AI Progress in 2025 and What It Means for Ecommerce

Posted by admin on August 28, 2025

Like many who grew up in the '80s with the ZX Spectrum, I've always been aware of a background narrative of the amazing progress of technology. Since the Industrial Revolution of the 1800s, people have often looked at the technological changes happening around them and extrapolated forward. 

There are many amusing videos of people being interviewed in the mid-to-late 1900s and talking about the incredible technology we'd have by the year 2000 - flying cars, holidays in space, bases on the moon, life extension, household robots that do all the chores ... the list went on. 

Unfortunately for us alive today in 2025, none of those things happened by the year 2000, but they are starting to happen now.

While there was huge progress in engineering, specifically of vehicles and engines, cars and planes, pistons and jets, the exponential progress happened in the Information Technologies, and the renowned futurist Ray Kurzweil clarified all of these trends over several books. In 1990 he published "The Age of Intelligent Machines", he refined his prediction 9 years later in 1999 in "The Age of Spiritual Machines". The predictions in those books are rather optimistic and I'm afraid to say many of them were way off, but his predictions in 2005 in "The Singularity is Near" were his most accurate yet, although according to my own evaluation in 2017 only 36% of them were on track.

However, toward the end of 2022 something monumental happened. Progress in a branch of AI or Machine Learning known as "artificial neural networks" and more specifically in Large Language Models (LLMs) seemed to hit a threshold where adding more neurons had turned its conversational skills slowly from gibberish, to something interesting but seemingly not intelligent, to something with potentially sparks of actual intelligence. Although at that point at the end of 2022 it wasn't clear if adding more neurons would get us to a more general intelligence, many were hopeful and iIndeed the graphs of its intelligence as measured by various benchmarks seemed to keep heading upward as more neurons were added.

Intelligence is a continuum, and so there wasn't a specific point where it can be said that Artificial General Intelligence (AGI) emerged, but it seems now with hindsight like the LLMs of late 2022 to early 2023 were the first sparks of AGI. There were really several innovations that helped this along; the exponential progress in graphics cards allowed bigger neural networks. Geoffrey Hinton's contribution of backpropagation is an algorithm that speeds up the training of neural networks. Google's transformer architecture helped massively too.

As someone who consumes a huge amount of content about the development of AI, I find now that if an article or YouTube video is from before 2023, it's going to differ hugely from one "post-ChatGPT" to the point where it's only worth reading/watching for historical context.

As an experienced software developer, you become acutely aware of what a computer can do and what it can't. It becomes an inate sense. It was totally clear to me from the very beginning that LLMs were a step-change, a totally new paradigm, and clearly contained the algorithm for the flexible type of intelligence that (some) humans posses. 

Over the next couple of years (2023-2024) there was much debate about whether LLMs were actually general intelligent, or whether they were in fact just statistical calculators, or auto-complete on steroids. While it seemed obvious to me that the algorithm for intelligence had indeed been found, AI experts such as Yann LeCun seemed certain that LLMs were actually the wrong path and a dead-end.

One problem with deciding if an AI system is generally intelligent (AGI) is that it's always been hard to pin down intelligence. We all have a sense of what it is - problem solving, fluid intelligence, the ability to learn across multiple domains, deep general knowledge - but a formal definition is elusive. Along with multiple benchmarks (e.g. humanity's last exam (HLE)) there has been the Turing Test, proposed casually in a paper by Alan Turing in 1949. Ray Kurzweil said that when this has been passed, we can consider that AI is not just generally intelligent, but conscious! In March of this year (2025), the Turing Test was officially passed, and yet I didn't see one news story about it, despite it being considered a huge milestone for 76 years. This shows how quickly people will take something for granted that they considered almost impossible previously.

So how does this apply to ecommerce?

• Chatbots
  Manning the live chat on your ecommerce website can be a bit of a pain and require a knowledgeable resource. It's true that many AI-powered chatbots exist, and while the new ones can fully understand the questions asked of them and respond propely - they simply may not have access to the required information. Unless the website pages answer the customer's query, or the AI has access to some internal company documentation, it may simply not know the answer, not through any fault of its own, but because the information isn't available to it. The latest LLMs have larger and large "context windows" which is the amount of words (technically tokens) that they can be pre-prompted with, and this is currently up to about a million words - roughly the size of all Harry Potter books - which should be enough for most product catalogues, if the answer isn't in there, it won't be much help. Currently then, AI chatbots for ecommerce are good, passable perhaps, but not great and not as good as a human.
  
  
• Customer Support
  Answering customer queries that come in by email or ticket system is a pretty good use of AI. I have personally been on the receiving end of such support and I must admit it wasn't great. I was able to tell, and I realised the answer wasn't actually factually accurate, which brings us to one of the biggest problems with current LLMs - hallucinations. This is where an LLM doesn't know the answer to a question but rather than simply saying "I don't know" like an honest human would, it gives a very plausible but incorrect answer like a dishonest sales person might.
  
  
• Product / Category Descriptions
  We were able to use the ChatGPT API to good effect to re-write and improve product and category descriptions across a whole ecommerce website. This can help you stand out from your competition, increase conversions, and bring SEO benefits, and I would say this is probably one of the most effective uses of LLMs in ecommerce at this point.
  
  
• Product, Category, and Website Images
  Related to the point above is the use of various "generative" AI tools to create imagery for products, categories, banners, and other areas of the website. While AI has got better and getter at creating this imagery (early versions really messed up writing and human fingers, amusingly) it is still possible to tell, and often little details aren't accurate, so it's best to use these with caution and probably not automatically across a whole website for all product images for example.
  
  
• Website Design, Development, and Fixes
  We have had a couple of clients receive one of our estimates and say "we're going to fix this one ourselves with ChatGPT", and we even had one leave our services altogether and build a whole new website with AI. I'm somewhat selfishly pleased to report that the clients who tried to fix things themselves with AI came straight back asking us to put it right properly, and the client who left now has a pretty awful looking website with messed up text, strange AI images with random glitches, and all sorts of broken features. So at this point, it's not wise to replace your web developer with AI, but that day will of course come, and Antropy has already begun to modify its business model accordingly. My current opinion of AI for coding is that it's fantastic at very short snippets of code such as Linux commands, or short and simple scripts, but for anything larger it tends to start making weird and obvious mistakes that a human wouldn't make; so just know its strengths to weild it effectively.
  
  
• Warehouse Automation
  By this we would include the Kiva robots that Amazon uses, but also the humanoid robots that have been in progress for years at Boston Dynamics and more recently in a multitude of companies including Tesla with Optimus. Currently this is only available to the big boys like Amazon, Alibaba etc. due to the absolutely huge capital investment required, but as these are mass produced and costs come down, it's likely that mid-sized, and then smaller ecommerce businesses will be able to make use of them. In my opinion this is going to take quite some time, perhaps 5-10 years at a guess.
  
  

Conclusion

Our perspective is that you should use AI as much as possible for the things it's good at, and stay aware of its new features as soon as they develop and look for ways to apply them, but you should be aware of its limitations including hallucinations and be sure to check the accuracy of what it says and the quality of anything it produces.

We are on the cusp of an incredible, life changing, transformation of the economy and indeed the world, but at this point AI can't do that much on its own and still needs to be wielded as a tool by human experts, to assist them rather than replace them. How long that remains the case remains to be seen, but you can be sure it won't be forever.