Analysis: Claude and the Dawn of the "Post App" era

Claude.ai, the AI unicorn currently coming closest to beating OpenAI’s GPT4 in general performance, recently released a new UX feature which significantly changes how the user interacts with AI. It is the improvements to their User Experience (UX) howevever that show us where AI is heading in the long run.

Claude.ai, the AI unicorn currently coming closest to beating OpenAI’s GPT4 in general performance, recently released a new UX feature which significantly changes how the user interacts with AI. It is the improvements to their User Experience (UX) howevever that show us where AI is heading in the long run.

Working with Claude’s artifacts over the last few weeks, I realized it has become faster for me in some cases to generate a completely new application using claude than to wade search engines, dozens of github repositories and reddit to find, setup and run an app.

Now granted, Singapore has the most sane tax code in the world - but I reckon we can see the real glimpses of fundamental disruption here.

Significant Productivity Increase

Much of current conversation focuses on the most direct impact of computer aided code generation: The massive productivty injection into certain roles will that has limited potential of converting into additional sales - after all marginal costs of entertainment and software are already rock bottom, the attention economy maximized. Job market of technology is notoriously hard to forecast accurately, but it seems clear in 2024 that there will be significant disruption of labor following the pattern of machines in industrialisation - at an accelerated rate due to limited logistical adoption hurdles.

Dawn of The Post App Economy

But there’s something else that comes into view as a new generation of more powerful, user friendly AI tools increases accessibility to code generation. Where front-end developers may be grappling with the impact of v0 or Claude’s ability to write robust modern web components today, the question it opens for tomorrow is much more fundamental:

If we can produce situational and personalized (“I have bad eyesight, please take that into account”) interfaces for any problem on demand and in real time, are “Apps” still the best paradigm?

Software Engineers been the gatekeepers to the awesome power of general purpose compute for decades, learning programming langages to invoke it, handing it out to the rest of the world in magical artifacts called interfaces and apps. Now, just as gunpowder democratized the ability to kill, the ability to harness general purpose compute is about to be taken from the gatekeepers and democratized as well.

We are at the start of this movement, it won’t dominate overnight or even next year, but it looks clear to me that the manufacture is software and it’s specialized roles will be more disrupted than most people anticipate in the long run and the interface is one of the first bastions to fall.

Interactive content as a new frontier

Below is an example of turning a random facebook post about a tabletop gaming mechanic into an interactive simulator by adding a single sentence at the end and feeding it to claude ai. We can imagine a world where Meta.ai, instead of suggesting questions to elicit engagments asks “would you like me to turn this into a shareable, interactive experience.

What will happen next is the same that has already happened to other forms of content with the advent of accessible AI tooles: Github and Appstores will see an exponential explosion of content that will extend to other areas, like games, as the technology moves forward. That explosion of content will assert it’s own pressures onto the market and ecosystem and fundamentally rewrite creator economics in the process.

What’s holding it back?

There’s a lot holding this future back at the moment, with many issues requiring fundamental research breakthroughs or expensive custom engineering to overcome:

• AI system reliability: Reliability of even the best LLMs is still hovering in the 80-90% range. Good enough for single step tasks, but not good enought for further automation
• Introspection / failure detection: We still lack robust ways to detect failure.
• Safety issues: Critical safety issues such as jailbreaks and prompt injection continue to prevent most end user facing usecases at scale and new challenges such as training data poisoning are going to start asserting significant challenges in an environment where we want to trust the AI to create and run code on our behalf.
• The limited 4 kilo-token output window currently shared by most major models is limiting more advanced usecases.
• Integration into existing tools and work pipelines.

That said, even if reliability was to stall where it is today though - there’s no technical limitation for much more porgress at this point. It’s merely an investment focus issue. Currently companies are more interested in the gains made by scaling to larger, more powerful models. We can expect this to change as compute scaling reaches it’s limitations.