In just three years, a new wave of researcher-led AI startups has collectively raised over $40 billion — proving that frontier-scale ambition is no longer strictly the domain of major incumbents.
To understand what is driving this massive shift in venture capital and elite talent, this week’s episode of Radical Talks tackles the phenomenon head-on. Radical Partners Rich Kotite and Aaron Rosenberg sit down with Molly Welch to unpack their recently published research, “The Rise of NeoLabs.”
In addition to mapping the six research paradigms these labs are chasing, they discuss why diffusion language models are a direction the big labs are unlikely to pursue given how productive their autoregressive roadmaps already are. They also explore why the strongest teams build cultures of deep loyalty, dig into why NVIDIA now participates in more NeoLab rounds than any institutional investor, and close on the risks of building at the frontier.
Listen now on YouTube, Spotify, or Apple Podcasts.
AI News This Week
-
These AI-Native Companies Have Tiny Staffs and Fewer Bosses (WSJ)
AI-native companies are building organizations that stay lean and flat by design rather than layering automation onto structures built for a different era. OffDeal, a Radical Ventures portfolio company using AI agents to bring investment banking services to small businesses, found that bankers lost hours deciding what to do next, scrolling the CRM, and switching tools. The team built an AI task management system that surfaces each banker’s highest-leverage action while minimizing context switching.
-
America’s AI Labs are Under Threat from Cheap Chinese Rivals (Economist)
The release of Kimi K3, a Chinese model only slightly behind Anthropic’s Fable and OpenAI’s Sol, marks how quickly open-weight systems are closing the gap on frontier AI. These models are far cheaper, with the most capable DeepSeek system averaging four cents per task against $2.75 for Fable, an efficiency partly driven by innovations in response to U.S. chip export restrictions. The three-week global outage of Fable pushed companies to reexamine their reliance on any single lab, strengthening the case for sovereign providers like Radical Ventures portfolio company Cohere, whose platform is model-agnostic and runs inside a customer’s own borders.
-
AI chip Startup Etched Defies Skeptics, Hits $10.3B Valuation from Big-Name Investors (TechCrunch)
Radical Ventures portfolio company Etched closed a new $300 million Series C at a $10.3 billion valuation, roughly doubling its worth in seven months in a round led by Sequoia and Andreessen Horowitz, with participation from Jane Street and Radical. Etched builds full inference systems around proprietary chips purpose-built for transformer models. Two components designed from scratch, a low-voltage prefill chip and shared-memory interconnect technology for the decode stage, raise AI inference speed while lowering cost. The company recently manufactured its first generation silicon chip, began client testing, and has booked $1 billion in orders.
-
Could A.I. Do Your Job? We Put Agents to the Test. (NYT)
AI agents are proving capable of collapsing tedious office work into minutes. In one hands-on test, an agent completed seventeen employment forms almost instantly by writing its own Python script, reading employee data from a spreadsheet and formatting Social Security and phone numbers correctly along the way. The frontier now sits at the handoffs between tools, where an upload step that a script cannot reach still requires human direction. Agents add value on structured, repeatable tasks today, with reliability climbing as the tooling matures.
-
Research: Self-Improvements in Modern Agentic Systems (Jilin/KAUST/UofA/IDSIA)
Researchers have made advancements in the field of self-improvement, the study of methods that enable an AI agent to turn its own experience into lasting capability gains. Updates from inference are committed to either the weights or the harness of the agent, and methods are sorted by where the learning signal originates from self-generated demonstrations, the agent’s own evaluations, or experience gathered from real and simulated environments. The most open-ended systems rewrite their own code with minimal human input.
Radical Reads is edited by Ebin Tomy (Analyst, Radical Ventures)
