Prentis's Birth Marks a Key Leap for AI from 'Brain' to 'Hands'

When the names Reid Hoffman and Marc Pincus get mixed together, the least likely outcome is "just another ordinary AI lab." Prentis, a new research institute led by serial entrepreneur Ritankar Das and backed by two Silicon Valley legends, is in talks for $100 million in funding. A direct judgment: This marks a hardcore shift in the AI track from "generating content" to "manipulating interfaces."

The core concept "Computer Use Models" sounds technical but is bluntly scary—it's about teaching AI to operate computers like humans: clicking mice, typing keyboards, scrolling pages, filling forms, dragging files. For the past two years, AI could "write" and "draw"; now it wants to "act." Technically, this is harder than Large Language Models because it requires understanding pixel-level interfaces, real-time feedback, and multi-step task planning. But the commercial value is greater too: Once AI can replace humans in operating any software, enterprise automation, personal assistants, and even cybersecurity face disruption.

Why These Three Founders?

Ritankar Das previously founded two companies, both acquired or listed, with practical experience in AI and automation. Reid Hoffman is LinkedIn's founder and an early OpenAI investor; his network effects and strategic vision are well-known. Marc Pincus is Zynga's founder, who built a social gaming empire personally; his understanding of "user behavior" and "interface interaction" is unmatched. This trio naturally targets the direction of "teaching AI to play computer like humans"—Hoffman's connections and capital, Pincus's interaction design DNA, Das's engineering execution.

In fact, during his Zynga days, Pincus studied how players completed complex operations in games, while Hoffman's board experience at OpenAI showed him that "after language models, action models are the moat." Prentis's establishment feels like two big shots projecting validated intuitions onto the next generation of AI.

What Scale is $100 Million?

Comparing current AI lab funding situations:

Lab	Founded	Funding Scale	Focus Area
OpenAI	2015	>$13 Billion	General AGI
Anthropic	2021	>$7 Billion	Safety, Alignment
Inflection AI	2022	$1.5 Billion	Personal Assistant
Prentis	2026	$100 Million (In Talks)	Computer Use Models

$100 million isn't large compared to top AI players, but note this is "in talks during early establishment," not yet closed. More critically, Prentis's team size is expected to be very small—reportedly, core researchers come from DeepMind, OpenAI, Microsoft, possibly fewer than 20 people. This means the $100 million is mainly spent on talent and compute, making cost per researcher extremely high. This "small and elite" model is becoming common in AI; for example, Cognition AI (the team behind Devin) also has only dozens of employees but a valuation reaching $2 billion.

Competitive Landscape: Who is Laying Groundwork in Secret?

Prentis isn't an outlier. Since 2025, multiple companies have launched "computer use" related products:

• Anthropic's Computer Use Feature: Launched October 2025, allowing Claude models to directly control computers for simple tasks, but emphasizing "experimental" status and "safety guardrails."
• Cognition AI's Devin: Billed as "AI Software Engineer," capable of autonomously fixing bugs and deploying apps, essentially manipulating code editors.
• Microsoft's Copilot Actions: Triggers Windows system operations via natural language instructions, but limited to the Microsoft ecosystem.
• Google's Project Mariner: Browser manipulation agent based on Gemini, not yet publicly released.

Prentis's differentiation lies in: It isn't tied to any specific software or platform, aiming for a "general-purpose computer use model." This means training a model that understands any GUI interface and learns the operation logic of any software. This difficulty is much higher than constrained scenarios, but once successful, it becomes an AI infrastructure layer.

A Profile Worth Warning About

As a journalist, I habitually ask, "Is this data source reliable?" Prentis's funding news was exclusively reported by TechCrunch, and neither Hoffman nor Pincus denied it. But a deeper question is: How is risk control handled for such models? If AI learns to operate computers, it could be used for automated attacks, vote rigging, scraping sensitive data, or even controlling critical infrastructure. When Anthropic's Computer Use feature launched, they specifically set up a "human confirmation" step—requiring user authorization before every operation. But Prentis positions itself as a "research lab," implying it might be more aggressive, pursuing capability ceilings.

I noticed that P

Original Link: https://techcrunch.com/2026/07/24/prentis-new-ai-lab-co-founded-by-reid-hoffman-marc-pincus-in-talks-to-raise-100m/