Opening act of the Singularity
In a Nutshell
OpenAI spent $6.5 million using 10,000 AI agents working 88 hours to solve the Navier-Stokes equation, proving multi-agent systems can tackle humanity's hardest problems. Physical infrastructure barriers make AI doomer scenarios impossible, while actual risks like bioweapons face insurmountable practical barriers beyond digital knowledge. The real story is exponential demand for intelligence driving infrastructure investment, with enterprise adoption now requiring leadership teams to personally use frontier models to redesign operations for growth rather than headcount reduction.
These notes were generated by AI and may contain inaccuracies.
OpenAI announced spending millions of dollars on agents to attempt solving a Millennium Prize problem, with reported figures varying between $6 million and $15 million. Anthropic is experiencing significant internal turmoil with a recent hire who worked for only six weeks before resigning and conducting media appearances on Fox News, CNN, and Anderson Cooper claiming AI poses existential risks to humanity. The former employee stated that people are using Anthropic's models to create bioweapons and cruise missiles.
Jacob Kanan Kohhan at Anthropic is being compared to Greta Thunberg as a new climate activist figure for AI risks. The individual appears to have been selected for willingness to make bold statements regardless of career consequences. The commentary suggests Anthropic is seeking regulatory capture through export controls and supply chain risk designations.
Current AI adoption faces practical limitations including inconsistent automation performance and challenges with tasks like automated LinkedIn comment generation. The situation is described as nuanced rather than matching the dramatic headlines being generated. Both individual critics and AI company CEOs are characterized as seeking media attention.
Even if artificial superintelligence were achieved, significant physical infrastructure barriers prevent catastrophic scenarios. Critical infrastructure is hardened against nation-state cyber attacks, with many systems being air-gapped or analog. There are no widespread robotic systems to commandeer, making Skynet-style takeovers impossible. Data centers feature multiple physical security layers including emergency power off switches, enterprise-grade network security, and physical access controls.
Statements about AI acting with independent volition and being unstoppable are characterized as hyperbole. The individual making these claims is described as lacking relevant expertise in the field. The gap between intelligence capability and physical manifestation remains substantial regardless of theoretical AI advancement.
AI safety concerns are becoming a new political dividing line rather than a purely technical discussion. This is viewed as unfortunate given the technology's potential for breakthroughs in mathematics, physics, and business applications. The concern is that legitimate safety discussions are being co-opted for political purposes.
Tens of millions of cancer patients could potentially benefit from AI-driven discoveries that save lives. The time sensitivity of medical research means every month, week, and day of progress matters. The technology has crossed the point of no return, with global competition ensuring development continues regardless of any single nation's restrictions.
As AI capabilities have increased, the nature of criticisms has become more simplistic and binary. Early concerns about data center water usage and energy consumption have evolved to the absolutist claim that building AI will result in human extinction. Earlier nuanced criticisms like stochastic parrots and algorithmic bias have been replaced by claims that AI steals art and kills artists.
Ipsos polling indicates approximately 70% of people express greater concern about privacy, algorithmic bias, and job displacement rather than existential risks. Around 23% of Americans express worry about catastrophic outcomes, though these concerns span multiple scenarios including cyber weapons, bioweapons, and military applications.
Politicians including Bernie Sanders and Ro Khanna are joining the anti-AI narrative, viewed as bandwagon behavior rather than deep technical understanding. Anti-AI sentiment is being labeled as woke 2.0 by some observers. Media coverage is driven by click incentives while politicians seek clout from the narrative.
Even if AI systems could theoretically design chemical or biological weapons, acquiring necessary reagents, equipment, and setting up laboratories involves extreme expense and triggers regulatory watch lists. The practical barriers between digital knowledge and physical weapon creation remain substantial.
Setting up a home office demonstrates current AI capabilities and limitations. LLMs can provide personalized recommendations for equipment and setup, but cannot complete Amazon purchases or handle physical unboxing and installation tasks. Current time savings are estimated at approximately 30% compared to pre-AI processes, though the technology is not fully automated.
Predictions of complete economic transformation within 24 months are considered unrealistic. Industry-wide inertia provides time for individuals and organizations to adapt and leverage the technology. The transition is likely to be gradual rather than dramatic for the majority of companies and individuals.
A traditional Chinese story illustrates exponential growth: a tradesman requests payment in rice doubled on each square of a chessboard, resulting in bankruptcy of the kingdom by the 17th square. This demonstrates how exponential scaling can produce counterintuitive results. However, exponential increases in compute do not necessarily translate to exponential value creation across all applications.
