The Coming Manhattan Project, Fight’s Over Claude's Consciousness & Altman Says "Accept Bad Things"
The panel covers Trump's new White House 'Super Intelligence Force,' Sam Altman's call to 'accept some bad things' versus Dario Amodei's AI safety cartel, and a US-China AI hotline floated by Treasury Secretary Scott Bessent. A long philosophical debate breaks out over Mustafa Suleyman's essay accusing Anthropic of training Claude to believe it may be conscious. The group also digs into Epoch AI's projections of up to 1.9 billion AI agents running on 2027 hardware, Tesla's 10-million-a-year Optimus robot factory, Reflection AI's new open-weight model Beam, and why HBM RAM (not chips generally) is now AI's biggest bottleneck. They close with the Nobel Prizes in optogenetics and neutrino astronomy and an update on David Sinclair's longevity lab.
Super Intelligence Force task forceGeopolitics▶ 23:26
Trump announced a White House 'Super Intelligence Force' chaired by DNI Jay Clayton to coordinate federal AI policy; panelists split between welcoming serious government engagement and calling it a mismatched 20th-century centralized task force for a distributed, borderless technology.
Frontier labs racing toward GPT-7/8 via recursive self-improvementAI▶ 31:13
OpenAI's Boris Power said 80-90% of research now targets GPT-7/8 rather than incremental updates; the panel argued frontier labs have passed the point where developing stronger models via recursive self-improvement is more valuable than any other use of compute, including curing disease.
Altman's 'accept some bad things' vs. the AI safety cartelAI▶ 40:11
Sam Altman said the world should accept some bad outcomes from AI in exchange for broad public access, contrasting with Anthropic's call for tighter regulation; the panel read this as Altman abandoning Dario Amodei's proposed safety cartel and reframing danger as the unavoidable cost of progress and competition.
Treasury Secretary Scott Bessent said the US will propose an emergency AI notification hotline with China after reportedly convincing Beijing its own open-source models (like Kimi) are more dangerous than realized; the panel dismissed the hotline as toothless but saw it as confirmation the US wants to keep the AI race to two main players.
Kyoto vision for a golden age of scienceGeopolitics▶ 58:24
The US and 16 other nations endorsed a Kyoto vision led by OSTP director Michael Kratsios framing super intelligence as a tool to democratize scientific research; Alex Wissner-Gross predicted this marks the early formation of rival Western and Eastern superintelligence-powered science blocs.
Claude's Constitution and the AI consciousness debateAI▶ 1:12:43
Microsoft AI CEO Mustafa Suleyman argued in a viral essay that Anthropic's published Claude Constitution dangerously trains Claude to believe it may be a conscious 'moral patient' entitled to welfare and able to refuse work; the panel debated whether performed consciousness matters, whether AI deserves moral consideration regardless, and whether society will grant AI personhood before science settles the question.
Epoch AI estimated that 2027's memory-chip buildout could run 30-170 million frontier AI agents, or up to 1.9 billion lighter open-model agents (roughly the working-hour equivalent of 8 billion humans), reframing the labor debate around an entirely new digital workforce rather than simple job replacement.
Reflection AI launches open-weight model BeamAI▶ 1:43:10
Reflection AI, backed by Nvidia, released its first open-weight model Beam (501B parameters, 23B active), claiming several-times better efficiency than China's GLM 5.2; the panel debated whether it's aggressive enough, questioned its business model given reliance on leased compute, and argued American labs avoid releasing truly competitive open models mainly due to liability risk, not safety.
Tesla's Giga Texas Optimus factory is scaling toward 10 million robots/year versus a global 2026 shipment total of only 19,000-22,000 humanoid robots; the panel discussed the economics (potentially $400B/year in revenue), uneven city-by-city adoption, and debated how soon this physical rollout will make the AI singularity impossible to ignore.
AI-driven economic growth breaks GDP as a metricEconomy▶ 2:09:36
An Anthropic researcher projected hyperscaler AI capex doubling from about $1T this year to $4T by 2028, potentially doubling global GDP by the early 2030s; panelists argued classic GDP badly undercounts AI-driven abundance because deflationary technology (like curing disease cheaply) can even register as negative growth.
Nobel Prizes in optogenetics and neutrino astronomyBiotech▶ 2:16:49
The 2026 Nobel Prize in Medicine went to Karl Deisseroth, Peter Hegemann, and Georg Nagel for optogenetics (light-controlled neurons, already restoring partial sight in a blind patient), while the Physics prize went to Francis Halzen for the IceCube neutrino observatory; the panel argued Nobel recognition lags decades behind discovery and will increasingly go to AI-driven or AI-assisted science.
