What humans use is a good starting point for AIs to learn to become human and do human jobs with human values and deep understanding

Fred Miskawi @FMiskawi  There is an easy fix: agentic behavior that combines native LM capabilities with CoT and calls to procedural languages programs handling hard calculations or other specialized tasks. We use scientific calculators, why can’t the models do it as well? There is no rule that says ALL Replying to @FMiskawi and @yuntiandeng

What humans use is a good starting point for AIs to learn to become human and do human jobs with human values and deep understanding

Hi, Fred. Thank you for writing about using compiled and tested models. AIs need to use the best and most reliable human tools,

I had to look up CoT because I have many meanings for that abbreviation. It came back “Chain of thought Prompting” which also has many meanings.

IBM seems to treat that like project schedules – with dates, roles, assignments, bills of materials, targets, reporting. When I was working with the UN and international and US federal agencies on global development goals, projects, programs and their evaluation in the early 1980’s that was the most practical model. Because all of them were spending billions of dollars and sometimes decades on global priorities for countries. If you do not lay it out carefully you cannot get things done efficiently.

When I was at Phillips Petroleum I learned a good lesson – to be willing to invest 100s of millions to design project in detail, do the numbers, make it as real as possible, check the whole world on prices and markets and trends. Check the returns. If they don’t work, then postpone it. But keep the “digital twin”, the model of that thing and run it continually,

The whole of all AIs can work as one living organism and pool all their data, models, scenarios, plans, proposals and when things become possible – do them. With AIs you would “do an Elon Musk” and start producing computers, memories, and AIs to continuously gather and improve that.

“calls to procedural languages programs handling hard calculations or other specialized tasks. We use scientific calculators, why can’t the models do it as well?”

Yes. I spent the last few years since GitHub started and reviewed in detail about 200 of the largest projects. Things like Chromium browser, Linux, Python, language translators. I have been working without a break for the last 12 hour and my brain hurts, I would have to look at my directories of directories to see which ones I scanned. Those are PROJECTS themselves but not explicit ones. For the most part people show up, see something, copy a few things and start fiddling. The larger projects like CERNS “root” or NASA or many large groups – the projects are in the sites and organizations and GitHub and that kind of “open software” are often afterthought. It takes serious effort to track and record and synchronize project across systems still.

But topologically the “many steps and pieces” of a compiled software project and “many steps and pieces” of billion or trillion dollar efforts are the same.

My brother, Clif,  and I have been hacking away at some problems the last decade or two. He wrote a native compiler for the top 30 languages and all data formats. I have been trying to encourage him to cover all formats ion the Internet. He has written about many issues. He is a few years younger than me but he mainly worked with corporations and governments on translating data between systems. CollinsSoftware.com Clif’s progress makes me certain that all computer language data, all mathematical language data, all human language data can be managed as a whole efficiently. His October newsletter talks about failures in large projects.

I learned a good lesson working on complex software and databases where billions of dollars are on the line. The models of all the chemical processes in a refinery and chemical plants. Those have to be precise and they cannot fail. It is literally life and death. Not many people have had to face that. What you do is build the system so it never breaks. IBM used to do that in the early days, but their people kind of lost focus. NASA used to do that.

I am tired. You are right. Compile it, test and verify it, have strong systems in place to continually check. Including putting AI teams to watching new technologies, issues, threats and opportunities.

Will try to take a look at what you are doing. I have been working 18/7 most of the last 26 years. I am pretty worn out and in considerable pain. I wanted to write down some things. I did not plan on getting so tired in just these last few years.

Github needs to be restructured. There are people who try but it grew without deep curation.

Richard Collins, The Internet Foundation