2026-04-30

Google Invests $40B Into Anthropic, GPT 5.5 Drops, and Google Cloud Dominates | EP #252

The panel covers an extraordinary week of AI releases (Kimi K2.6, GPT-5.5, DeepSeek V4) arriving at a pace of roughly two major models per week, and debates whether US labs still hold a real lead over Chinese open-weight models. They dig into Google's compute dominance (8th-gen TPUs, a 960,000-GPU A5X cluster, ~25% of world AI compute) alongside Anthropic's huge new funding deals with Google ($40B) and Amazon ($33B), framing compute and energy — not algorithmic breakthroughs — as the real bottleneck, with TSMC/Samsung/Intel fabrication capacity as the ultimate constraint. Other threads include AI surveillance and ambient computing privacy risks (OpenAI's Chronicle, Microsoft's abandoned Recall), the rising deepfake/identity-fraud arms race (World ID, a Zoom deepfake-detection push, a Grok-generated fake French woman with a reflective ID), the Musk v. OpenAI trial opening in Oakland, and the UAE's push to run 50% of government on agentic AI within two years. A long biomedical segment showcases ChatGPT for clinicians, AI-assisted organ-transplant matching, a pancreatic cancer mRNA vaccine with striking survival results, a single-shot CAR-T therapy curing melanoma, and AI-driven drug repurposing. The episode closes with robotics/transportation updates (a table-tennis robot, Tesla's Cybercab entering production, Joby's first NYC air-taxi flight) and an AMA on the future of work, consulting, P(doom), and universal income.

▶ Watch on YouTube

Topics

AI model race acceleration (15 releases in 8 weeks) AI ▶ 3:34
The panel discusses the dizzying pace of frontier model releases (Kimi K2.6, GPT-5.5, DeepSeek V4 among 15 major releases in 8 weeks) and debates whether the real winners will be whoever builds the best abstraction layer over the models rather than the model providers themselves.
Kimi K2.6: trillion-parameter open-weight MoE model AI ▶ 14:02
Moonshot AI's Kimi K2.6 is a 1-trillion-parameter, 32B-active mixture-of-experts model that natively processes text/image/video, was trained for a reported $4.6 million, and runs at a fraction of closed-model API cost -- discussed alongside how MoE/sparsity works layer by layer.
GPT-5.5 release and benchmarks AI ▶ 24:53
OpenAI's GPT-5.5 shipped 7 weeks after 5.4 with big gains in long-context reasoning, token efficiency, and lower hallucination; the panel reads it as intended largely to strengthen Codex, and highlights Frontier Math Tier 4 gains implying ~1%/month progress on research-grade math.
Google Cloud and 8th-generation TPU dominance Compute ▶ 31:33
Google unveiled 8th-gen TPUs (TPU8T for training, TPU8i for inference) with 3x training speed and 80% better performance-per-dollar, and Sundar Pichai reported 16 billion tokens/minute processed and 75% of Google's code now AI-written.
Google's A5X supercluster and Vera Rubin GPUs Compute ▶ 34:35
Google committed to 960,000 Nvidia Vera Rubin GPUs for its new A5X bare-metal instance, delivering 10x lower inference cost and 10x higher token throughput; the resulting cluster is described as 2x the size of Colossus 2 and 2.4x the size of Stargate Abilene.
Anthropic's mega funding deals with Google and Amazon Economy ▶ 36:48
Google committed $40B total to Anthropic ($10B now at a $350B valuation, $30B more on performance targets, plus 5GW of TPU compute over 5 years); separately Amazon is investing $33B total in exchange for Anthropic committing $100B+ in AWS spend and running Claude on Trainium chips.
Compute, energy, and the TSMC bottleneck Compute ▶ 40:52
The panel argues the real constraint on the entire AI race is semiconductor fabrication capacity (TSMC, Samsung, Intel) and, in the near term, energy/power permitting -- and that almost nobody besides Elon Musk discusses the fabrication bottleneck publicly.
Elon vs. Sam/OpenAI trial begins in Oakland Geopolitics ▶ 54:43
Jury selection began in the Musk v. Altman/OpenAI federal trial in Oakland; the panel discusses the two-phase trial structure, reports of political influence in jury selection, and what a loss could mean for OpenAI's nonprofit-to-for-profit conversion and its invested capital.
AI surveillance and ambient computing privacy AI ▶ 59:29
OpenAI's Chronicle takes continuous screenshots processed by cloud agents to build a memory of the user's activity; the panel compares it to Microsoft's abandoned Recall product and argues this functionality should be built securely into the OS/hardware rather than sent to remote servers.
Deepfakes and identity verification arms race AI ▶ 1:05:48
Rising video-call deepfake fraud (a $25M Hong Kong wire-fraud case) is driving projects like World ID's Zoom integration (retina-scan Orb + face/video verification) even as Grok can generate a fully convincing deepfake person holding a reflective, realistic ID.
Token maxing and the AI spending economy Economy ▶ 1:12:34
Startup CEOs bragging about huge AI compute bills ('token maxing') are debated as either a warped vanity metric or a healthy sign of aggressive AI adoption; Dave Blundin argues for targeting roughly a 1:1 ratio of token spend to human payroll.
UAE's agentic government at scale Geopolitics ▶ 1:16:55
Sheikh Mohammed bin Rashid announced UAE will run 50% of all government sectors and services on agentic AI within two years; Salim Ismail describes direct work with the Prime Minister's office and cites a golden visa issued in 5 hours as an example of the pace this authority structure enables.
Biomedical and longevity breakthroughs Longevity ▶ 1:19:59
A run of medical AI/biotech stories: free ChatGPT for clinicians outperforming human doctors on HealthBench, an AI tool (TopHeart) to reduce wasted donor hearts, a pancreatic cancer mRNA vaccine with strong 6-year survival results, a single-shot CAR-T therapy curing melanoma, a repurposed blood-pressure drug fighting MRSA, and David Fajgenbaum's AI-driven drug-repurposing nonprofit Every Cure.
Robotics and autonomous transportation Robotics ▶ 1:45:33
A table-tennis-playing robot (ACE) beats human opposition using multi-camera vision; Tesla's driverless, $30,000 Cybercab entered production aiming for 2 million units/year; and Joby Aviation flew its first NYC air-taxi demonstration flight from JFK to Manhattan in about 7 minutes.