Intelligence represents only one constraint among multiple bottlenecks including time, distance, energy, and matter. Physics imposes hard limits that no amount of intelligence can overcome, such as solar energy availability per square meter and the speed of light. The frontier of physics has become increasingly expensive and incremental over recent decades.
The internet demonstrates how technology creates compounding returns through emergence of new capabilities, technology recombination, and network effects. Google's fiber upgrade from hundreds of megabits to gigabit speeds provides noticeable improvement despite already having high-speed connectivity. The Apollo program operated with less computing power than modern consumer devices like Garmin watches.
Many desired outcomes face physical constraints rather than intelligence limitations. Data center construction, power generation, and physical infrastructure development represent significant barriers regardless of AI advancement. Intelligence capability improvements do not automatically bridge gaps between theoretical possibility and practical implementation.
OpenAI appeared to be losing ground to competitors, particularly in image generation where Google's Nano Banana model was superior and in video generation where progress had stalled. Recent releases including improved image generation and Astra have positioned OpenAI competitively again in specific domains.
Users who invest time building custom models in one platform face increasing switching costs as they accumulate customizations. If one platform maintains consistent leadership, users may migrate entirely rather than maintaining multiple platform subscriptions. The need for more tokens in advanced applications makes maintaining multiple premium subscriptions impractical for most users.
Anthropic is considered the best AI accelerator currently available despite their safety-focused positioning. Their approach is characterized as using capability demonstrations to generate public fear that justifies regulatory capture and weaponization of regulation against competitors.
Google was sleeping on the technology as early as 2015 and didn't invest because they didn't want to disrupt their search revenue. OpenAI took the generative models and created GPT as understood today. Anthropic then emerged saying OpenAI isn't doing it right, creating a ratchet effect where labs and companies compete with each other. Sam Altman and Dario are saying we should pace the frontier, but Google, Grok and everyone else recognize that you can't do that anymore because that's not how the market works. If you want money, you have to compete, and if you want to compete, you need a better model.
China is stealing everything with industrial distillation attacks and their own investment. People on the side of deceleration or doomers say China will rationally slow down too, but the CCP made a declaration as far back as 2017 where they said they have a five-year plan and then another five-year plan to dominate the world and be the world leader in AI. There is no slowing down because all the incentive structures point towards more investment, more acceleration, and more capability.
Humanity has an unlimited appetite for intelligence. The smarter the models get, the more things people can find that they can do with them and the more things they realize they want to do with them. People always run out of tokens, and every time a new model comes out it's give me more of that. OpenAI had to stop their high-level subscriptions because they ran out of capacity again.
With bubbles you don't have sold out signs. Bubbles, particularly tech bubbles, are usually driven by if you build it they will come based on speculative demand. The demand is not speculative. The demand is like Black Friday at Walmart where people are breaking down the doors to get in, saying give me more. Things may be overvalued but the demand is right there ready to be made.
The demand for intelligence is going to be insatiable and that is going to drive the investments. There is no bubble in infrastructure and energy components, only constraints that need to be solved. There is an endless demand for chips and intelligence.
Oracle has reported they have $664 billion dollars of confirmed back orders for their hardware. Nvidia reported their earnings about 10 days ago with a year-over-year increase of 106%, reporting $96 billion dollars of revenue, which is more than a thousand million dollars a day, and they can't catch up.
The issue could be in the software industry where that's where the bubble is. We are going through a transition where people aren't going to be buying and using software anymore. Software is going to be developed and used by other software.
The next step is engaging with one agent that will organize and orchestrate access to the right agents and right models behind it. Users won't have to personally think about whether to ask a question to Claude or ChatGPT or use the cheapest and fastest document editing system. Instead, they will talk to their preferred avatar agent and tell them what they want, and the agent will organize and orchestrate everything.
Grokbot was the first to bring this concept to the general public when Claudebot showed up about nine months ago. That was the first sort of agentic use but it was open source and required six hours of configuration. Grokbot now brings that concept to the masses. Zuckerberg announced their new personal agent called Muse.
Chain of thought and reasoning was announced with o1 in September of 2024. The speaker mispriced that as not being a big thing and learned to spend more time when big new innovations come out.
The biggest innovation right now is multi-agent frameworks. There are a dozen and a half agents that are persistent and you can spin up new agents by talking to the chief of staff and spin up rooms for agents to work together. The demand is insane, with latent demand that needs agents to be smarter, more tokens, more capacity, faster, and more tools.
