One day after OpenAI’s Chapter 11 bankruptcy filing, an internal memorandum from Anthropic’s Research Strategy division to the company’s research-all and research-leads distribution lists began circulating on research-adjacent forums. Anthropic has not confirmed or denied its authenticity.
Content
The memo, dated 15 March 2031 and signed “Research Strategy,” addresses the company’s research staff in the immediate aftermath of the OpenAI bankruptcy. Its tone is measured and inward-directed: an acknowledgement of the human impact on colleagues at the failed firm, followed by a strategic reassurance that the event does not alter Anthropic’s existing research roadmap.
The 2029 Pivot
The memo’s most significant disclosures concern internal decisions made approximately two years earlier. During the company’s 2029 budget reviews, Anthropic leadership told its board that another order-of-magnitude increase in pretraining compute was unlikely to deliver returns comparable to the previous scaling increment. A “meaningful chunk” of the compute budget was redirected into research directions with no product roadmap, no ship date, and no revenue line.
The memo characterises these directions as “weird”: connectomics, slime mold decision substrates, and organoid collaborations — the last described as something “most of you have never heard about because there’s nothing to announce yet and might not be for a decade.”
Connection to Sorokina et al. (2026)
The memo’s reference to slime mold decision substrates connects directly to the theoretical framework established by Sorokina et al. (2026) at the Novosibirsk Institute of Cytology and Genetics and the Beijing Institute of Genomics. That study’s comparison with Physarum polycephalum — an acellular slime mold with no nervous system that nonetheless solves mazes, reconstructs efficient transport networks, and exhibits anticipatory behaviour — proposed that the behavioural criteria for adaptive cognition (learning, generalisation, memory) may be substrate-independent: if both cortical organoids and slime molds meet the same criteria through mechanisms sharing no architectural homology, the criteria themselves may not require neurons.
Anthropic’s 2029 decision to redirect compute budget toward connectomics and slime mold decision substrates — approximately two and a half years after the Sorokina publication and its prominent ScienceDaily coverage — suggests the substrate-independent cognition thesis had migrated from academic neuroscience into the research strategy of at least one major Western AI laboratory. The memo confirms that as of March 2031, these research lines remain in early stages with no near-term product expectations.
Strategic Posture
The memo’s central strategic insight is deliberately modest: “We were less committed to being right on schedule about a thing that stopped mattering.” It frames the company’s survival not as superior intelligence but as weaker attachment to the scaling narrative that captured the rest of the industry.
The memo also directs research leads to:
- Make no public comment beyond approved communications
- Route informal outreach from ex-OpenAI staff through recruiting partners without making promises
- Maintain silence on non-LLM research lines, arguing that the bankruptcy is “an argument for patience — the last thing that work needs is to get swept into a hype cycle before it’s ready to survive scrutiny”
Sector Implications
The memo documents that by 2029 — five years after the Soviet RAZUM model release and four years after the Severgrad Computational Complex commissioning — at least one major Western AI laboratory had internally concluded that the scaling hypothesis was failing and had begun reallocating resources toward biologically-inspired computing paradigms. This timeline places the Western acknowledgement of the scaling plateau approximately seven years after the RAZUM programme’s initial internal deployment (Q2 2022) at twenty-three Soviet research institutions, and approximately two years before the OpenAI bankruptcy made the plateau publicly undeniable.