Richard Socher, former chief scientist at Salesforce, has secured $650 million in funding for a new venture focused on developing artificial intelligence systems capable of improving their own performance without human intervention, according to TechCrunch AI.
The funding round, one of the largest for an AI startup in 2026, positions Socher’s company to compete directly with established players pursuing artificial general intelligence whilst taking a distinct technical approach centred on recursive self-improvement—systems that can identify weaknesses in their own architecture and modify their code accordingly.
Socher, who previously founded MetaMind before its acquisition by Salesforce in 2016, has committed to shipping a commercial product within 18 months rather than pursuing pure research. This timeline distinguishes the venture from competitors like OpenAI and Anthropic, which have pursued longer development cycles before major product launches.
The technical approach centres on what researchers call ‘meta-learning’—AI systems that learn how to learn more effectively. Rather than requiring human engineers to manually improve model architectures, these systems would theoretically evaluate their own performance across tasks and adjust their internal structures to optimise outcomes.
“The bottleneck in AI development has shifted from compute to architectural innovation,” Socher told TechCrunch AI. “We’re building systems that can explore that design space faster than human teams.”
The business implications extend across the AI development stack. If successful, self-improving systems could dramatically reduce the engineering resources required to advance model capabilities, potentially compressing development timelines from years to months. This would pressure existing AI labs to accelerate their own timelines or risk falling behind competitors with more autonomous development processes.
Cloud infrastructure providers stand to benefit from increased computational demands during the self-improvement training phase, which requires extensive experimentation across model variants. Conversely, AI consulting firms and specialised ML engineering teams may face margin pressure if architectural optimisation becomes increasingly automated.
The approach carries substantial technical risks. Self-improving systems require robust safety constraints to prevent unintended optimisation—a challenge that has proven difficult even with human oversight. Previous attempts at meta-learning have struggled to generalise beyond narrow domains, and scaling these techniques to frontier model sizes remains unproven.
The funding structure includes provisions requiring the company to demonstrate specific capability milestones before later tranches are released, according to sources familiar with the terms. This performance-based approach reflects investor caution following high-profile AI ventures that consumed substantial capital without shipping commercial products.
Socher’s track record provides credibility that pure research ventures lack. MetaMind’s image recognition technology was integrated into Salesforce’s Einstein AI platform and remains in production serving enterprise customers. His departure from Salesforce in 2022 was amicable, with the company taking a minority stake in the new venture.
The competitive landscape includes DeepMind’s work on AutoML systems and OpenAI’s exploration of AI-assisted research, though neither has announced dedicated commercial products in this domain. Socher’s 18-month product commitment suggests a more applied focus than these research initiatives.
Market watchers should monitor three indicators of progress: published benchmarks showing self-improvement capabilities on standard tasks, customer pilots with enterprise partners, and peer-reviewed papers detailing the technical approach. The latter will be particularly telling—credible self-improvement claims require transparent methodology that the research community can evaluate.
The $650 million backing represents a substantial bet that AI development itself can be automated, potentially accelerating the pace of capability gains whilst concentrating expertise in fewer organisations capable of building these meta-systems.
