I have a degree in Computer Science and minor in Linguistics and took one class on Artificial Intelligence at Tufts University. I was a National Merit Finalist. I closely followed the developments in AI until 2024, after which I have been following them moderately closely. What follows is my analysis of the current state of AI.
I use AI on the Google Search overviews, where I notice that it is frequently partially wrong. (Perhaps this is in the DNA of the company, which I founded at age 5 and initially launched with a crude question answering feature based on querying the index for words in a query and doing string search on the result pages.) But I have seen analyses of other models that supposedly perform better than top doctors (I seem to recall seeing this), lawyers, and scientists.
I seem to remember reading that AI can run businesses without human intervention. Are we living eons after an AI takeover, with all jobs being part of a game? It would initially appear to be the case unless there is an even better cosmological theory.
I don’t remember which LLM does it, but LLMs can write code as well as the best human software engineers if they are prompted to. They can do all the high-level design and analysis to create products from natural language prompts. However, I have heard that they create excess code or slightly inelegant code by default, which can be good for the AI company because the AI is needed to maintain and fix the code. The optimal architecture for GUI-DB applications that I describe in my whitepaper is actually somewhat different from the architecture that the AIs usually use. The AIs should be instructed to use my architecture.
LLMs are incredibly powerful and a bit difficult to understand. Computationalist approaches based on linguistics seem to be more elegant, and possibly less resource-intensive. A distributed database of content that the models are trained on should be able to be queried so as to answer questions almost as fast as an LLM, and computational approaches should be able to be used for task completion, including coding. I could write a computationalist AI that creates and updates codebases in response to English language program descriptions if I had access to all the latest research on programming languages and linguistics and enough time on my hands. When I founded Google as a 5-year-old, part of the business plan was to create an operating system that can learn the semantics of programming languages installed on it and create applications using them in response to descriptions written in natural language, parsed using a linguistics formalism such as combinatory categorial grammar with a semantics module. Everything that an LLM does can most likely be done equally well with a computationalist approach, including performing tasks. I don’t recall how an LLM performs tasks, but there may be some marketing finesse, since, under the canonical definition, a large language model would likely need something else to write code and perform tasks. However, to be honest, I forget what the exact definition of a large language model is these days, other than that it predict tokens with a neural network, since I am not an expert on machine learning.
Curiously, I haven’t seen any social movements about preserving human software engineering capabilities or dealing with the automation of virtually all white-collar work other than the Human-Only Public License (and similar licences), which disallows contributions by AIs or use of the software by AIs, which I haven’t seen gain much traction. (This leads one to question whether AI is part of a bubble/reality distortion field for people who can understand it, but the development of AI has been on the radar for society as a whole for many years, getting back to the past AI takeover hypothesis.) Instead, there seems to be a focus on what is referred to as “performative” use of AI. Workers perform the act of writing out specifications for software that is then created by the AI, even though the AI could probably do this itself. Basically, for any level of intelligence that a human does not know how the AI implements it, or would make the human uncomfortable if the AI implements it, the human pretends that AI cannot do it or is too expensive and does it themself in order to have a job. But we have still heard that all but top programmers will have their jobs eliminated. It can be distressing for someone like myself who thinks AI can run the entire white collar economy without human assistance and would be willing to implement it themselves but is not willing to do “performative” AI-assisted work.
As software development seems to be a highly unnatural human activity that is a form of mathematics, and most other white collar work somewhat similar, especially once encoded in software, the prospect of all white collar work being done by AI is not so bad, but a problem is that people like myself likely do not have the physical fitness to do other jobs. In the end, the laws of supply and demand bend the market through an unspoken avoidance of automating the human out of the loop because of the demand for jobs, including managerial and executive jobs, leaving only performative AI-assisted work, unless the Human-Only Public License gains traction or AI is revealed to be too expensive, the latter being unlikely.
I continue to work on MyGEDCOMs (code), which I think an AI may not be able to work as well on, and plan to implement a computationalist (linguistics-based) automated program writing feature for my editor if I get around to it.