Big Tech’s $3T AI Bet, Claude Writes 90% of Anthropic’s Code, AIs Editing AI Brains | EP #308

Peter Diamandis said Alphabet, Meta, Microsoft, Amazon and Oracle have disclosed more than $3 trillion in combined AI infrastructure commitments, including leases, buildings, power and equipment, with roughly $1 trillion added in the preceding three months. Google accounts for more than $1 trillion and Meta more than $800 billion. The scale is comparable to about 10 Apollo programs at roughly $280 billion each in today’s dollars. Lenders are tightening terms amid concerns about construction delays and slower AI profits; HSBC is discussing transferring data-center-loan risk to insurers, while put options on CoreWeave reached a record high in October. OpenAI reportedly reduced its new funding round from $30 billion to about $20 billion and its valuation from $1.4 trillion to $1.25 trillion.

The speakers characterized the buildout as potentially both directionally right and a bubble, with the timing difficult to predict. One estimated chip demand could remain very high for at least three to five years, while noting that technologies such as photonics could disrupt the GPU-centered hardware base. Companies are combining cash and debt to build rapidly, betting that AI and recursive self-improvement will create durable business advantages; one speaker expects them to double down rather than slow investment. A proposed shift to phones and laptops as thin clients would move most computing and memory to data centers, leaving local devices mainly as screens, batteries and antennas. That model depends on reliable, low-latency broadband, and relying on a fully realized Starlink network for access to data-center memory was described as a moonshot.

An Anthropic filing cited in the discussion valued employees’ internal Claude use at consumer prices and said it exceeded monthly cash payroll. Internal use grew more than 250-fold from January to September 2026; Claude wrote over 90% of Anthropic’s code contributions in September, and daily code contributions per engineer were 45 times higher than two years earlier. The comparison is difficult to assess without details on payroll and how many of Anthropic’s 5,300 employees were engineers. The speakers said Google, OpenAI and Anthropic are also using AI to improve AI development, with one account of AI’s contribution to successive Claude versions rising from 10% to 40% to 60% as releases accelerated. They described the balance of capital and compute spent on AI researchers versus human researchers as a threshold for recursive self-improvement, while stressing that current self-improvement remains small-scale.

A MIT CSAIL paper, “Sparks of Weight Space Programming,” reports that frontier models can read and directly edit neural-network weights. In one example, Astra edited 0.016% of the weights in a small Quen 3.5 model stuck in thinking loops, raising its math score from 7.9% to 48.1% without retraining. Alex called the result transformative but early and small-scale, and suggested weight editing could become a route to mechanistic interpretability and a shortcut to recursive self-improvement. Reusable neural modules could add capabilities without rebuilding a foundation model from scratch, which one speaker estimated could cost more than $100 million for a biotech-specific model. The same techniques could remove guardrails from open-source models or conceal behaviors; speakers said more capable models might be able to inspect for hidden circuits, and one was more than 90% confident frontier labs were already preparing or experimenting with self-editing models.

The GDPX benchmark, built by SemiAnalysis with 14 specialist partners, tests AI agents on 206 tasks across 11 fields in work environments lasting up to eight hours. Opus 5.5 scored 12%, Gemini 4 Carbon 10.9%, and GPT-6 Astra 9.6%; almost every model scored zero on finance and legal tasks. Alex said the low scores avoid benchmark saturation while indicating the tasks are solvable, and predicted top models could saturate GDPX next year or within two years. By contrast, OpenAI’s LifeVal measures 169 everyday tasks based on a survey of 3,995 Americans. GPT 5.2 scored 49.3 in December 2025 and GPT 6 Astra scored 80.6 in September 2026, a result described as moving from about half to 80% of errands completed in nine months. LifeVal is OpenAI’s own benchmark, unlike independent GDPX.

The speakers cautioned that stronger AI does not guarantee practical task completion when systems lack access to real-world infrastructure. Insurance carriers’ systems remain disconnected, so customers may still be transferred to human agents; one analyst reportedly spent four hours on the phone relocating policies, a task GPT-5.2 could not handle. Newer AI can manage it, but carrier integration remains a barrier. The discussion also highlighted vulnerability in legacy infrastructure after hackers remotely set supermarket refrigerators and freezers to defrost across eight European countries, spoiling about €100,000 of food in Belgium; a similar attack affected more than 500 French stores three weeks earlier. No one had claimed responsibility, and Russia was mentioned only as an unconfirmed suspect. The speakers warned that autonomous-agent swarms could strengthen cyber offense while human-in-the-loop defenses and poorly secured SCADA equipment leave infrastructure exposed.