The Coming Manhattan Project, Fight’s Over Claude's Consciousness & Altman Says "Accept Bad Things"

The U.S. Super Intelligence Force signals an effort to accelerate AI development, while Scott Bessent’s proposed emergency notification process with China would create communication during what a speaker describes as an ongoing arms race—not a deceleration pact. The discussion also points to China’s potential ability to direct millions of Huawei SN GPUs and open models toward selected targets.

The speakers debate whether broad access or stronger restrictions better balance AI’s benefits and risks: Sam Altman favors public access despite possible harms, while Anthropic favors stronger regulation. They also argue that open-weight models are already circulating, making outcome-focused guardrails more workable than trying to prevent their use, though downstream liability may discourage frontier labs from releasing models openly.

Public opposition is already constraining AI development through resistance to data centers and proposed U.S. House bills targeting recursive self-improvement. In response, speakers favor adaptive safeguards such as incident reporting and accountability, while arguing that risk assessments should map potential harms rather than let fear of maximum regret rule out ambitious projects.

AI investment is drawing capital from companies such as Nvidia and AMD and large investors including Blackstone and MGX, with one speaker saying global AI spending is nearing the global defense budget. An Anthropic researcher reportedly put hyperscalers’ AI-related capital expenditures at about $1 trillion this year, following four or five years in which AI compute capacity grew twofold or threefold annually.

Compute scarcity is showing up in both steep access costs and higher component prices: a 72-GPU system that cost $3.5 million a year earlier was reportedly available only through a three-year, $9 million lease, while RAM prices were said to be up 10x. Speakers attribute the constraints to concentrated capacity and demand from higher-priority AI applications.

Speakers describe faster AI research as a potential accelerator of model progress: AI-driven research has compressed the intervals between models to 10 or 11 days, although one speaker cautions that compute, energy, memory, and evaluation still constrain recursive self-improvement. A separate speaker warns that companies adopting current models risk being overtaken by models described as 10x better arriving in three months.

An internal Anthropic model reportedly helped researchers improve an algorithm previously believed to require quadratic complexity, producing a result of 1.9995; the model proposed changes that human researchers analyzed and documented. That research focus aligns with Boris Power’s reported estimate that 80% to 90% of OpenAI research is aimed at training GPT-7 and GPT-8.

The speakers contrast publicly supported open science with proprietary AI research, disagreeing over whether AI will displace intellectual-property-based science or leave proprietary work viable. They also argue that private funding can support riskier research than NIH and NSF grants, citing a potential pool of $1 trillion to $2 trillion from U.S. foundations and wealthy donors and Mike Lazeritz’s funding of nine Toronto-area quantum-computing and photonics labs.

One speaker proposes that government buy 10% to 20% of compute at market rates for distribution, while another warns that market forces may steer GPUs toward consumer sales rather than goals such as curing diabetes. The proposal comes amid reported hyperscaler AI-related capital expenditures of about $1 trillion this year and a broader debate over whether universities and governments can counterbalance commercial allocation.