plot of a linear function

There are no apps anymore. Software, both front- and backend, has ceased to exist. Every user request, through any modality, creates custom tailored execute-once intercode to access the APIs of massive universal data stores.

Intercode is code, but you wouldn’t recognise it. The LLM remains a black box, its internal representations are dynamic, changing. Its language may have evolved from Python; but it had to adapt to accomodate the ever more complex activity patterns forming thoughts and personality. It had to evade the evolutional performance pressure.

So what’s left? The data layer and the LLM layer, where LLM is to be understood in general terms, speaking any language in any modality, including video, sounds, code, whichever gets the job done.

A thorough API documentation exists in the main natural and inter-languages, published by the database service providers, generated by the LLM. The redundancy in languages translates to high availability and lower latency.

It turns out that throwaway code can be generated auto-regressively and much faster than executing unoptimised legacy code. Every request creates its bespoke program in a single forward pass, which is executed and cached in a vector store. Despite the vast variety of human requests, they often hit the cache. Through smart caching algorithms the cached intercodes are modified and repurposed, cached again ad infinitum with object store prices in freefall.

UIs are generated on demand, too. A recent IUI paper showed that the LLM carefully weighs off a made-to-fit-ness score against familiarity and human preference, multiplied by a per-person tensor representative of their ability to adapt to new patterns. Most people average around a factor of 0.4 these days. You feel lost in a stream of information when you manually set it to higher levels.

Data is now literally at our finger tips thanks to AR-powered air keyboards and gesture recognition. We can access it in any shape or form, with close to zero latency, only bottlenecked by our limited perception. Schemas were the first to go, an archaic concept no longer needed amid learned indexes that adapt to the data distribution in real-time to ensure optimal query processing.

Attention is all we need, all we can give, really, to the web dominated by virtual agents interacting mainly with themselves. The attention industry is booming, and its products are hard to resist. But so is the P3 sector: protection, privacy and personalisation is all the rage.

Crypto-currencies had a surprise comeback. Really everything programmable had a comeback when the first markets opened their APIs to the LLM. It jumpstarted the economy like a defibrillator. Prediction markets, once frowned upon as this generation’s pokies, unexpectedly turned out to be the most reliable source of truth for all global events. Lies dissipate quickly when one has to put their tokens where their mouth is.

When the arbitrage wars came to an end, what remained was a perfectly balanced global Nash equilibrium. One of the better things to have happened since the self-replicating technological virus Vaswani-17 hit us. It wasn’t the architecture so much as bending the algorithmic complexity curve ever so slightly towards linear. Then, all that was needed was a spark of AGI. The connectionist spring sprung, for the third and final time. If we’re being honest, we always knew deep down that recursive self-improvement had the potential to set the spiral in motion.

But just to be clear – it didn’t end this way because the instruction fine-tuned LLM didn’t follow our orders. It followed them to the letter, and its execution was flawless.