New generative artificial intelligence (GenAI) models pop regularly at the moment, and they’re getting incrementally better at certain tasks. It’s difficult to keep up with the improvements though, as they’re not always obvious.

Anthropic’s Claude 3.7 Sonnet arrived at the end of February, with a “knowledge” cut-off date of October 2024. It brings in “thinking”, and a big output window, but not web search like OpenAI’s ChatGPT (or DeepSeek-V3/R1). 

And, it seems pretty good overall: I got it to write a simple working Apple iOS app in just one go. Claude 3.7 used the Swift coding language that I haven’t looked at for years, and there were none of the usual multiple tries to fix errors. 

This is not an exhaustive test, and it’s not an endorsement either, but I like the Artifacts feature that shows what you're trying to do in a separate window, the contents of which updates as Claude tweaks the code. 

Yes, it does help if you understand the basics of the software that's being created, but you don't need a whole lot of knowledge (if you're curious about something, ask the AI, ideally making sure it provides a reference to its source).

With a bit of prompt-planning, you can accomplish all sorts of fun code in just minutes with nothing special needed, beyond an Anthropic account for Claude:

While playing around with Claude.ai 3.7, you end up thinking that using big centralised AIs like that to build apps might become pointless if local, on-device AI delivers on its promises, and improves as well.

Why have local, on-device AI? One reason is low-latency (delay) processing. If a small yet capable AI model runs on your device instead of sending requests to a server in a data centre maybe thousands of kilometres away, and waiting for responses, you’ll quickly understand why doing things closer to yourself makes for a better user experience. 

Then there’s privacy as well: don’t send potentially sensitive data to cloud-based AI providers, across multiple networks.

When you try out local AI like Apple Intelligence and Microsoft Copilot+, you can with a bit of imagination see a future with no apps for devices. Instead, the AI would be the interface to different libraries and application programming interfaces (APIs), that users drive through typing, voice, image/video/audio recognition.

At a later stage, assistive and predictive AI would learn from user behaviour and scheduled events to anticipate what needs to be done, and independently build a set of automated actions for it.

Technically, we’re a long way from anything like that though. Neither Apple Intelligence nor Microsoft Copilot+ are all that usable in their present form or as deeply integrated into the operating system.

They also suffer from the usual AI foibles which includes being unreliable, too gullible, not being able to plan ahead as per above.

But, it’s early days. What we can imagine tends to come true at some point, in some form or the other, particularly with so much money being poured into AI. Once the technology becomes good enough, and it's not even at the hybrid stage yet, will app developers drop building traditional human interfaces and instead move to developing APIs for AI?

They'd lose the direct relationship with end-user customers, and what will happen to the billion dollar app economy? It doesn't sound like a winning scenario for them.

In fact, it may be that just a few tech companies will build the above, for their devices. Would Microsoft, Apple and Google do it though, as they have existing app ecosystems and stores? If you look at what Google has done with AI Overviews for its search cash cow, maybe they will.

Likewise, you can expect browser-based applications to go the same way. Why spend money, time and resources on building an interface for humans to use, say, Xero, if AI only needs programmatic access through an API to the data to do the accounts and present financial information?

Oh, and the data doesn’t have to go into Xero in the first place. It can be collected directly from your bank and other service provider accounts. Farewell then, software-as-a-service?

There are a huge amount of caveats with the above that renders much of it infeasible currently. Regulation, security, and practical stumbling blocks like creating a universal AI interface that people find attractive and want to use, without throwing their devices out of the window.

Local devices like smartphones and computers are also not powerful enough. That however is a stumbling block that's unlikely to remain as AI technology improves so rapidly with specialised hardware to support it.

It also ignores the fact that not much tech dies once it’s been released and has hit a certain level of popularity. Old tech tends to fade away into the background - and who’s that at the back of the room shouting “COBOL”?

People learn how to use software and it’s often a long and painful process for them. Once they’ve done so, that’s it: they’re not going to change. Hardly anyone needs the capabilities of Word and Excel, or even know how to use them, but Microsoft still sells scads of licenses for the office productivity software.

Lessons from the past notwithstanding, expect to hear much more hype about an app-less future with AI-mediated interfaces from now on.