2026-06-29

Anthropic vs. Alibaba, OpenAI Delays Its IPO, and the US Government Blocks GPT-5.6 | #267

An emergency weekend episode covering the US government's first-ever national security hold on commercial AI, which forced Anthropic to restrict Mythos/Fable 5 and OpenAI to throttle GPT-5.6 to a small list of approved customers. The panel debates whether this protects the US or hands China an opening, discusses Anthropic's accusation that Alibaba ran a massive distillation campaign against Claude, and covers OpenAI's decision to delay its IPO despite a looming trillion-dollar valuation target. Additional threads include GPT-5.5 Cyber ('Daybreak') shifting AI from offense to defense, ByteDance's SeaDance 2.5 video model threatening Hollywood, a quantum computing executive order and funding round, a possible photonic-computing breakthrough that could replace GPU data centers, and an orexin-based drug (following Eli Lilly's $6.3B Synthesa acquisition) that could let people function on 4 hours of sleep. The episode closes with an AMA covering UBI economics, moon vs. orbital data centers, and post-human forms.

▶ Watch on YouTube

Topics

US government gatekeeps frontier AI releases Geopolitics ▶ 6:54

For the first time in US history, the executive branch placed a national security hold on commercial AI, restricting Anthropic's Mythos/Fable 5 to ~100 approved companies and OpenAI's GPT-5.6 to ~20 approved companies, with the White House now involved in customer-by-customer release decisions.

GPT-5.6 launch and the regulatory 'endgame' AI ▶ 10:35

OpenAI announced three GPT-5.6 tiers (Sol, Terra, Luna) roughly matching Mythos on cyber/bio benchmarks. Alex Wissner-Gross argues government throttling has become the coordination mechanism frontier labs failed to establish among themselves, synchronizing OpenAI and Anthropic's capability releases.

China's open-weight models and the harness advantage AI ▶ 13:58

Emad Mostaque describes how GLM 5.2, run through a strong multi-model harness (Blitz's Frontier SWE harness), already outperforms GPT-5.5 on the hardest coding benchmark despite costing far less compute, suggesting harnesses now matter more than base model gaps.

Anthropic's Mythos red-team exercise and Project Glass Wing Geopolitics ▶ 26:17

Mythos, red-teaming for the US intelligence community under Project Glass Wing, reportedly broke into almost all classified government systems within hours (per Senator Mark Warner), which the panel debates as the real or pretextual reason for the government clampdown.

GPT-5.5 Cyber ('Daybreak') and the AI defense-offense shift Compute ▶ 33:34

OpenAI's new defensive cybersecurity model scored a record 85.6 on the CyberGym benchmark; Sam Altman frames the real prize as automatically patching vulnerabilities across infrastructure, raising the question of who is trusted to merge AI-written code fixes.

OpenAI delays its IPO Economy ▶ 41:44

OpenAI is reportedly choosing to stay private rather than go public below a $1 trillion valuation, weighing a 2026 IPO under $1T against waiting until 2027. The panel debates whether SpaceX's stock volatility, cash needs, leadership structure, or SEC disclosure burdens are the real driver.

ByteDance's SeaDance 2.5 and AI-generated film AI ▶ 1:07:44

ByteDance's SeaDance 2.5 generates 30-second, 4K videos referencing up to 50 inputs (characters, video, audio, cinematographic controls). Emad Mostaque, ex-CEO of Stability AI, says this fulfills his 2023 forecast of Hollywood-level, full-length AI movies by 2027.

Real-time AI video and the enterprise/gaming opportunity AI ▶ 1:14:58

Panel discusses the video game industry (hundreds of billions) as a bigger prize than Hollywood ($50B), with China ahead on 'world model' style physics understanding, while Alex Wissner-Gross argues real-time interactive video-to-video AI (citing Alibaba's 'One Streamer' demo) could unlock a $30 trillion enterprise market.

Anthropic accuses Alibaba of a massive Claude distillation campaign Geopolitics ▶ 1:19:31

Anthropic alleges Alibaba used 28.8 million fraudulent exchanges across 25,000 fake accounts to extract and copy Claude's capabilities via distillation. The panel debates whether this constitutes theft/espionage or is simply how China's economy has always operated, and whether it becomes the pretext for broader Chinese AI suppression.

