The ever-optimistic Peter Diamandis writes,
In the three months from November 2025 through January 2026, Americans filed 1.56 million new business applications: the highest for any three-month period since tracking began in 2004.
Solo-founded startups surged from 23.7% of all new ventures in 2019 to 36.3% by mid-2025. Four out of five entrepreneurs have now integrated AI into their operations.
In the future, human & AI collaboration will be the norm. But we know AI is fallible. Therefore, we should require students to document how they used AI, the prompts they employed, the outputs they received, and what they accepted, rejected, or modified and why. When students know their reasoning will be evaluated, they engage differently.
He has other good ideas, as well. I think there are two reasons for AI misuse by students. One is that students in general are motivated to get the credential without learning (Bryan Caplan has made this point often and well.) The other is that Google taught us to treat an Internet-connected computer as an AnswerBot: type in a query, and get a result.
I want the student to feel like he is working with a tutor. Your personal tutor is not someone you are trying to fool in order to get a good grade. And you don’t want your tutor to just hand you the answer. You want your tutor to help you gain a grasp on the concepts.
Here is the taxonomy I use. Two axes. How much autonomy does the agent have, from delegated to sovereign? And how broad is its scope, from specific to generalist? That gives four quadrants.
The first is the Specialist Tool. Narrow, task-specific, fully under human control. Price scrapers. Report generators. Automated data pipelines. Useful, proliferating rapidly, but commoditizing quickly. This is the robotic process authomatic of the agentic era. Necessary plumbing, but not where durable value accrues.
…The fourth is the Autonomous Corporation. The endgame. Fully independent entities that manage diverse operations, allocate capital, set long-term strategy, and hire other agents. An AI-run investment fund. A content studio with no human employees. A distributed manufacturing network of autonomous nodes coordinating through agent marketplaces. This is the furthest frontier and the most speculative, but also potentially the largest. If agents can create value autonomously, the addressable market is bounded only by energy and compute.
…The most asymmetric returns of the next decade will not come from building better copilots. They will come from building the identity, financial, legal, and governance infrastructure that a new class of economic participant requires to function.
A long essay, and I do not think I understand all of it.
Also difficult for me to grasp is Andrey Mir. But I think I get this:
The evolutionary meaning of the internet seemed to be
1) the transfer of all human knowledge and all human speech into the digital, where they became available to future AI, and
2) the rise of the high-tech industry capable of creating AI. Media evolution needs our digital activity for AI to learn from us, just as a meadow needs the activity of bees, so it has made digital platforms engineer our engagement.
Before you could have LLMs, you needed to have content in a digitized form, so that the AIs could learn from it.
substacks referenced above: @
@
@
@
One might usefully contrast Diamandis’ entrepreneurial explosion with the planned economy that the National Science Foundation is delivering and how NSF activities under the Bayh-Dole Act are narrowing the tax base and making it harder for unconnected private entrepreneurs to succeed.
Yesterday, there was a link Dr. Kling posted to an article about robotic delivery of health and social services. The NSF, unsurprisingly given its vast resources and unlimited scope to interfere in the economy, also is shaping our future in this regard. NSF’s National Robotics Initiative (NRI) 3.0 and the NSF SBIR/STTR Robotics and Digital Health programs are funding development robotics in home and clinical settings, in particular in home-based and long-term care settings.
And the approximately $120 million in funding so far has selected some winners. Diligent Robotics, founded by Georgia Tech professor Andrea Thomaz who received an NSF CAREER award, developed Moxi, a hospital logistics robot that has completed over 1 million deliveries in acute care settings. The company followed a standard NSF commercialization path, progressing from NSF research awards to the NSF I-Corps program and SBIR Phase I and II grants, eventually raising private seed funding.
Other notable commercialized health robotics startups include GuidaBot, a joint venture between the University of Houston and Fannin Innovation Studio, which received an NSF SBIR grant to develop a robotic manipulator for MRI-guided interventions using solid-media transmission. Additionally, PSYONIC developed durable, touch-sensing bionic hands for amputees, leveraging NSF funding to build its technology before appearing on Shark Tank in 2024 to expand its business.
Ain’t the free market grand?
More generally, the National AI Research Institutes program awards money to a plethora of mainly academic centers to get in the AI game. For example, the NSF and NVIDIA trecently gave a $152 million joint grant to the Allen Institute for AI (Ai2) for an Open Multimodal AI Infrastructure to Accelerate Science (OMAI) project. The Allen Institute previously was able to spin out NSF funded assets under the name Xnor.ai which it then sold to Apple for $200 million. Tax free. This is similar to how Stanford sold out its NSF funded interest in the google search algorithm for $300 million. Stanford still apparently receives licensing fees.
It seems that if we really wanted to see Diamandis’s vision realized we would repeal the Bayh-Dole Act and shut down the NSF.
In some respects, same as it ever was.
Back in my Econ undergrad, a management professor offered that while ethics would prevent him from following through, that he is willing to bet that he could make a lot of extra money by simply auctioning off As (or even Cs) in his class, making education (for many students) one of the rarest products: one that the consumer would pay more to get less of.
With AI, one doesn't need to bribe the professor, but in either case, those students unwilling to take the deal -- those there to actually learn materials -- will be advantaged over those willing to take the shortcut. This won't be evident based on CVs, but will be evident after someone is hired and work is underway.