Startup Bios Sampler

Brian McCullough introduces Hone, a five-month-old startup that raised a $60 million seed round led by Benchmark and Index Ventures at a company-reported $285 million valuation, whose pre- or post-money status is unknown. CEO Moritz Stephan describes the product as AI that “owns outcomes,” not agents that wait for individual tasks. Hone’s “engines” receive goals such as improving customer retention, learn a company’s systems, people, history, and constraints, then decide what actions to take and whom to involve. The company compares its approach to Kubernetes: maintaining a desired state rather than following step-by-step instructions.

Hone argues that current agents remain tools inside human-run workflows because people must decompose goals, provide context, and monitor failures. Its manifesto compares today’s AI to Edison’s dynamo: the productivity breakthrough comes from reorganizing the factory around the power, not merely producing more power. Hone wants full-time engines that retain knowledge, improve over time, and treat the customer’s accumulated information as its intellectual property. An engine should be judged like an employee—“what did you get done this week?”—by its effect on the profit and loss, such as reducing onboarding drop-off to 8% without damaging other metrics.

Because Hone’s engines can rewrite their tools, code, and memories, the company simulates decisions before allowing them to go live. In its example, an engine changes “ask George before ordering from a new supplier” into an exception for a downed production line after replay reveals that blindly waiting for George would delay needed bearings. Hone’s “director” simulates the outside world, while fake versions of software tools reproduce real facts about 90% of the time; weeks can be replayed in minutes. In 125 internal scenarios, Hone says replay reduced breaking changes from 33% to 5%, though nobody outside has verified the result.

Stephan previously launched Vienna book-delivery startup Lobu, co-founded a Mars rover project called Tumbleweed that grew past 50 people across four countries, researched AI at Stanford, co-founded anonymous campus app Fizz, and helped build Cognition’s Devin before becoming Scott Wu’s chief of staff. Fizz co-founder Carlo Chua and former Harvey and Mercury engineer Oliver Brady joined him at Hone. Known customers are Cognition and Modal; Cognition is also an investor through Wu, creating a close founder-customer connection. McCullough notes that Hone has disclosed neither revenue nor customer results, making the round primarily a bet on the team and a difficult long-horizon problem.

Ghost, founded by 19-year-old former quant Zain Javed and fellow 19-year-old Nicholas Chua, emerged from stealth with an $11 million Andreessen Horowitz-led seed round. Its Core device is a screenless, stainless-steel computer with an NVIDIA RTX Pro 4000 Blackwell GPU with 24GB of memory, an AMD Ryzen 5 six-core processor, 64GB of RAM, and a 1TB SSD. It costs $3,499 with no subscription, runs local models including Alibaba’s Qwen 3.8 family, Google’s Gemma 431B, and Ghost’s Muse Glimmer 30B, and accepts models from Hugging Face. First-batch shipping is scheduled for October 31, with a 30-day return window and one-year warranty.

Ghost says Core keeps models, memory, source code, and logic on the device, so the user owns the system even if Ghost disappears; Javed contrasts it with renting an assistant from a cloud provider. The device can connect email, calendars, finances, files, browser history, health records, smart-home equipment, and wearables, while a firewall is intended to block suspicious credential use or data exfiltration. However, the homepage says “no data ever leaves your home,” while the support page and privacy policy acknowledge that web search, email, updates, and internet tasks can send information through Ghost’s gateway to outside services. McCullough also flags untested security, in-house manufacturing challenges, and the risk that local models trail larger cloud models.