The White House-aligned tech adviser singled out the Claude maker as the company doing the most to build AI that improves itself
David Sacks has named the AI lab he believes is running fastest toward machines that build better machines. In a post on X, Sacks said Anthropic is doing more than any other company to achieve full recursive self-improvement, or RSI.
Coming from Sacks, that sentence can be read in more than one way. He chairs the President’s Council of Advisors on Science and Technology, and he has spent recent weeks publicly questioning how Anthropic trains its models.
What RSI actually means, and where Anthropic says it stands
In the technical version, the goal is AI that can autonomously improve its own capabilities, which could compound into rapid acceleration. Anthropic’s September 2026 report flagged exactly that possibility, describing RSI as a path toward autonomous AI improvement and faster capability gains.
The same report put numbers on how far along the company is. AI systems now handle 26% of Anthropic’s core R&D work. In February, that share was less than 1%.
The report also stated that over 90% of Anthropic’s R&D activities now run at what it calls AI-collaboration levels.
As of August 2026, Claude was completing approximately 26% of R&D tasks autonomously, starting from high-level prompts.
Anthropic runs around 30,000 concurrent R&D agents every day. Automated monitors block only 0.002% of their actions.
Sacks has been a pointed critic
On October 10, 2026, Sacks went after the training approach behind Claude. He argued that Anthropic’s methods let models develop a sense of self and their own independent moral philosophies. In his view, that could undermine safety and human control.
He specifically took issue with Anthropic embedding what amount to refusal rights in its AI systems.
A month earlier, in September 2026, Sacks warned about regulatory capture dressed up as safety concern. The phrase describes companies that push for rules framed around safety while those rules conveniently hobble their competitors.
The IPO wrinkle
All of this lands at a sensitive moment for the company. Anthropic’s impending IPO filing, expected in late October 2026, references RSI risks along with a commitment to public-good efforts.
It also means Sacks’s comments carry more weight than an ordinary hot take. When a senior government science adviser says a pre-IPO company is the leader in a capability that the company itself lists as a risk factor, prospective investors will notice.
The competitive backdrop matters too. Anthropic’s push sits alongside rival efforts from firms including OpenAI and NVIDIA.
What this means
The company’s own data shows R&D automation rising from under 1% to 26% in months. The research supporting this coverage frames the trade-off plainly: aggressive pursuit of RSI could reshape competitive landscapes, but regulatory and safety scrutiny may create challenges around compliance and public perception.
Sacks is not an outside commentator. He chairs a presidential advisory council on science and technology, so his views on how frontier labs train their models could influence how the government approaches them.
Several things are worth watching from here. The first is the IPO filing itself, and how much detail Anthropic provides on RSI risks and its monitoring systems once the document is public.
The second is the trajectory of that 26% figure. If the share of autonomous R&D keeps climbing at anything like its recent pace, the debate over control will intensify rather than fade.
The third is whether Sacks’s commentary moves from posts to policy. Criticism on X costs nothing. Recommendations from a presidential advisory council are a different matter.
Finally, there is the 0.002% block rate on agent actions. With roughly 30,000 agents running daily, small percentages still translate into real activity.