Predictions made

open Alex Wissner-Gross: At the current pace of roughly 1% per month improvement on Frontier Math Tier 4, essentially all professional research-grade math problems will be solved.
EP #? · · due: 4-5 years from April 2026 (~2030-2031) · ▶ watch
“even just at the present pace, we're talking about essentially all frontier math tier 4, all professional research grade math problems being solved in the next four or 5 years.”
Your call:
open Salim Ismail: Humanoid robots will reach mass-scale, widespread adoption.
EP #? · · due: five to seven years minimum (~2031-2033) · ▶ watch
“I think humanoid robots are five to seven years away minimum at in in mass at mass scale uh in widespread adoption.”
Your call:
open Salim Ismail: A consistently operating fabrication lab on the Moon will exist.
EP #? · · due: fifteen years minimum (~2041) · ▶ watch
“I think a fab lab on the moon and consistently doing fabrication and all that stuff is 15 years away minimum.”
Your call:
open Peter Diamandis: Fabs on the Moon doing manufacturing and launching material to Earth orbit via mass drivers will be operational.
EP #? · · due: roughly 15-20 years (~2041-2046) · ▶ watch
“Well, okay. Maybe 15, but it's not the next 5 years.”
Your call:
open Dave Blundin: Anthropic will hit between $40 billion and $70 billion in annual revenue, per a board-meeting conversation with an Anthropic investor.
EP #? · · due: end of 2026 · ▶ watch
“Anthropic under the covers is thinking they might hit between 40, 50 up to 70 billion in revenue by the end of the year.”
Your call:
open Alex Wissner-Gross: Ray Kurzweil's version of the singularity -- superintelligence collectively smarter than all of humanity, which Kurzweil pegs at 2045 -- will be reached well ahead of that schedule.
EP #? · · due: well before 2045 · ▶ watch
“I think we're going to hit that so far ahead of of 2040.”
Your call:
open Alex Wissner-Gross: The current circular-economy-looking compute/cash deals concentrated among the top 10-12 AI and hyperscaler companies will diffuse throughout the broader economy.
EP #? · · due: next few years · ▶ watch
“that's going to diffuse throughout the economy over the next few years would be my prediction.”
Your call:
open Alex Wissner-Gross: If deepfaking causes severe enough societal problems, market demand will unlock existing but currently unused technological solutions, including hardware-level cryptography for cameras that verifies video/image chain of custody.
EP #? · · due: unspecified (conditional on deepfake harm reaching a societal breaking point) · ▶ watch
“if ever this the situation of deep faking gets so bad that it's causing real problems at a societal level that'll just unlock all of these technological solutions.”
Your call:
open Dave Blundin: AI-driven clinical diagnostic tools will face huge regulatory and professional 'immune system' backlash as they roll out.
EP #? · · due: unspecified, 'over the next few years' · ▶ watch
“just let's my prediction is huge uh regulatory and immune system backlash on this one.”
Your call:

Numbers that matter

Worth digging into

🕳️ Noam Brown's 'weights matter less than compute for reasoning' thesis
If inference-time reasoning scale genuinely dominates raw model weights, it reframes the entire AI competitive landscape as a pure compute race rather than a weights/IP race -- with major implications for open-weight models and export-control policy.
🕳️ Anthropic's Project Deal internal AI marketplace
It's an early, real-world test of letting an AI autonomously buy, sell, and negotiate on employees' behalf inside a live organization -- a preview of agentic commerce before it hits the broader economy.
🕳️ UAE's two-year agentic-government rollout
This would be the first national government attempting to run the majority of its services on agentic AI at civilizational scale, serving as either a template or a cautionary tale for other states -- including whether Western democracies can or should attempt anything similar.
🕳️ Deepfake fraud vs. identity-verification arms race
Deepfake fraud losses are projected to reach $40 billion by 2027, and the panel expects an escalating 'white hat / black hat' arms race between video-generation tools (like Grok's fake reflective ID) and verification tools (World ID's Orb/Zoom integration, hardware-level camera cryptography).
🕳️ David Fajgenbaum's AI-driven drug repurposing (Every Cure)
The mismatch between 18,000 recognized diseases and only 4,000 FDA-approved drugs is a striking, tractable gap that AI-driven matching could close quickly, and Fajgenbaum's own citizen-science origin story (curing his own rare disease with repurposed rapamycin) is a compelling proof point.
🕳️ Elon vs. Sam/OpenAI trial in Oakland
The outcome could unwind OpenAI's conversion from nonprofit to for-profit and put hundreds of billions of dollars of invested capital in question; the panel also flags reported political influence in jury selection as notable.