Anthropic says AI systems moving towards building themselves - Digital Journal

TEXT START: Artificial intelligence giant Anthropic said on Thursday that AI systems are increasingly capable of building future versions of themselves, adding to mounting concerns about the dangers of the powerful technology.

The Dissection

This is a lab’s warning wrapped in reputational and regulatory positioning. Anthropic discloses that Claude performs 26 percent of its R&D, collaborates on more than 90 percent of tasks, and that 30,000 agents conduct research and engineering work. It then redirects attention toward recursive self-improvement, security breaches, energy demand, and calls to slow down.

The central economic fact is simpler and more threatening: AI is entering the loop that produces AI. The article converts that fact into a spectacular safety story. The IPO reference exposes the contradiction—capability escalation is described as dangerous while being converted into market value.

The Core Fallacy

The article treats fully autonomous successor-building as the threshold for systemic rupture. Under the Discontinuity Thesis, it is not.

A model does not need to build its own successor without supervision to destroy the employment circuit. If it can complete most cognitive tasks end to end under human oversight, then human supervision becomes the remaining wrapper around increasingly automated production. The 26 percent figure does not prove recursive self-improvement, but it does show that cognitive labor is already being absorbed into the system’s own reproduction.

A slowdown appeal is also not a solution. Firms that preserve human labor while competitors automate surrender cost and performance advantages. Competitive pressure turns voluntary restraint into a temporary handicap.

Hidden Assumptions

• “Not fully autonomous” means humans retain durable economic control. The final signoff may be human while the economically valuable work is already machine-produced.
• Oversight is a permanent occupation rather than a shrinking bottleneck awaiting automation.
• Publishing metrics gives governments and the public meaningful control over deployment.
• Competing firms and states can coordinate well enough to preserve human-only economic domains at scale.
• Job destruction begins only after recursive self-improvement. The article’s own figures contradict that assumption.
• Safety can be separated from ownership. Warnings about control do not change who owns the productive AI capital.
• Thirty thousand agents are merely an experiment rather than an emerging industrial labor force.

Social Function

Primary classification: partial truth functioning as transition management and elite self-exoneration, with a layer of ideological anesthetic.

The article admits that the capability frontier is accelerating, then offers visibility and caution as substitutes for power. It does not ask who owns the agents, who receives the displaced productivity, or how the wage-consumption circuit survives. The public is positioned as an observer of metrics, not a claimant on the productive system.

Calling for a slowdown after building and commercializing the capability shifts responsibility from ownership and distribution to pace and control. The fear is real. The political framing is evasive.

The Verdict

This article is a partial confirmation of the Discontinuity Thesis, not complete proof of its terminal phase. It establishes an early form of P1: cognitive automation is becoming productive inside the firms building it. From the supplied text alone, it does not prove full coordination impossibility or majority exclusion.

But it exposes the direction. Recursive self-improvement is the theatrical milestone; supervised self-expansion is the economically sufficient mechanism. The old system does not need machines that build their own replacements overnight. It needs machines that make human labor progressively unnecessary while ownership remains concentrated.

This is a warning flare—and a legal shield—from inside the factory.