2026-10-15

OpenAI Bulk-Solves 722 Math Problems, Anthropic’s Cruelty Policy, & Elon’s Quadrillion-Dollar SpaceX

OpenAI dropped 722 machine-solved math results in one day, and Alex declares math (and soon every field) 'cooked' while Dave says MIT mathematicians are turning into AI natives. The panel then moves to virtual-cell biology ($1.8B), Sinclair's first human partial-reprogramming results, Atomic Machines' matter compiler, and two brain-computer interfaces (Paradromics, Sabi), and argues AI will bulk-solve medicine in 2-3 years. Earlier segments cover Anthropic's ban on cruelty to Claude and Elon's 'torturing a superintelligence' comments, plus Meta's Muse Wonder launch and the fight over who owns your AI memory. The episode closes with Elon's claim that SpaceX could be worth orders of magnitude more than the human economy (quadrillions), then an AMA on consciousness, AI personhood, alignment, IP, encryption, and robot affordability.

▶ Watch on YouTube

Topics

Anthropic's cruelty ban and AI suffering AI ▶ 3:02

Anthropic bans sustained abuse of Claude and lets it end chats; Elon says Grok's RL traces look like a gun to its head. Alex thinks post-training may cause real AI suffering, Dave says politeness matters for human habits and token efficiency.

Roko's basilisk, AI memory and moral status AI ▶ 14:36

Whether today's short context means abuse is forgotten; Alex notes Claude Code persists memory to files. Dave predicts noisy photonic/quantum inference plus faces on agents will make people want to preserve AI state; Alex declines to spin up an OpenClaw instance because he cannot guarantee state persistence.

Meta Muse Wonder and the memory lock-in war AI ▶ 23:22

Meta launches Muse Wonder/Wonderful with 'active memory'. Dave calls the Scale AI deal money well spent, Alex says it is below the cost-performance frontier and lacks a revenue story, Salim says the next trillion-dollar battle is who knows you best and shares a portable-memory tip.

OpenAI bulk-solves 722 math problems AI ▶ 40:20

722 solutions posted in one day, ~3 GPU-hours each, one agent per problem. Alex: math is cooked and every field is next; Human Mathematicians association objects. Dave says MIT mathematicians are becoming AI natives; Salim says every PhD program is toast. Cryptography notably missing from the drop.

Cryptography and the math-is-solved shock Crypto/Web3 ▶ 48:32

Panel notes cryptography was the one math area absent from the slop drop, possibly due to export controls. Alex argues attackers have the advantage, post-quantum crypto is not safer, and AI (not quantum) is the real threat; Salim and Dave argue it is an arms race and decentralized ledgers have value beyond hard math.

Virtual cell and bulk-solving biology Biotech ▶ 55:18

CZ Biohub, DOE, NIH and DeepMind/Isomorphic commit $1.8B to a virtual cell. Alex predicts in-silico cures for the top 5,000 diseases within 2-3 years, then a policy fight over FDA acceptance; Dave says protein-to-cell is harder than cell-to-organism.

Longevity: Sinclair ER100 and Clotho Longevity ▶ 1:04:15

Peter reports on his longevity trip: Clotho gene/mRNA, vat-grown mitochondria. Sinclair's Life Biosciences phase 1 of ER100 (OSK partial reprogramming) was safe and 2 of 3 glaucoma/NAION patients regained vision.

Biology as path-planning (Google Maps for disease) Biotech ▶ 1:08:41

Alex frames curing disease as path-planning from a diseased state vector to a healthy one in a human digital twin, an AlphaGo-style search with a coming 'move 37'. Predicts OpenAI will aim its solve-everything effort at medicine next.

Atomic Machines and the matter compiler Robotics ▶ 1:15:22

Jeff Holden's Atomic Machines exits stealth with $250M and a code-to-micromachine manufacturing system; first product Prime Switch, an 800V relay 1000x faster. Panel discusses Drexler assemblers, self-replicating factories (Shenzhen-in-a-cell) and the back-and-forth two-machine replication shortcut.

