Google just put forty billion into Anthropic and the cloud wars look essentially over. Alex breaks down why this is really about maximizing economic value per token, why Google now controls roughly a quarter of all AI compute on the planet, and why TPUs are now designing TPUs in a recursive self-improvement loop that goes all the way down to the silicon. Then OpenAI counter-strikes with GPT 5.5: a massive jump on Terminal Bench 2.0 making Codex a real Claude Code competitor, plus 1% per month gains on Frontier Math Tier 4 meaning research-grade math is cooked in four years at current pace. Demis used to say five breakthroughs remained between us and AGI a decade ago. Alex's update from his Frontier Lab friends: it is half a breakthrough now, maybe zero. Plus: why Noam Brown thinks weights matter less than compute, why Anthropic is hunting compute everywhere, lunar AI data centers needing fewer regulatory approvals, fabs on the moon in maybe ten years not twenty, the OpenAI nonprofit-to-PBC trial getting jury-influence drama, Microsoft Recall as architectural atrocity, WorldCoin as Minority Report retina ID, professions cooked starting with law, mRNA vaccines as the actual Drexler nanorobots that finally arrived, and Henry Intelligent Machines turning every human into a tastemaker overseeing AI conglomerates. The singularity is not a point. It is an interval. We are right in the middle of it.
TIMESTAMPS
0:00 Cold open: Google commits $40 billion to Anthropic
0:09 GPT 5.5 drops, math is cooked
0:36 The three-way Western frontier race and the weights-versus-compute pivot
3:10 The average consumer is no longer in the equation
3:59 Where Kimi K2.6 and DeepSeek v4 actually fit
5:15 Sparsity as the path to a million-parameter diamond model
6:39 GPT 5.5 Terminal Bench 2.0 jump and Frontier Math Tier 4 cooked
9:20 Half a breakthrough left to AGI, per Demis
9:45 Google now controls ~25% of all AI compute, TPUs design TPUs
10:37 Powered land as the new strategic asset
11:36 The compute circular economy among the top ten labs
12:57 Lunar data centers and fewer regulatory approvals
13:26 Fabs on the moon in maybe ten years, not twenty
13:51 Anthropic's unifying principle: maximize economic value per token
15:31 OpenAI versus Anthropic as the next Sorkin Apple-vs-Microsoft
16:14 The OpenAI nonprofit-to-PBC trial drama and jury selection
18:40 Microsoft Recall and OpenAI Chronicle want to be in the OS
19:43 Privacy as a baked-in property of the silicon
20:30 WorldCoin as Minority Report retina ID
21:29 Hardware-level cryptography for camera chain-of-custody
23:08 The human-to-AI labor ratio trends to one-to-infinity
24:35 Professions cooked, GDPVal mapping all of knowledge work, law next
26:45 Of course this is about replacing doctors
27:32 Clinicians could not resist EMRs, they cannot resist superhuman AI
28:28 Clinicians hate EMRs but love AI, the surprising flip
29:02 Donor organs becoming unnecessary
30:09 mRNA vaccines as the actual Drexler nanorobots that finally arrived
31:57 CAR-T as the horse-and-buggy era of immunotherapy
32:36 Off-target effects as a secret weapon, drug repurposing on tap
33:29 Why AI table tennis took so long
34:13 Why did everything take so long
34:51 Henry Intelligent Machines and humans as tastemakers
36:50 Singularity as an interval, not a point
38:43 Revealed preference on superintelligence estimates
39:10 Where to follow Alex's daily takes: alexwg.org / Innermost Loop