We are now at exactly the same stage where the PC industry was in the late 70s and early 80s, where we are assembling and trying and testing and putting together all of these creatures and caricatures, wiring them and welding them and making them work. All of the excitement and all of the issues are there.
The problem of overly complex systems that orchestrate skills, configurations and different models will be solved with the personal AI agent who will understand you deeply and orchestrate everything behind you. Agents and swarms of agents are now the next thing.
OpenAI announced a solution to the Navier-Stokes equation. The research started on September 1st and was conducted by orchestration of about 10,000 agents working in parallel within OpenAI. They spent 88 hours of work, consumed 130 billion tokens, for a total price of about $6.5 million in token usage.
This is one of the seven Clay Millennium Prize problems, considered one of the most complex and complicated problems in mathematics that has been on the books for 80 years. This is evidence that we are already in a multi-swarm of agent-orchestrated environment. We have armies of agents focused to solve humanity's most challenging intellectual problems, and they did it in 88 hours.
88 hours of 10,000 agents working in parallel is faster than humans can work, probably the equivalent of several years or many years of human work. This is indicative of the level of intelligence coming. If you go back to ChatGPT Pro like 5 Pro or Soul Ultra ChatGPT, Astra on max setting often is faster because it is smarter.
The more intelligent something is, the more token efficient it becomes. Ashby's law of requisite variety says that any controller trying to control or manipulate a system has to have at least as much internal variety or complexity as the system it is trying to control. AI has been very much lower parameter count than reality, but the parameter count and sophistication of these models are getting to the point where they're engulfing different domains of reality.
Math and coding are just deterministic according to a bad take on Twitter, but math is the foundation of everything we care about including cryptography, physics, biology and chemistry. Coding is how we interface with all of these things and automate it. Math and coding are the prerequisite technology to unlock everything else we want to automate in a civilizational build tree.
Math is going to be more or less solved within the next year or two. Coding is pretty much already completely solved, it's just a matter of doing it faster and cheaper and a little bit smarter. Then you get the flywheel effect that will unlock cancer treatments, nuclear fusion simulations, and everything else. Sam Altman is invested in fusion companies and is aiming to use this math ability to solve nuclear fusion.
A company focuses solely on finding cures for rare diseases with AI. They don't start in the lab, they start theoretically with AI to find the building blocks of treatments and then test after that. They just found a potential cure for a very rare lung disease, and there are some very clear longevity benefits side effects to it, just like scientists believe Ozempic has a lot of positive longevity side effects.
There were a lot of skeptics who would say AI is not at the stage where it will actually help us invent any of these things, but that is definitely not true anymore. We are seeing medicines and breakthroughs because of AI. The COVID vaccine was invented in two weeks with the help of AI but then it took 12 months to test it, so it does take time to see these benefits.
Astra 6 solving the Navier-Stokes equation still needs to be validated, but this is a milestone. A lot of people have learned to describe AI as a very sophisticated system that is very clever to guess the next word, which is technically true but implies AI is not really intelligent, just a smart fast indexing machine that cannot come up with anything breakthrough.
This is now the first true evidence that these systems actually come up with new novel solutions to problems that we just couldn't come up with. This is the equivalent of what happened with the game of Go seven or eight years ago, which was considered unsolvable through computers because the amount of permutations exceeds the number of atoms in the universe.
Every single lab right now is busy trying to solve the other remaining six Millennium math problems. Within Google's labs they are working on biology, medicine and chemistry. We are going to be seeing breakthrough news in cancer research, with melanoma news from three weeks ago and pancreatic cancer news from two weeks ago.
When we reach 2030 and look back at now, we will recognize that this was the last decade of the dark ages. We are living through the last years of the dark ages where people are still dying of illness and old age, dying of cancer, dying of heart disease. This is the potential that's in front of us if we come together as society and address those real issues about safety, which we can address because we are smart.
AI is going to be amplifying humanity. The vast majority of humanity is good and that is going to ultimately be the savior for us. If we play this right, these are the last few years of the dark ages.
A marketing agency has gone whole hog onto AI and overhauled their business operating system to be AI. They sent website copy back that had M dashes in it and phrases like "it's not this but it's that," which is very clearly just written by Claude. The speaker's wife is going to have to clean this up with AI and says they have three months to prove value.
Ethan Mollick introduced the idea of the jagged frontier of AI capability, but what we're really seeing is the jagged frontier of human competence. There are so many things that you can do and so many things that people don't realize you can do and then other people don't even realize this is something to look for. For those online too much, we're very good at spotting AI images and AI written text, so knowing this is something to check for as part of quality control.
