On AI & Software

AI & Software: A conjecture to disprove.

It was interesting to read the prompt behind disproving the Dinitz-Garg-Goemans conjecture by using a counterexample and how Terrance Tao talked to ChatGPT about it later. The reported cybersecurity incident between OpenAI and Hugging Face is also quite interesting. These two incidents have shifted me from a more neutral stance on AI to a more extreme one. AI really is a tsunami.

To work less, to work easier, to work faster, to work better we humans introduced leverage. First leverage was human (hiring labor), then physical strength (machines), then capital (borrowing and lending; please note that this order of levers could be different), then systems (modern management, software, etc.) and now we have introduced a lever for intelligence and knowledge.

All existing knowledge will be in a bottle. Further knowledge-work will not be making use of learned knowledge in a setting but rather pushing the frontier of knowledge and making the bottle of knowledge bigger and better. Knowledge work will remain exclusive to those who have the talent and know how to challenge the frontier. Today making use of existing knowledge and applying it is a valuable skill; tomorrow it will be less so.

I read on a HN thread that we are in the COBOL era of AI. The world will change. Embedded intelligence will be mainstream. The distinction between deterministic, probabilistic, and hybrid algorithms will be more pronounced and formalized. In a few years the AI supply chain will become more efficient. Specialized inference and training chips, architecture embedded in hardware (for example, specialized hardware for KV cache), and other such innovations will solve hardware inefficiencies. Better optimization techniques, better algorithms, more domain specific architectures. Emergence of extremely large (tens of trillion or params) and extremely small language models will emerge.

Software is just getting started, and the next step is meta-software. The coding AI harness will build inside meta-software. Software engineers will be responsible for the meta-software, traditional software engineering (building domain application) will regress to domain experts. Meta software is dynamic, on-demand, and integrates with other software. Meta software houses other software, grows organically with the organization, it is deeply integrated with the underlying infrastructure. Security, governance and controls are baked into the meta-software. AI works with and morphs meta-software in real time. There will exist software marketplaces on top of a meta-software platforms, where software will be commoditized. Migration from one software to another will be made simpler, thanks to AI. The most valuable of software will be characterized by interoperability, integrations with hardware and software, and the world.

In an interview I took last week, I instinctively asked the question "What does your Claude Md look like?". The nature of work is evolving. People should continue to learn domain – engineering, medicine, finance – however, they should be pushing frontiers and should not stay an incumbent of knowledge passed down from predecessors. Engineers will generally move up a level, from software to meta-software, and domain experts to users of software to builders of software on top of meta-software.

The economic effects of further AI development are also to be reckoned with. A smaller working population can support a larger ageing population. Population will further shrink. Consumption per person will skyrocket – a declining population, and more abundant resources. Work becomes optional.