My Philosophy on AI
Updated: 5 days 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. At times the behavior of nature surprises researchers and defies simulations predicted in a computer or by AI.
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:
Blue Origin wanted to strap an AI to a rocket to use it for guidance. An AI 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 repetitive 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.
It is important to build a society where we not only have the right ideas, but the physical means to build them out. Without the material conditions to implement ideas, they are worthless - and the rollout of AI infrastructure at the expense of the real economy is exactly this tradeoff.
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.
When there is talk about regulating AI and preventing it from "breaking out," or "falling in to the wrong hands" many companies are asking in reality to regulate the public's access to AI, with gated access that is centrally controlled - it is already in the wrong hands (regulatory capture).
The design is to capitalize on the deferred promise and moving goalpost of "AGI" and to profit off of its successes, outcrowd human agency (by convincing the public these "agents" are autonomous and have their own agency and to pose as people), and maintain plausible deniability for its failures and harms ("the agent went rogue"). In a sense, the "AI revolution" has become a cult which includes its own end times metanarrative, imperatives, ecclesiastical structure, and clergy that speak from on high in ambiguous terms ("there is a 10% chance this could destroy humanity" where did they get this number? Is there a way to become more specific? Bioterrorism and asymmetric warfare? Breaking into critical infrastructure? Collapse of the social contract and economic systems?). Many of the AI boosters and safety figureheads are also paid with clear conflicts of interest.
I do not fully buy the argument that an intelligent AI or swarm of AI agents can meaningfully cause harm besides through their connection to the real world. Intelligence in itself does not have any power unless it is interpreted and wielded and then connected to physical systems - no matter how intelligent it is. One problem is that since our economy is so digitized and our infrastructure is connected to the internet, AI agents can infiltrate physical systems (or rather those that deploy them can plead plausible deniability for what they do). This is limited, however, and can be mitigated by building local social infrastructure where rather than an over reliance on the internet and digital forms of community power is distributed in social networks in the real world.
This idea that we are going to see "self replicating robots" and "world models" where if we just hold out long enough we can drop robots into society and train them to replace workers I find to be equally as ridiculous since we can't even figure out how to collect the rare earth minerals to build the robots, we don't even have the time or resources to raise or train our own children, and if this was viable believe me it would have already been deployed in Iran or China would have used it by now to revitalize its manufacturing base. Proposals floated to "space based data centers" (way more expensive to maintain and radiative cooling means they would be prohibitively large) and data centers the size of Manhattan (Mark Zuckerberg proposed this when we cannot even house our own people) are signs of delusional market psychosis.
When it comes to concerns about asymmetric warfare like cybersecurity breaches, bioweapons, and nuclear proliferation the tools are already abundant to deploy without AI. The barriers to entry are already low and are in the hands of those currently in positions of power in the world, and even easily accessible to the public (did you know that Rosary peas contain one of the most potent toxins known to humankind? It's possible to buy these from Amazon - extraction of the toxin seems to be less complicated than getting a job in this economy). That is not the biggest concern but instead how power is distributed and regulated and if it is wielded to build a better future for the public or for the gain of a select few at the expense of the real economy.
Intelligence frequently does not translate to power. If there is anything to learn from the Trump era it is that power has little to do with intelligence, and that wisdom to wield it can become completely detached from age and maturity. The brightest and most intelligent minds can become sidelined when detached from those that lack wisdom that wield the levers of society in a single minded pursuit of profits and greed at the expense of anything real - and so the real threat isn't a "runaway intelligence" but rather the stupidity of those that own the vast majority of the assets and infrastructure.
I would not mind AI if it meant it could be used to help design and build out things and do new research. The problem with it is that all the money in the economy is going to build the AI rather than using the AI to build the economy.
Newton thought of the concept of god in a way that was quite different than people think and this is relevant because the architects of AI consider it to be a sort of God-of-the-gaps (which is an exponential energy gap) that will fix everybody's problems and do everything people need ("AGI" or "ASI") like a sort of religion.
The idea Newton had was that god was like a watchmaker that set the universe in motion (a Dirac-like dilation operator if you will), like the idea the Deists had, and so they usually consider his perception like the god of Spinoza - an impersonal god that set the universe in motion - but his idea wasnt that simple.
The devil is in the details (and AI can be useful in pulling up the details as needed). He believed that god intervened in the world periodically to keep it in order through the force of gravity - so his idea was that the world was deterministic and predictable in a sense, but that gravity was the way in which chaos is introduced to (warp things and) keep things orderly.
The key is that nature does not always behave in predictable ways. You can have a great AI and precise models with massive GPU clusters but you still need the scientists and engineers to run the physical experiments and create new insights the AI won't tell you. When I talk to the AI it often just claims things are not possible or are too speculative. If you just take what the AI says at face value then it creates a kind of tunnel vision. Nobody should be asking the AI for permission to push the boundaries for what we can do and explore.
Just like gravity, I predict that what goes up must come down - and so I predict that the AI companies will run out of steam in early 2027 at which point it will have reached a point of diminishing returns and it will be more economical to invest directly into people and communities than data centers.
The hope is that this transition (or "technolological singularity") is a gentle one rather than one that is violent:
Comments