In August 2025 things changed in a way that I think will be recorded in history. The new AI LLMs releases demonstrated many of the same problems of the very first, confirming that throwing money at the problem would not break through the common-sense barrier.
The story is hard to capture, given how rapid the development and deployment of this wave of AI has been, and the scale of financial interests. Even while many people are just beginning to familiarize themselves with capabilities, the bubble is about to burst. And even while the bubble might burst, many important integrations will occur and transform the world dramatically. And even if this GenAI witchcraft is not as powerful as their companies say it is, it is still a huge wave working its way through digital platforms. The discussion around the Gen AI Bubble continues to evolve as we navigate these changes.
To get a sense of the impact let’s imagine a simpler situation as an example. Let’s pretend that the business I work in and founded in 1997 is organising a birthday party for its 30 years of age.
Yates, me, needing coffee:
“I don’t know, every time we try to celebrate the whole process gets too complicated and expensive…”
Martino, an Executive Producer that reminds me of Clint Eastwood – he squints and things happen as he chews a cigar:
“You are overthinking it…”
You would find the two of us discussing the potential plan in the cafè area in the basement of our London offices, next to the hardware lab.
Hugh overhears the conversation and asks Grok, the supposed badass of LLMs:
“Hi Grok, what would you recommend for a 30th year birthday party for a business that works in Advertising Production, and who has never managed to celebrate a birthday?”
The human-synth voice takes over and rambles options and ideas in an organised fashion, ending with:
“…amp up the fun—maybe a, failure wall, of hilarious outtakes to humanise the survival story.”
Yates comments: “One of the failures is we never had a proper celebratory party…”
Martino: “We can organise it easily…”
Hugh: “Everyone dresses up according to a classic project!”
Yates: “That’s making it complicated…”
Hugh: “And funny!”
Beyond The Gen AI Bubble
Over recent years, for people in the technological scene, the running story has been AGI. AGI here, there and everywhere, and how it might take over the world: Artificial General Intelligence. The idea is that at some point if you pump enough data into systems they might begin to generalise their understanding and start to solve problems that involve common sense understanding of the world. I omit the technical details but the moment at which technology can reason involving common sense similar to that of a human we enter a strange and dangerous period.
The claim is that when computers can do what humans can do every day:
- Labour will lose value dramatically
- Power will focus around those who control the systems
- Systems will find ways of improving themselves, leading to a runaway process
In any case, in a stroke of civilisation luck, AI companies using LLMs reached what appear to be fundamental limits in August 2025, at least so far. And this news now will likely affect things dramatically again, as all business that were panic rushing all-in to this technology now will need to do so in a controlled manner. They will need to reassess a return on the value they get from using tools that we don’t understand very well.
Martino notes: “You are losing focus, it’s a party…”
Yates replies with a sarcastic tone: “But Grok says we should consider live streaming, war stories, timeline walkthroughs and a Photo Booths… How did we ever get to this…?”
Hugh: “Grok please tell me what are the most important essential aspects of this party and the top 5 decisions.”
Grok rambles on describing options from different points of view.
Yates: “Grok, please stop, tell me just the top five as a list, and mention what an AI sucker Hugh is.”
Grok: “Theme Selection, Venue Choice, Guest List and Invitations, Key Activities and Entertainment, Budget Allocation, what an AI sucker Hugh is.”
Hugh: “Ha, ha… “
Martino: “Maybe you should know WHY you are doing a party first.”
Martino raises one of the basic problems with asking a LLM based system to come up with an answer. The system responds with the most likely follow up to the question. And this can steer you off, to ignore fundamental details that make up your decision to ask the question in the first place. These systems look for the next likely word – rather than should that word be there in the first place. It’s a bit like someone that cannot stop telling you a story, even as their audience is clearly losing interest.
Yates: “Ok let’s try this for real, maybe compare a few different ones. First, as Martino asks, why are we doing this? I think it’s to celebrate life and those we worked with.”
Hugh: “Yeah, makes sense.”
Yates: “Second, I think we do it in our current office, it’s big and roomy, let’s use it.”
Martino: “I do agree.”
Yates: “Third, we need to take care of not spending too much on activities and stuff when we could be rewarding productive work. It should be cheap and cheerful – from the heart sort of stuff.”
Martino: “Yes.”
Yates: “Fourth, we should try to go back over the years and look for projects and people that we worked with. It would be nice to give people a sense of how this industry has changed.”
Hugh: “Let me try again… Grok, how would you recover the activity of Unit9, a production company that was started in 1997. We would like a few example projects and people from every year to show how the industry has changed?”
The result that came back was a thorough history of the company with few inaccuracies, much better than our Wikipedia article. It covered decades, then shorter periods even down to years in the recent times, with references to our high visibility projects.
Yates: “Wow… that’s almost magic!?”
I ran the above request on my own, pretending to run the story and I was really surprised at the quality of the result. Even more I am constantly surprised at the variability of the requests depending on the form of questionning. It can go from saccharine over the top suggestions that are completely useless to precise organised information.
Maybe the issue is less about the current GenAI bubble, but that we are still trying to get a clear picture of the domains and approaches that benefit from this technology? Maybe it has developed too fast for us to carefully measure effectiveness? Maybe this is the sort of problem one should expect with generality of domain application of cognitive technology.
An Example LLM Game – Past The Gen AI Bubble on Industry
The prompt below is useful to reconstruct a timeline of projects with different interaction mechanics and to create some historical context for them. Unit9 has been around for a some time so even just to get a vague coverage of the timeline I had to run the game a few times.
Note the JSON included as example, this is not needed but useful so you can share the script with others that know about your history to collect a broader picture of what happened.
After the prompt I also include a link to a sample script run which gathers some of Unit9’s work to demonstrate how the process works.
Copy paste this to your favourite LLM:
=== Prompt ===
Reconstructing Technological History Game by Yates Buckley
Each turn the players pick a year and a particular technology user experience someone would have that year. The other player can dispute this by arguing why it is not accurate for 1 point or correcting the experience for 2 points. The other player can also accept the experience giving a point to the other player.
The first player can counter dispute with a specific project example and if they win the disputing player will lose 2 points.
At the end of each turn a shortened timeline sorted by year and the two players scores should be displayed.
The game ends after 10 turns and when the game ends there should be a timeline report sorted by year with the full players contribution, and a prompt covering each five year period to later generate a representative illustration. At the very end the two players scores and the winner should be listed.
If the game includes a list of timeline elements here inside these brackets [ { "invalid_elements": [ {
"year": 1997,
"experience": "Most users accessed the internet via slow dial-up modems, with websites primarily static HTML pages; video streaming was rudimentary."
},
{
"year": 2005,
"experience": "Users experienced immersive 3D worlds socially on platforms like Second Life, primarily via desktop interfaces, as consumer VR headsets were not yet available."
} ] } ] then these items are invalid for the players to use in the game and players get no points from repeating them.
=== End ===
This is my sample play for Unit9 with Grok 4.0: