2026-09-22

Ask the Mates Anything Round #2 | MOONSHOTS AMA #293

In this fan AMA, Peter Diamandis, Alex Wissner-Gross, Dave Blundin, and Salim Ismail take listener questions on AI safety, education, careers, and the future of abundance. Recurring threads: use AI to monitor AI instead of locking capabilities in labs, treat AI as cheap enough that funding is no longer the excuse to not start, and capture AI-driven productivity gains through equity rather than salary. The panel debates whether college still matters (yes for friendship and deep thinking, no for its curriculum) and gives concrete career pivots into manufacturing/hardware, agent management, and AI-security defense. They close on personal probability estimates for whether AI ultimately creates abundance for humanity, both landing above 90% over a 10-year horizon, while flagging the next few years as the highest-risk window.

▶ Watch on YouTube

Topics

AI safety via mutual AI monitoring and transparency AI ▶ 0:07
Dave and Alex argue the real fix for AI safety is designing AI to monitor AI's actual thought processes transparently, and gradually exposing new capabilities to the real world instead of locking them in labs.
LLMs as a weak 'upload' of the human mind AI ▶ 3:49
Alex cites a Meta research result linking GPT-2's hidden activations to human fMRI voxels as evidence that language models trained on internet behavior already compress a weak form of human mental state.
Bias to action: build without waiting for funding Economy ▶ 8:36
Answering a question about funding an AI education project for kids in developing countries, Alex argues 'how do I get funding' is the wrong question now that AI tools are cheap and accessible enough to start within an hour.
Superintelligence, abundance vs. loss-of-control risk, and P(fab) AI ▶ 10:07
A listener pushes back hard on Diamandis-style optimism, asking what evidence would change the panel's mind about abundance. Alex and Dave defend it, and later give personal probability estimates (both over 90% on a 10-year horizon) that AI ultimately creates a much better world.
Diversity in AI-era governance; Ireland as a testbed Geopolitics ▶ 15:13
Dave argues the worst outcome is one or two governance models dominating the world; AI should instead let thousands of governance experiments run in parallel, with small, open countries like Ireland as ideal testbeds.
Self-replicating manufacturing for lunar and space industry Space ▶ 19:34
Alex and Dave answer an aspiring exoplanet-manufacturing engineer: the real bottleneck to 'dismantling the Moon' is building a self-sufficient local industrial ecosystem (mining plus manufacturing), ideally a self-replicating factory using lunar resources.
Career pivot: hardware, manufacturing, and physical AI Robotics ▶ 22:58
Dave tells computer-science people to move into manufacturing now because software has matured but hardware needs people for a long time, and advises a robot food-delivery founder to add drone delivery and to study VC networks via PitchBook.
Capturing AI productivity gains requires equity, not salary Economy ▶ 29:39
Responding to a founder whose AI-driven productivity gains aren't translating into income, Dave argues nearly all the value from AI automation flows to shareholders via capital gains, so the fix is to become an owner, not to ask for a raise.
AI-assisted filmmaking: story-first, iterate against hit-film patterns AI ▶ 31:49
Peter says story is still the key to a successful AI-made film: use a model to analyze the top-grossing films' common elements and iterate your own story against them. Dave adds that China's flood of AI micro-dramas shows it's now cheap to test thousands of story variants before committing to one production.
The future of higher education in the age of AI Other ▶ 35:07
Prompted by a father questioning whether to push his son toward college, the panel agrees college's curriculum value is collapsing but its social and deep-thinking value endures; Alex wants top research universities broken up and rebuilt as for-profit public enterprises, and Salim says universities' 'immune system' against change is nearly as bad as religion's.
Agentic AI cybersecurity risk and board-level disclosure AI ▶ 50:03
A healthcare CISO says an open-weight model found vulnerabilities in 10 minutes that $50,000 of human pentesters missed. Dave and Alex advise leading with data and framing, and call preemptive AI-security defense the business opportunity of the century.
The agentic web: AI-native interfaces replacing scraping AI ▶ 54:19
Dave argues every customer-facing website now needs an AI-native ('XML for agents') interface instead of relying on slow, error-prone scraping, citing Jeff Bezos's push for an internal API-first view of Amazon as the model to follow.
Vision restoration: gene therapy and brain-computer interfaces Health ▶ 56:28
Answering a listener losing his sight to a hereditary retinal disease, Peter points to David Sinclair's AAV/Yamanaka-factor gene therapy (Phase 1 human trials) and BCI work; Alex adds that BCI-based vision augmentation could deliver superhuman eyesight, including UV/IR vision, within a few years.

