The Information Launches Third Episode of AI Deep Dive on AI Recursive Self-Improvement
In-Depth Interview on Chip Recursive Self-Improvement
Technology media The Information has released the third episode of its AI deep dive series, focusing on recursive self-improvement: AI designs better chips, and better chips retrain stronger AI. Guests include Anna Goldie and Azalia Mirhoseini, co-founders of Ricursive Intelligence.
The two previously co-led Google’s AlphaChip, a system capable of generating chip layouts in hours and participating in multiple generations of tensor processing unit designs. They left Google in 2025 to establish Ricursive, advocating for compressing the design cycle from two to three years down to weeks. The first phase accelerates human design, the second phase automates chip production for clients without in-house teams, and the third phase uses stronger models to design the next generation of chips under human supervision. The company has raised approximately $335 million, with a valuation of about $4 billion, backed by investors including Sequoia, Lightspeed, DST, and NVIDIA.
The episode covers topics such as what recursive self-improvement means, whether chip design can be automated, co-design of hardware and software, surpassing existing models and chips, how much efficiency can be improved, what happens after loop acceleration, whether alignment can keep pace, and whether progress should be slowed. They emphasize that they are not creating a replacement for traditional electronic design automation tools, but aim for an end-to-end model: ingesting workloads and outputting manufacturable layout files.
Recursive self-improvement is viewed in the industry as the hardware entry point to intelligence explosion. I.J. Good described superintelligent machines designing better machines as early as 1966. DeepMind's AlphaEvolve has been used for network architecture, data center scheduling, and chip-related optimizations. Anthropic claims that engineers are delivering code at rates several times higher than in previous years and warns that fully autonomous design of successor systems may occur before institutions are prepared. OpenAI also has internal models for writing training kernels and optimizing agent collaboration.
Ricursive pins the loop at the physical layer: algorithm iterations outpace tape-out, leading to wasted computing power; shortening tape-out with models is necessary to synchronize algorithms with silicon. Currently, human setting of goals and acceptance is still required, rather than models rewriting their own weights.
In market mechanisms, this is a narrative of content customer acquisition layered with hardware startups, not a chip shipment announcement. Buyers are model companies needing customized accelerators but unable to afford a complete backend team; sellers are startups compressing design cycles. Funding flows from venture capital into GPU training hours, then attempts to charge for tape-out level layouts. Beneficiaries are the former Google team with layout generation experience, while traditional chip routes with two-year cycles and electronic design automation vendors charging by suite face pressure. The first phase of interviews does not change TSMC's schedule.
Source: Public Information
ABAB AI Insight
Goldie and Mirhoseini have commercialized their layout generation experience from Google. AlphaChip demonstrated that local steps can be completed in hours, while Ricursive aims to connect placement, routing, and verification into a learnable end-to-end model that can be reused across projects. The $335 million funding primarily buys computing power, indicating that the product has not yet replaced the cash flow of electronic design automation, first proving generalization on card burning. NVIDIA's investment turns a potential competitor into a design layer within the ecosystem, rather than waiting for it to grow into a rival chip manufacturer.
The capital path is "design as a model." Whoever first maps workloads to manufacturable files can charge per instance or performance improvement, without needing to build their own wafer fab. Once the loop starts turning, advantages will accumulate along with data: each new chip makes the next one faster. This contrasts with old vendors that only sell tool licenses.
Analogous to DeepMind using reinforcement learning for data center layout and OpenAI using models to write CUDA kernels: software first consumes the design process, then feeds back into hardware rhythm. The industry phase is collaborative design, not the emergence of fully autonomous self-replication. The alignment chapter is included in the same issue because after hardware loop acceleration, the calendar for safety assessments will be dragged along by the design cycle.
Structural judgment belongs to technological substitution. Chips are transforming from scarce artisan products into artifacts that can be iterated by models. The mechanism is: if algorithmic progress outpaces tape-out, the computing power structure will become outdated; speeding up tape-out is equivalent to releasing the physical brakes on model advancement. Recursion is not a mysterious singularity, but the compound interest of design feedback changing from years to weeks.
ABAB News · Cognitive Laws
If models can draw chips, the upper limit of the next training will not only depend on algorithms.
As design cycles compress from years to weeks, the calendar for safety assessments will also be dragged along.
First sell to those without chip teams to bypass the licensing walls of old tools.