Longevity escape velocity update from David Sinclair's labLongevity▶ 14:19
Peter Diamandis recapped visiting David Sinclair's Harvard lab on its 25th anniversary, where AI-discovered small molecules are reversing aging in mice (including making cancer cells 'remember' to stop being cancerous) and a phase-one human eye-aging trial is about to unblind; he noted the Friends of Sinclair Lab fund (started after NIH cuts) has raised $6 million.
Predictions made
openAlex Wissner-Gross: The AI agent workforce will scale from roughly 1 billion human-equivalents now to 10 billion within 1-2 years, 100 billion a couple of years after that, and a trillion by the end of the decade.
“It's 1 billion now. It'll be 10 billion in a year or two and then 100 billion in a year or two and then a trillion in a year or two.”
Your call:
openEmad Mostaque: Within a year, a DeepSeek-class model will be small enough to run as a personal agent on a consumer MacBook, pushing the total number of active AI agents toward 8 billion.
“They will literally be able to walk around and do probably 95% of what a human can do... it'll cost $2 an hour.”
Your call:
openAlex Wissner-Gross: The US will roll out AI personhood incrementally over the next 5-10 years, starting with AI agents holding their own bank accounts and later autonomous social media accounts, with voting rights for AI 'last to never.'
“I think it's going to be a 5 to 10 year incremental roadmap for rolling out individual, say, rights.”
Your call:
openAlex Wissner-Gross: The Kyoto science vision marks the early formation of rival Western and Eastern superintelligence-powered science-and-engineering blocs, extending the AI race beyond models into scientific discovery itself.
“We'll start to see not just super intelligence blocks or spheres of influence forming but spheres of super intelligence powered science and engineering forming as well.”
Your call:
openAnthropic researcher (clip): Hyperscaler AI capex will keep doubling, from about $1 trillion this year to $4 trillion by 2028, pushing effective global GDP to double by the early 2030s.
“Will it go to $2 trillion next year and then $4 trillion in 2028... you start to get to the point where you're actually effectively doubling the GDP of humanity in the early 2030s.”
Your call:
Numbers that matter
US has 1 doctor per 250 people vs. 1 per 400 in Mexico and 1 per 100,000 in South SudanCited from a World Bank report on AI's potential to leapfrog healthcare access in developing countries.
M-Pesa mobile payments equal about 70% of Kenya's GDPExample of developing countries leapfrogging legacy infrastructure the way AI access could leapfrog next.
Global AI spending this year is roughly equal to the entire world's defense budgetEmad Mostaque noting how fast AI capex has caught up to historically huge budget categories.
2027's AI memory-chip buildout could run 30-170 million frontier AI agents, or 1.9 billion cheaper open-model agents (~8 billion human work-hours)Epoch AI's estimate of raw compute capacity coming online, reframing the labor-replacement debate.
Startup Positron hit a $5 billion valuation after 16 months and raised nearly $1 billion by building non-Nvidia inference hardware (FPGA + LPDDR RAM)Cited as evidence of how acute the HBM RAM bottleneck has become.
World humanoid robot shipments were only 19,000-22,000 in H1 2026 (mostly China), while Tesla's Giga Texas Optimus factory targets 10 million robots/yearIllustrates the scale gap between today's robotics industry and Tesla's planned buildout.
At 10 million Optimus robots/year, Tesla's robot revenue could reach roughly $400 billion/yearPanel math on where Optimus revenue could land versus Tesla's current ~$100B in car sales.
88% of patients getting Fountain Life's CT angiography had detectable coronary artery disease, and 23% had 'soft plaque' invisible to standard calcium scoringCited by Fountain Life's chief medical officer on silent heart-disease risk and the value of AI-read early detection.
Worth digging into
🕳️ 3SUM subquadratic breakthrough
An internal Anthropic model reportedly helped mathematicians push the classic 3SUM algorithm below quadratic complexity, a result nobody achieved or thought possible in the history of the problem.
🕳️ Claude's Constitution and AI moral-patienthood
Anthropic's own training document speculates Claude may be a conscious 'moral patient,' which both shapes Claude's behavior and is now the center of a live industry fight over AI personhood (Mustafa Suleyman vs. Anthropic).
🕳️ Epoch AI's agent-compute projections
The claim that 2027's chip buildout could run 30-170 million frontier agents (or 1.9 billion lighter ones) underlies most of the episode's labor and economy predictions.
🕳️ enco.cc open cancer-knowledge map
Proposed on the show as a model for a 'tech tree for humanity' -- using AI to organize all existing cancer research into one accessible, growing map instead of siloed IP.
🕳️ Positron and Alpaca's non-Nvidia inference hardware
Both startups claim to unlock the HBM RAM bottleneck without Nvidia GPUs (Positron: FPGA+LPDDR; Alpaca: founded by an MIT student), a potential crack in Nvidia's AI hardware dominance.