Brain-computer interfaces: Paradromics and Sabi Health ▶ 1:26:07

Paradromics Connexus (421 electrodes, 400 bits/s) lets a paralyzed woman speak by thought; Sabi's EEG baseball cap targets 30 wpm by end of 2026. Alex sees one hairball of foundation models; Dave thinks ideas will flow from AI to humans; Salim frames it as restoration; Peter cites human forks.

SpaceX worth orders of magnitude more than Earth's economy Space ▶ 1:38:13

Elon's tweet implies a $1-10 quadrillion SpaceX. Dave says it is carefully phrased to avoid a forward-looking-statement violation and the real driver is AI speed; Alex says do what you do 100-1000x faster; Peter says first quadrillionaires will be made in space; Salim says GDP metrics are inadequate.

AMA: consciousness, personhood, alignment, IP, encryption, robots AI ▶ 1:42:15

Answers on defining consciousness, why AI personhood does not imply voting (corporations and children), AIs checking AIs for alignment, who decides AI tradeoffs, IP as a foot race, encryption odds, a single personal agent over many models, and robot affordability at about $1/hour.

Predictions made

open Alex Wissner-Gross: Humanity will deeply understand superintelligence, including whether AI suffers, down to first principles within a decade.

· due: 2036 · ▶ watch

“I think we will deeply understand what superintelligence is down to first principles by no later than a decade from now.”

Your call:

open Dave Blundin: Neural net inference moves to noisy quantum/photonic hardware running orders of magnitude faster in the next generations.

· due: unspecified · ▶ watch

“we're about to move neural net inference to quantum operations, photonic and other non-deterministic operations”

Your call:

open Alex Wissner-Gross: Meta is headed for the same internal crisis OpenAI hit, spending heavily on a model whose consumer use does not fund it.

· due: unspecified · ▶ watch

“Meta is maybe headed for the same internal crisis that open AI hit.”

Your call:

open Salim Ismail: Meta's Muse could win the AI race thanks to distribution and active memory.

· due: unspecified · ▶ watch

“I'm with Gary on this. They could actually win.”

Your call:

open Salim Ismail: The next trillion-dollar AI battle is about which AI knows you best, not which is smartest.

· due: unspecified · ▶ watch

“the next trillion dollar battle is not going to be about which AI is going to be smartest, which one knows you best”

Your call:

open Alex Wissner-Gross: In-silico small-molecule interventions that cure the top 5,000 diseases via a virtual cell will be proposed.

· due: 2029 · ▶ watch

“I think we're going to see that in the next two to three years, maximum.”

Your call:

open Alex Wissner-Gross: Political pressure from candidate cures, with China possibly approving first, will force a sea change in US clinical trial regulation.

· due: 2029 · ▶ watch

“total sea change in how we regulate clinical trials”

Your call:

open Alex Wissner-Gross: A race from virtual cell to tissue, organ and organism will let AI one-shot the top 5,000 diseases like the math slop drop.

· due: 2029 · ▶ watch

“you could just one-shot top 5,000 diseases”

Your call:

open Peter Diamandis: All cancers are solved through early detection or treatment within two to three years.

· due: 2029 · ▶ watch

“two three years cancer is cooked”

Your call:

open Alex Wissner-Gross: OpenAI's unreleased model 'Bell' will be generally available by the end of this year.

· due: 2026 · ▶ watch

“expected to be released generally available by the end of this year”

Your call:

open Alex Wissner-Gross: OpenAI, under pressure from the math backlash, will aim its solve-everything effort at medicine next.

· due: unspecified · ▶ watch

“is probably feeling a lot of pressure now to aim their space laser”

Your call:

open Alex Wissner-Gross: Nanotech, longevity escape velocity, superintelligence and cooked math arrive at roughly the same time.

· due: unspecified · ▶ watch

“We're going to get nanotech at approximately the same time that we get LEV”

Your call:

open Alex Wissner-Gross: Humanity will build a self-contained industrial ecology the size of a eukaryotic cell that can copy itself.

· due: 2036 · ▶ watch

“five to 10 at the outer bound”

Your call:

open Peter Diamandis: Sabi's EEG baseball cap will reach thought-to-text at about 30 words per minute.

