Updated: 51 minutes ago
I wanted to write on the topic of my general opinion on AI use (as in typed by hand on my Galaxy Note 20 rather than formatted by Claude) to provide my comments and general philosophy regarding AI, which I've also written about on my website and which I presented on at the IPI conference in the UK:
(Yes I did use AI to generate the PowerPoints but the script is 100% written by me; and while I did recieve job offers at University of Chongqing and Jinan and Shenyang I've mentioned in the presentation I declined them).
In my opinion this is extremely important to address because my opinions are somewhat unconventional and not really shared by anybody else I know in tech. I've thought about this topic for 16 years... what makes a person different than a computer? What is different about the physics? What makes us uniquely human? I've had discussions with many prominent researchers in this area over the years and this topic became central to my Project BLUEBIRD and Project SIDECHANNEL pitches:
I've discussed this topic with many of the "AI Successionist" "LessWrong" (even if you are less wrong you are still wrong) think tanks in Silicon Valley and I think that there is just a general misunderstanding about what AI is and the physics:
When I talk to claude about my ideas, at least 50% of the conversations I have with the chatbot it will tell me that my ideas and intuitions are "analogies" or "too speculative" or even wrong, and what I've found that the way I use the chatbot is more as a way to correct the chatbot than it telling me what to do or think. I will at times if I am working on a project spend a few hours with it to work out a problem and guide it where I want it to explore. Based on my experience with it, there is a 0% chance that it can "autonomously" act as a stand-in to replace intuition. Sometimes intuition or inspiration comes like a bolt of lighting or can only be obtained in physical ways that cannot be replicated by a computer program.
I, for example, spent over 16 years building tesla coils and noticed patterns in the plasma filamentation, and I discussed this with claude and it tried to claim that the patterns were just caused by random pertubations in ionization channels in the air. I knew this was wrong, but the chatbot is designed to think in a tiered way based on how much you probe it and challenge it. It turns out that the patterns I witnessed in the tesla coil streamers can be modeled by the Navier-Stokes equations and Maxwell's equations, with statistics which could be modeled by the Riemann zeta function.
This is not something I gathered from the chatbot but came from direct observations in the real world and publications I read many years ago, and so I reached out to Ed Witten about this and to one of the researchers in this area to chat (Dr. Migdal, who is a member of IAS who has claimed to have solved the Millennium Prize problem of Turbulence). I chatted with him and I reviewed his publications and conceived of the MILLENNIUM Project:
https://www.trevornestor.com/post/why-the-riemann-hypothesis-and-fluid-turbulence-may-be-the-same-problem
Sometimes intuition comes as a feeling that has not yet fully been articulated.
The week after my discussion with Dr. Migdal, OpenAI claimed to have solved this millennium prize problem - by Dr. Migdal is skeptical. So really chatbots have a place in the process, but cannot replace interactions with real people. This is much in the same way that books have a place in society, but cannot replace people (just consider the sorts of things that have been rationalized when people have been left to interpret the Bible on their own as a sole source of information):
The performance of AI trained on its own outputs without material or social feedback decays. Collective intelligence (CI) in people scales faster than AI per unit of energy and AI scaling reaches thermodynamic limits of scale (written about by Joseph Tainter). AI does not operate on the physics needed for interbrain synchrony (as a part of its business model) which is measured in people. In fact, interbrain synchrony underlies social adjustment and success on teams, and is one reason AI cannot simply replace teachers.
Linguistic intelligence is only a narrow kind and also requires human-in-the-loop interpreters to be intelligible. An LLM in itself cannot become conscious or replace people and stories of it "breaking out" are properties of its stochasticity and deliberately designed through its implementation with physical systems:
I was rejected for a project with Blue Origin because of my opinions on this (consequently the FBI also more than once wanted to consider me to become an agent and also to work on special programs - they also rejected me in all cases based on my strong opinions on the Epstein files):
Blue Origin wanted to strap an LLM to a rocket to use it for guidance. An LLM left on its own is not intelligent and will make the least intelligent decisions in the real world without constant corrective feedback (sort of like if you had a pilot flying without GPS then the trajectory of an airplane will begin to diverge from its intended path). The hype is largely a marketing trick:
As far as projects are concerned, typically my work flow is to select topics which I am familiar with and/or which are interesting. From there I will use claude to help me identify possible approaches based on published literature, and will look at the publications.
Afterwards, I select what approach I think might be best as a starting point and work with claude to help me put together a technical volume. At times I may combine several approaches (for EmberCell I was exploring sodium ion battery technology and another high capacity battery technology, and blended them together with carbon nanotube based sheets as an example).
After a discussion at times researchers will have feedback and I will ask them if they are interested in partnering or if they have any feedback to the draft, and often times they will (for Project BOLT for example researchers identified a publication which recently proposed an architecture for Nd:YAG lasers that did not require cooling) I will rework the draft and consider the changes.
I can spend hours with claude developing a project and claude cannot simply build a project, and when I use it I work through problems and make design decisions, but it is an easier interface than typing everything line by line. Sometimes I have arguments with the chatbot because it insists that a design idea or idea I've had is wrong or too speculative or just an analogy and I often prove that its wrong and the chatbot has to sheepishly admit it was incorrect.
Claude can help with identifying contradictions, searching publications, finding the contact information of possible vendors, researchers, and customers. Claude can assist in organizing notes and formatting project documents. Claude can help to make applying to grants in portals more conversational so that when I am doing the applications I can maintain my sanity during long hours in a basement filing the necessary information (it is just easier to have a digital companion to talk things through and search information rather than sit sifting through materials).
Claude, cannot, as far as I can tell, replace the human-in-the-loop. My general philosophy is that Claude is a glorified Google search or autocomplete (this is also the term linguist Noam Chomsky used to describe it). Rather than sitting down reading a book or a publication, you can simply talk to it - and that is what I think is most valuable about it. Claude makes mistakes but it is an excellent brainstorming and organizational tool that can be beneficial for those with my A(u)DHD. AI can be a tool to democratize science but if power over it is too centralized and it is used as a tool of financial engineering then it can be deployed to destroy a society because it flattens social capital - not because the flow of information itself that it facilitates is dangerous or because it in itself contains "agency." If AI development is balanced in tandem with material and social conditions in a society, and is used to build relationships, and it is open sourced and free, that is the best way to avoid this scenario.
There is a lot of talk about the need to scale up AI to "compete with China." China already has the winning business model - outsource the costly research and development for the models to the United States and then just cheaply distill the models, and then rather than monetizing and gatekeeping it, open source and distribute the models freely. The models can then be used for what is actually important which is building things - building infrastructure, manufacturing, and things that people actually need rather than wasting time and energy gatekeeping information and frightening people to sell a cure.
Many of these metanarratives (AI, climate change, terrorism, pandemics, etc) are cynically hoisted onto the public conveniently by the same people that sell the cure - that's no accident. That's not to say that these things can't be problems, but it's obvious that it's become a ploy to rationalize self destructive policies (for example, Sam Altman made a comment attempting to appeal to climate change for why it is better to replace people with data centers since it allegedly requires less energy for compute than to raise a new person to the age of 20):
Ultimately, in order to progress we need to have the tools and information to design things and actually build them. We need ways to share information collectively that aren't gatekept. Science by its nature is a democratic and universal process, but AI should be used to help build social capital and infrastructure rather than flatten it. Perhaps rather than worrying about whether AI tools are used we should be more concerned about who gates access to AI capabilities and how building out AI infrastructure is used as a tool tool divert capital way from everything people need to survive and centralize control and power in a society. Unfortunately, decentralizing AI also conflicts with its business model.