Trump's quantum computing executive order and funding Compute ▶ 1:30:37

Trump signed executive orders to supercharge US quantum computing and shield quantum research from foreign espionage (treating it like nuclear secrets). $2 billion in CHIPS Act venture funding was allocated: IBM ($1B), D-Wave/Rigetti/Inflection ($100M each), PsiQuantum ($140M).

Photonic computing as the next compute leap Compute ▶ 1:40:19

Dave Blundin says highly quantized neural nets now perform on par with FP32 nets, opening the door to photonic compute at roughly 1/100th the mass of equivalent Nvidia GPU compute, which he expects SpaceX/Elon Musk to begin deploying within a year to 18 months.

Orexin drugs and the short-sleeper opportunity Longevity ▶ 1:45:14

Following Eli Lilly's $6.3 billion acquisition of Synthesa Pharmaceuticals (targeting the orexin wakefulness pathway), the panel discusses orexin-based drugs as a potential lifestyle treatment letting most people function on 4 hours of sleep, following the GLP-1 playbook of a narrow drug expanding into a mass-market therapy.

AMA: UBI economics, moon vs. orbital data centers, post-humans Economy ▶ 1:52:50

Audience questions cover why UBI dividend math from AI-company equity doesn't work at scale, whether companies will flee to jurisdictions like Argentina to escape US AI restrictions, why the moon may beat orbit for data centers (security, mining feedstock), and what physical/legal forms humans and AI persons might take post-singularity.

Predictions made

open Alex Wissner-Gross: The capability gap between Chinese open-weight models and the two remaining Western frontier open-weight models is on a trajectory to shrink to zero.

EP #? · · due: Christmas 2026 · ▶ watch

“the number is on a trajectory to go to zero by Christmas of this year”

Your call:

open Emad Mostaque: Open models (like GLM) will close the gap with closed frontier models given the current rate of improvement.

EP #? · · due: December 2026 · ▶ watch

“definitely by December, um but maybe even now”

Your call:

open Emad Mostaque: The US government will ban Chinese open-weight models from use by US corporations and require licensing and KYC (including prompt retention) for access to frontier or new-frontier models.

EP #? · · due: unspecified/near-term · ▶ watch

“I think that there's a good chance that the US government bans Chinese open weight models from being used by corporations and requires a license and KYC for any frontier or new frontier model including retention of your prompts”

Your call:

open Alex Wissner-Gross: Real-time, 4K-quality AI-generated interactive video games will become achievable.

EP #? · · due: 2027 ('next year') · ▶ watch

“I think what you've just described David is next year. Real-time 4K video games.”

Your call:

open Dave Blundin: The video game market (currently a couple hundred billion dollars) will grow 3x to 10x once real-time AI-generated interactive video is cracked.

EP #? · · due: unspecified, following the real-time breakthrough · ▶ watch

“I wouldn't be surprised to see it three, four, five, 10x”

Your call:

open Dave Blundin: SpaceX/Elon Musk will begin launching or building out photonic compute infrastructure to replace Nvidia GPU-based data centers.

EP #? · · due: December 2026 to mid-2027 (within 12-18 months) · ▶ watch

“I would be shocked if by the time we're talking to Elon next December, I'd be shocked if we're not talking about launching... the terrafab started on the photonic compute at about 1/100 the mass”

Your call:

open Peter Diamandis: Elon Musk will go on an acquisition spree, buying companies broadly, now that SpaceX has liquid public stock.

EP #? · · due: unspecified/near-term · ▶ watch

“He's going to be on acquisition rampage. Uh watch him buy companies left, right, and center.”

Your call:

open Emad Mostaque: The EU AI Act has a good chance of being repealed following the fallout from the Anthropic Fable/Mythos restriction episode.

EP #? · · due: unspecified · ▶ watch

“So I think there's a good chance the EU AI gets repealed.”

Your call:

open Elon Musk (clip): Neuralink may attempt the first direct human-to-human telepathic (brain-to-brain) communication.

EP #? · · due: later in 2026 · ▶ watch

“Elon announced that Neuralink may make an attempt later this year for the first direct human-to-human telepathic communications”

Your call:

open Elon Musk (clip): Full-length motion pictures generated entirely by AI will emerge.

EP #? · · due: end of 2026 · ▶ watch

“Elon said by the end of this year, full-length motion pictures coming out of AI. Hollywood is cooked.”