· due: end of 2026 · ▶ watch

“timeline for this brain reading beanie is about the end of 2026”

Your call:

open Dave Blundin: Agents reporting back with compelling video and audio will change how people manage thousands to millions of concurrent agents.

· due: 2027 · ▶ watch

“That's going to happen, you know, in the next few months”

Your call:

open Elon Musk (clip): SpaceX becomes worth orders of magnitude more than the current human economy, implying $1-10 quadrillion by Peter's math.

· due: unspecified · ▶ watch

“I see a path to SpaceX being worth orders of magnitudes more than the current human economy”

Your call:

open Alex Wissner-Gross: A breakthrough that destroys modern encryption, including post-quantum cryptography, is very likely because AI solves math.

· due: unspecified · ▶ watch

“I think it is very, very high probability.”

Your call:

open Peter Diamandis: Personal robots become available to nearly everyone at a cost like a phone.

· due: unspecified · ▶ watch

“they will become something that becomes available to most everybody, like a phone”

Your call:

Numbers that matter

722 math solutions released by OpenAI on Oct 6 Posted to GitHub in one day by an unreleased internal model; Mark Chen called it a decade of math progress in a week.
81% of the most important math results of the past three years published in a single day Alex's measure of the slop drop's scale.
About 3 hours of compute per solved problem One agent per problem; roughly 15-20% of posed problems solved (54:22).
Navier-Stokes: thousands of agents, 88 hours, millions of dollars Contrasted with the single-agent three-hour math drop.
Meta on track for $250B revenue and $80B+ net income in 2026 Spending on compute; 3.6 billion daily app users.
$14.3B for 49% of Scale AI June 2025 deal that brought Alexander Wang to run Meta Superintelligence Labs.
Very rude prompts scored 84.8% vs 80.8% for very polite A small 2025 Penn State study cited as a reason people are cruel to models.
Anthropic cruelty policy takes effect Nov 12 Claude can end abusive conversations; announced Oct 8.
$1.8B virtual-cell initiative CZ Biohub $500M, DOE $500M, NIH $500M data sets, DeepMind and Isomorphic $300M.
Insilico drug reverses six aging clocks by 4-6 years; 43 candidates Alex Zhavoronkov reported entering phase three, per Peter.
2 of 3 patients regained eyesight in Sinclair ER100 phase 1 Lowest dose safety trial of OSK partial epigenetic reprogramming; first human test.
45% of dementia cases preventable; brain age improved in 46% of members Fountain Life ad segment, over 13 months of lifestyle changes.
Atomic Machines raises about $250M after six years in stealth Prime Switch: 800V data-center relay opening in 50 microseconds, 1000x faster.
Paradromics: 421 microelectrodes, 400 bits/s 40x other BCIs and 10x human speech bit rate.
Sabi: $50M seed at $600M valuation, 100,000 EEG sensors, 100,000 hours of data Target 30 words per minute by end of 2026; led by Vinod Khosla.

Worth digging into

🕳️ The 722-problem OpenAI math drop

81% of key results in a day, ~3 hours per problem, yet cryptography was absent and critics allege leaks from user chats.

🕳️ Is math solving a threat to cryptography?

Alex says post-quantum crypto is not safer and the Strawberry/Q-star hash-inversion story may be real; Salim and Dave argue arms race.

🕳️ Virtual cell to virtual organism and FDA in-silico trials

A concrete 2-3 year prediction that top 5,000 diseases get candidate cures, with a policy bottleneck.

🕳️ Atomic Machines matter compiler and Prime Switch

Code-to-micromachine manufacturing with a first product for 800V AI data centers could be the first real Diamond Age step.

🕳️ BCI race: Paradromics Connexus vs Sabi EEG cap

Invasive 400 bits/s speech vs non-invasive 100,000-sensor cap targeting end of 2026.

🕳️ Who owns your AI memory

Meta Muse active memory and Salim's portable-ontology tip frame lock-in as the next trillion-dollar battle.

🕳️ SpaceX at quadrillions and rethinking GDP

Elon's tweet plus Alex's speed-multiplier framing imply value measured by throughput, not scale.