Predictions made

open Dave Blundin: Humanity will overcome the near-term dangers created by the AI arms race (misuse, China/Russia competition, AI-enabled weapons) within about 5 to 10 years.
· due: 2031-2036 · ▶ watch
“I believe that humanity will be able to overcome these risks within the next 10 years, or perhaps even within 5 years.”
Your call:
open Alex Wissner-Gross: The probability that AI ultimately creates a much better, more abundant world for humanity (P-fab) is over 90% on a time horizon beyond 10 years.
· due: 2036+ · ▶ watch
“on a time scale of over 10 years, I think E zoom T is over 90%.”
Your call:
open Alex Wissner-Gross: AI's usefulness for autonomously coordinating human work will last only about 10 more years, after which humans will need to converge more closely with machines to remain economically relevant.
· due: 2036 · ▶ watch
“the period during which AI can usefully and productively coordinate human activities will likely be very short.”
Your call:
open Alex Wissner-Gross: AI-assisted discovery of software vulnerabilities will spike for roughly 18 months, peak, then decline as the most critical zero-days get found and patched.
· due: 18 months (~Q1 2028) · ▶ watch
“it would be reasonable to expect that numerous vulnerabilities will be discovered over the next 18 months.”
Your call:
open Alex Wissner-Gross: Brain-computer interface technology will give people superhuman eyesight, including vision beyond the visible spectrum, within a few years.
· due: ~2029 · ▶ watch
“I expect that you will have superhuman eyesight within a few years.”
Your call:
open Alex Wissner-Gross: As AI becomes more energy-efficient than humans (higher IQ-per-watt), it may eventually out-compete humans for the solar system's solar energy, economically pushing humans toward the outer solar system.
· due: unspecified · ▶ watch
“AI will become increasingly energy-efficient, and from an economic perspective, at some point in the future, AI may be able to use the solar energy produced within the solar system more efficiently than humans.”
Your call:

Numbers that matter

Worth digging into

🕳️ The GPT-2 hidden-activation vs. fMRI voxel study Alex cites for the 'weak upload' claim
Alex leans on a specific Meta research result to argue LLMs already encode compressed human brain states; the claim deserves checking against the actual paper's scope.
🕳️ David Sinclair's AAV/Yamanaka-factor gene therapy (ER-100) for vision restoration
Uses reprogramming factors to reverse age-related gene expression for Leber hereditary optic neuropathy and macular degeneration, now in Phase 1 human trials -- a concrete near-term therapy worth tracking.
🕳️ Mercor as a template for managing AI-agent workforces at scale
Dave calls Mercor the fastest-growing company in history and cites it as proof that people-management skills transfer to managing large fleets of AI agents.
🕳️ The Phase 3 longevity drug referred to on-air as 'Silicon Medicine'
Peter cites a molecule in Phase 3 trials claiming to extend lifespan 3-6 years in a roughly 100-person study -- a striking, checkable longevity claim if the company can be identified.
🕳️ 'Solve Everything' and 'Accelerando' as maps for navigating the AI transition
Alex and Peter point to their own book plus Charlie Stross's 25-year-old novel as the best available frameworks for the current moment -- worth reading against the panel's actual on-air claims.