat similar predictions and offered up: healthcare, education and software development. Again with similar arguments – personalized medicine, emergence of AI tutors, and for software development optimism on handling more and more complex programming tasks
Gemini (2.0 flash experimental)
Again made similar predictions: healthcare, education and finance. Similar argumentation: personalized medicine, intelligent tutoring systems and in finance AI-powered robo-advisors for personalized financial planning.
Being critical
I don’t think these are the area that will be “disrupted” or changed very much in 2025, and here is why.
Healthcare has been primed for disruption and personalized medicine for more than a decade, and it is a slow change, not a disruption. The siloed data systems and patchwork of infrastructure makes data aggregation difficult. The data is noisy as different hospitals and providers have different standards for documentation, there are variations in diagnostic opinions (second opinions) and a ton of regulatory hurdles and data security and privacy issues to work out. Once you have done that, you need to get into the whole FDA / EMA race of approvals. This is not going to happen over night.
Education has proven resistant to change for decades. Nothing much have changed between my elementary school and the teaching my son is now receiving. There is still an over-reliance on rote memorization, standardized tests which fails to capture the student’s understanding and abilities to problem solve, and the introduction of AI assisted learning risks undermining critical thinking and the ability to learn to learn. What I mean by that is that learning is partly suffering, struggling to understand something that is just beyond your current abilities and keep at it. AI assisted learning could easily nudge you along with hints making that struggle far less, and hence remove your ability to get there yourself.
Manufacturing and supply chain is already pretty far along the spectrum of integration of robotics, AI / ML forecasting models, and predictive maintenance. I don’t think we will see big disruptions here, more incremental improvements. Similarly, with finance. Algorithmic trading, automated credit scoring, etc is already run of the mill – this will improve but not fundamentally change the industry.
Software programming is an interesting beast. Tools like Github co-pilot, cursor and others have made significant progress in code completion and boilerplate generation. Moreover, coding is a very structured task, following specific rulesets, design principles and have a very limited output space. For instance if I in python write
"for ..."
I have limited options. Either "for i in range..."
"for i, elem in enumerate..."
this makes coding a much more tractable problem to solve in prediction of the next tokens. Lastly, code is tested my unit and integration tests, meaning the AI could write these tests to validate that the program written is actually producing the expected results. This could be an area of disruption.
Retail and e-commerce. Yes! I completely agree. If you look at e-commerce sites from a decade ago and today they have not changed much in terms of the user experience (see my post on skincare e-commerce). A lot of the hyper-personalization has yet to happen, and I think this area is ripe for change. Partly, because the objective functions (what trains AI) is more or less baked into the e-commerce experience:
- Clear ROI metrics: conversion rates, cart value, customer life time value
- Clean transaction and behavioral data at scale
- Every small increment in search / recommendation directly impacts revenue
- Competitive pressure forces adoption – if your competitors us AI to optimize margins others must follow
- Infrastructure is modern and built for real-time personalization
My predictions
So I agreed with one of the predictions from the SOTA LLMs — and here is the rest of my predictions.
Social media is already awash with AI generated content and content creators are increasingly using AI to generate content to stay ahead of the game. The whole point of social media is engagement time and eye-balls, and the more you can trick people to click, view and stay the better. Hit those dopamine release buttons fast than ever and your revenue will grow. In addition:
- The data barrier is solved through clear user consent and massive amounts of engagement data
- Network effects amplify AI impact – as models improve user engagement, it creates more data for further improvements
- Relatively light regulation compared to healthcare/education
- Direct monetization potential through improved ad targeting and content recommendation
- The infrastructure is already built for large-scale data collection and deployment
AI research and development. This is the holy grail of many of the leading AI providers. Given that models and architecture is done by humans there are only so many experiments you can run. As compute continues to grow, and models gets faster to train and evaluate you are bound to hit an inflection point where your man-power and talent is the limiting factor. If you can get AI to improve AI, you have the potential for exponential gains in performance – all just at the expense of compute. So my predictions here boils down to the following arguments:
- Meta aspect of AI improving AI development creates potential exponential gains
- Research labs have the technical expertise to implement and test new approaches
- Direct financial incentive through improved model performance and efficiency
- Clean, structured training data from previous models
- Modern infrastructure built specifically for AI workloads
More pedestrian observations
I think AI is still very much in the hyper-cycle and I hope 2025 will start to normalize out expectations and claims. I hope to see many more laser-focused application of AI within small domains that really excel at solving just a single consumer pain point. I think there is a lot of success to be had in that approach.
We will without doubt see some spectacular failures as AI agents are released too early and given to much power in execution. Hopefully not something that will take down the stock market, disrupt power supply or similar big scale events.
I fear that our ability to distinguish between AI and human generated content will almost disappear, and this will be used for misinformation, propaganda and fraud at a scale we have not seen before.
What are your hope and fears for AI in 2025?