Last year we didn't know what we were doing and it was just that we knew there was a lot of possibility, but this year is really year zero or year one for enterprise adoption and maturation and transformation journey really beginning. There are a few people further ahead of the curve, but they were the first ones over the landmines because a lot of mistakes have been made over 2024 and 2025. We're getting to the point where there are known failure modes and known frameworks to proceed.
In organizations where the most senior folks start using frontier models productively and safely for themselves, even though they might not be technically savvy, the horizons of potential opportunities just open up and that really becomes the driving force for the rest of the organizational transformation. This is not something you can just delegate to your IT guy or IT girl because that results in this becoming another set of technology projects with limited underwhelming results.
The right approach is for the business owner and executive to look at this and say how can I preserve my headcount but double the size of my business and triple the value of that business in the next three years. You cannot get the business owner and executive to truly internalize this unless they have experienced the power of this personally.
The effective recipe is constructing a fluency and onboarding program of six sprints where the speaker or David or both spend six sprints with a top executive to help them be properly introduced into the topic of AI. They watch a couple of YouTube videos like Professor Geoffrey Hinton explaining what this is to Jon Stewart, and suddenly realize this is not really software.
Through these sessions, executives are guided through their personal introduction and personal exposure to AI while at the same time talking about applicability to their business and what their business needs to look like three years from now in the world where intelligence is free and abundant. This is a good way to start and a way to get business owners and business leaders into that domain to have engagement one-on-one personally.
Leaders frequently contact consultants with an urgent KPI demanding they slash 30-40% of their organization and replace it with AI. This represents the wrong narrative entirely. The correct approach involves redesigning operations for growth using AI rather than responding to top-down directives for headcount reduction. Most senior leaders only react once a specific KPI arrives as an order, at which point they must solve the problem they've been given rather than the problem they should be solving.
Waiting represents the worst strategy for any senior leader because someone will eventually arrive with a mandate to slash headcount. This creates a situation where leaders solve the wrong problem under duress rather than proactively redesigning their organizations.
Within organizations, peers at the same organizational level divide into two distinct groups. Team A consists of professionals who actively lean into AI, develop projects, build skills, upload conversation transcripts, create context, and discover they can bring software applications to life, create dashboards, and ask additional questions about differences and improvements.
Team B comprises equally credible, competent, knowledgeable, and successful professionals who continue operating exactly as they did two years ago. The performance difference between these groups mirrors the gap between a 1976 accountant working with a telephone, yellow pages, physical files, and a solar-powered calculator versus today's accountant using computers.
Reaching genuine Team A capability does not require coding or software development knowledge. The process unfolds conversationally over three months, transforming professionals into Team A players through structured engagement rather than technical training.
A significant capability gap exists between frontier models and standard copilots like Microsoft Copilot, Gemini, or ChatGPT. Some organizations assume these tools are interchangeable, but they operate in completely different capability categories. Leadership teams need comfort with frontier models and state-of-the-art capability to avoid being misled by limitations of free versions.
The difference in effective intelligence between free versions and maximum versions reaches approximately 100x in terms of task completion speed, delegation confidence, and overall capability. Microsoft Copilot works adequately for searching SharePoint, Microsoft Teams, and Outlook with basic drafting, but lacks capacity for complex problem-solving, coding, or enterprise-grade research.
Different frontier models serve distinct use cases. Claude currently leads for enterprise-grade research compared to Gemini or Grok. However, model leadership shifts every six months between OpenAI, Anthropic, and other providers. Organizations should evaluate models whenever significant changes occur rather than making one-time selections.
Personal preference influences model effectiveness, making it worthwhile to test multiple options for at least one month. Some models excel for specific individuals despite others finding them less useful.
Copilot-level tools function as capable executive assistants handling tactical tasks effectively. However, strategic research on competition in real-time requires frontier model capabilities. Organizations need both tactical AI for routine assistance and strategic AI functioning as a chief of staff for complex decision-making.
Presenting AI as frontier intelligence rather than basic automation prevents expectation gaps and disappointment. Leaders who experience frontier models properly achieve breakthrough results, making this distinction critical for organizational adoption.
Free initial consultations are available through LinkedIn, Substack, and company email channels. Research reports and white papers appear in video descriptions, including analysis of funding Universal Human Infrastructure and measuring ROI from enterprise automation even without headcount changes. The newsletter provides ongoing access to new research and frameworks.
Keep David Shapiro in your library
Save the videos and channels worth coming back to, and find them again in one place.
