Kinzey Capital Management: Anthropic AI Slowdown Plan

Contributed content: this article was written by a third-party contributor and does not necessarily reflect the views of ABC Money. Editorial and Advertising Policy

Anthropic’s Dario Amodei proposes independent safety audits and shared protocols with authoritarian states to pace frontier AI as OpenAI adopts embedded evaluators, sharpening questions on liquidity, infrastructure payback and concentration

  intelligence development, setting out a three-step plan in a 3,800-word essay published on 12 September. Independent auditors would verify safety standards, democratic nations would coordinate industry rules and authoritarian governments would agree to shared protocols. Amodei warns that autonomous systems could take over the entire internet within six months to a year at the current pace, with damage running into the hundreds of billions of dollars. Kinzey Capital Management Pte. Ltd. reads the plan less as a verdict on AI than as a test of the obligations each pool of capital must meet.

Amodei argues that AI has accelerated sharply since the summer, driven by the technology’s growing capacity to build its own successors. His first step asks each frontier AI company to give embedded third-party evaluators, METR among them, ongoing employee-level access. Coordination with authoritarian states, the hardest step, would begin with agreements that serve every party, such as a ban on using AI to develop biological weapons. A meaningful pause needs the US and China to act simultaneously, and AI training is easier to conceal than missile silos, making enforcement harder than nuclear arms control.

Anthropic commits to that first step unilaterally and with immediate effect, granting evaluators permanent, employee-level access to its systems. They may publish findings without Anthropic’s editorial control. OpenAI chief executive Sam Altman says pacing the frontier has already been a primary topic of discussion inside his company, which will adopt the same embedded-evaluator commitment. Much as bank regulators have long stationed supervisors inside the institutions they oversee, the evaluators would audit internal AI deployment, information security and safety decision-making as well as public products.

Jacob Coxon’s resignation from Anthropic on 15 September, after three years of pretraining research at OpenAI and Anthropic, exposes the unease inside the laboratories. His post argues that neither company is acting responsibly, describing both as racing straight to self-improving superintelligence. Evan Hubinger, Anthropic’s safety lead, backs the post and puts his own probability of AI killing all humans within the next decade above 10%.

Liquidity is the first test for money that must stay reachable, since institutional portfolios periodically sell holdings to meet capital calls. A broad revaluation would squeeze the liquidity cushion just as obligations fall due, marking “the point at which a paper loss becomes a forced sale”, according to David Nilson, who serves as Director of Private Clients at Kinzey Capital Management. More than 50% of S&P 500 sectors show AI-linked exposure at present, while the Philadelphia SE Semiconductor Index emerges from this week’s sell-off still up nearly 60% over more than eight months of trading.

Investors with decade-long horizons weigh the proposal against the roughly $858.8 billion in AI capital expenditure by hyperscalers over the past two years. Billy Leung of Global X Management argues that pacing AI leaves spending on chips, power and infrastructure intact while extending the timeline over which it pays.

Joint accounts face a coordination problem the market does not resolve, with every holder jointly liable for fees, taxes and penalties. Investment decisions typically require agreement from all owners, so where one co-owner favours reducing AI exposure and another prefers to hold, the structure offers no mechanism for a split decision. Creditors of either owner can freeze the entire balance.

Concentration magnifies each of these effects, as actively managed equity funds have raised their AI weighting from 9% to 14% over roughly three years. The market value of those positions has doubled over the same span, and US equity investors have roughly half their capital in technology stocks at current index weights once Amazon, Tesla, Alphabet and Meta are counted. The US currently makes up more than 70% of the global developed market, so global portfolios carry almost the same tilt and one sector’s revaluation drives the whole account.

The arithmetic hardens once withdrawals begin, since a $38,168 withdrawal equals 4% of a $954,210 portfolio at the moment it is drawn. The same sum equals 6.67% once that portfolio falls to $572,526, and sequence-of-returns risk concentrates in the first decade of retirement. Kinzey Capital Management treats the plan’s coordination thresholds as a range of outcomes, not a settled course, with Nilson framing the question as “how much of that range each obligation can hold before the audits reach full depth”.

Kinzey Capital Management, on the Record

Singapore-based Kinzey Capital Management oversees discretionary multi-asset portfolios for private clients, companies, families and foundations, treating shares, bonds, funds and cash as a single book. Every portfolio is set by what its capital must achieve, when it may be needed and how wide a range it can absorb, with instruments chosen to match. Reporting measures each portfolio against its assigned duty.

Growth Portfolios, Income and Withdrawals, Corporate Reserves and Joint and Family Accounts continue while the underlying duty lasts, whereas Concentrated Shareholdings and Second-Opinion Reviews are one-off assignments around existing holdings.

Kinzey Capital Management Pte. Ltd. is registered under UEN 202105652G.

https://kinzey.com

Chloe Lim handles press enquiries at c.lim@kinzey.com.