Your call:

Numbers that matter

28.8 million fraudulent exchanges across 25,000 fake accounts Anthropic's alleged scale of Alibaba's distillation campaign against Claude, which it calls the largest AI model theft/acquisition ever if the claim holds up.
GPT-5.5 Cyber ('Daybreak') scored 85.6 on the CyberGym benchmark Reported as the highest single-model score ever posted on that real-world cybersecurity vulnerability benchmark.
GLM 5.2 represents roughly $25 million worth of training compute Cited by Emad Mostaque to illustrate how cheaply an open model reached top Frontier SWE performance with the right harness.
SpaceX stock fell from a high of $202/share to $153 (IPO priced at $135, opened at $150) Used by the panel to debate whether SpaceX's stock volatility was a legitimate reason for OpenAI delaying its own IPO.
OpenAI recently raised $120-122 billion Cited as the real reason OpenAI doesn't need to rush an IPO for cash.
OpenAI is forecast to lose $26 billion this year Noted alongside OpenAI's $40-50 billion revenue run rate as evidence the company can absorb losses without an imminent IPO.
$2 billion in US quantum computing venture investment via the CHIPS and Science Act; IBM got $1B, D-Wave/Rigetti/Inflection each got $100M, PsiQuantum got $140M Breakdown of Trump administration quantum computing funding announced alongside the new executive orders.
A $3,000/month UBI for US residents alone would cost $12 trillion/year versus a $7.4 trillion total US federal budget Peter Diamandis's math illustrating the scale problem with a US-only UBI.
Funding basic UBI purely from AI-company dividends at a 5% yield would require roughly $10 trillion in AI company value, and covering even half of that would require owning about half of the biggest AI companies in the world Emad Mostaque's calculation showing why dividend-funded UBI doesn't scale.
Sleeping 6 hours or less regularly is linked to a 48% increase in coronary heart disease risk, 15% increased stroke risk, 12% increased all-cause mortality, 5% increase in beta amyloid, 17% higher type 2 diabetes risk, and 4x higher risk of catching a cold Health section stats on the danger of chronic sleep deprivation.
Eli Lilly acquired Synthesa Pharmaceuticals for $6.3 billion Synthesa's orexin-targeting drug (originally for narcolepsy) is framed as a potential future 'lifestyle drug' for reduced sleep needs.
Sleeping only 4 hours instead of 8 yields 28 extra hours/week, 58 extra days/year, and roughly 2 extra 'months' of life-equivalent time per year Peter Diamandis's math on the value of becoming a short sleeper.
Video game industry is a couple hundred billion dollars vs. Hollywood at roughly $50 billion; video-generation AI is estimated at $4-5 billion in revenue versus roughly 20x that for code-generation AI Market-size comparison used to argue where AI value is currently concentrated versus where it could expand.
Enterprise software is estimated as a $30 trillion market Cited by Alex Wissner-Gross as the market real-time interactive AI (e.g. an AI meeting participant) could ultimately unlock, versus a roughly $1 trillion opportunity in narrower video generation.
Photonic compute could deliver the same computation at roughly 1/100th the mass of equivalent Nvidia GPU-based compute (described as a conservative estimate) Dave Blundin's estimate for photonic computing efficiency gains over current GPU data centers.

Worth digging into

🕳️ Anthropic vs. Alibaba distillation/espionage claim

If substantiated, this would be the largest alleged AI model theft ever and could set a legal precedent for whether distillation via fake accounts constitutes espionage rather than just a terms-of-service violation.

🕳️ US government's customer-by-customer gatekeeping of frontier AI releases

This marks the first executive-branch national security hold on commercial AI products in US history, effectively making the government a release-approval body for frontier models.

🕳️ Photonic computing as a GPU replacement

Dave Blundin claims quantized neural nets now match FP32 performance, enabling roughly 100x mass-efficiency gains via photonic compute, which could undercut Nvidia's dominance within 12-18 months if true.

🕳️ Real-time interactive AI video (Alibaba's 'One Streamer' demo)

The panel frames this as the bridge from consumer video generation to a $30 trillion enterprise market (AI meeting participants), a much larger prize than either Hollywood or gaming.

🕳️ Orexin-based short-sleep drugs

Following the GLP-1 playbook, Eli Lilly's $6.3B Synthesa acquisition suggests a narcolepsy drug could become a mass-market lifestyle product letting people function on 4 hours of sleep, with major economic and health implications.

🕳️ Argentina's 'non-human AI corporation' legal regime

Raised in the AMA as a real, novel legal experiment allowing AI persons/humanless companies to be chartered, which intersects with the UBI funding problem and US AI export controls.