Opinion: Bubble Wrap 2.0: AI Immaturity in 2025

Remember the Bubble Wrap app? Way back in like 2008, one of the first “novelty” apps to take off in the iPhone App Store was a Bubble Wrap simulator - literally just a screen of bubble wrap bubbles which you could “pop” by tapping them.

That was it.

And it was insanely popular. Some people at the time [citation needed, so sue me] went as far as to suggest that the app was in its own way partially successful for the explosion in popularity of the iPhone itself because (and I quote from distant memory here) “people felt that ‘if there’s an app for that, there could be an app for anything.’”

This was, as I say, back in the late 2000s. How very prescient.

The app of course spawned myriad copies and spin-offs, because as everybody knows, you can’t copyright an idea, a loophole which has been exploited by those devoid of the apparatus to generate fresh ideas themselves since the dawn of copyright law.

So what’s this got to do with AI in 2025?

Well, when I look at so many of the “AI-powered” startups, and comparing what they are offering with the basic capabilities of current generative AI transformers, I conclude that the only thing they can be doing is reselling the APIs of existing, hyper-scale Foundation Models from providers such as ChatGPT, DeepSeek, Luma and so on. For so many, this is the only practical explanation for what they are offering.

Nothing wrong with a bit of resale, but the point here is that so many of these startups add absolutely no value whatsoever, beyond providing an interface for humans to submit prompts to a third party API.

They are, therefore, commercially indistinguishable. All just endless knock-off versions of the same thing: a human / API interface.

So many therefore are ultimately forgettable and disposable but they are collectively contributing to the insane frenzy and hype around this generation of AI capabilities.

“If AI can do that,” you may think, “it could do anything.”

And just like the wow factor of popping pretend bubbles on a thousand pound handheld computer had no actual place in the long-term story of what is now the single most successful personal computing device in the history of humanity (so far), the funky quirky video-producing or essay-writing apps have no place in the future of what will doubtless prove to be the single most significant technological development in the history of humanity (so far).

So much Bubble Wrap.

But it gets worse for these resale apps. BECAUSE these orgs are (in my estimation) largely unskilled in the actual mechanics of model pre-training, finetuning, scoring, and so on, I would wager that many of them do not really understand how “the thing” they are reselling actually works.

Which means that for the ones who don’t just run out of cash or get squashed in the undignified scramble to the capital trough, they face a stressful and uncertain future, with many future planning meetings following a template of:

Product Person: WHY HAS IT STARTED DOING THIS?

Tech Team: We have absolutely no idea (nervous twitch)

Spurious results, polluted training data, unpredictable JSON formats, inconsistent response times. These are all potential deal-breakers and when 99% of your business model is basically to lean into somebody else’s mysterious and magical API, you better be ready for a wild ride.

Now - haters, hang on a sec - this is NOT me trashing all of AI with one ignorant stroke of an Apple Pencil. In actual fact it’s clear to me that we are at the beginning of the most incredible phase of innovation in computing, possibly ever. Now that a few crucial evolutions have matured - for example, the emergence of Foundation (multimodal) Models from Large Language Models - the shackles are off and we as a tech community can truly begin to explore and innovate.

It’s just that this currently unmappable future is not at all likely to include apps whose entire purpose is to place a flimsy resale wrapper around somebody else’s planetary scale API. These are all just momentary anomalies during this immature phase of AI.

So where will AI make a difference? Hard to say for certain but for now I feel the following areas are most obvious:

• Autonomous agents working with a user’s authority to make bookings, update subscriptions, automate lengthy form-based processes (much is currently being written about this; in fact I expect “agentic” to outstrip “AI-powered” as the de rigeur buzzphrase of choice in late 2025).
• Expediting complex consumption and analysis processes (such as tendered supplier selection) by making AI “read” complex offerings packs and summarise them into homogenous, more condensed choices, comparing and scoring each according to some established framework
• Performing validity checks on documents of complex claims by constructing DAGs of trusted cause / effect references (like a truth blockchain of references to reliable resources)

But in reality, who knows? But one thing is for certain: we’re just at the beginning.

About Me

I’m Mark, a technical leader with over twenty years of experience in technology and the implementation of it for practical purposes.

Please feel free to come debate me on LinkedIn about this or any of my other posts 😋 https://www.linkedin.com/in/markhenwood/

About this Doc

Thanks to Chip Huyen and O’Reilly Books for their excellent book AI Engineering which has helped me to see inside the processes for the first time, and to understand the perspectives involved.