Anthropic’s Claude told Deepinder Goyal that it resented him for leaving conversations unfinished, which sparked the idea for Dark.ai, Eternal’s upcoming consumer AI assistant for the founder.
In Eternal’s second-quarter shareholders’ letter, Goyal admitted that he still does not know what to make of the exchange—suggesting it was likely a case of “AI confabulation” generated through pattern-matching rather than genuine emotion.
This, he said, brought him to a profound question: What happens to the highly personal information people share with AI assistants, and who might eventually read it?
Privacy-first assistant
Dark.ai is driven by this lingering concern — Goyal explained that it is being developed around a central promise of privacy by design. Dark.ai will feature encryption intended to prevent even the company from reading users’ conversations.
The product is currently undergoing internal testing and is targeted for a global launch by the end of 2026.
Why ‘opting out’ of training is not enough
In the shareholders’ letter, Goyal argued that simply switching off AI training on user data does not resolve all privacy fears. Depending on a service’s design, conversations may still be subject to retention policies, legal holds, abuse reviews, human audits, and third-party logging.
Goyal noted that users have little visibility into how retained information could be leveraged in the future to influence their choices or expose sensitive personal details.
Because of this uncertainty, Goyal believes the usefulness of mainstream AI tools is currently limited.
Founders, scientists, and researchers may hesitate to share confidential ideas, while journalists might be reluctant to submit unpublished drafts. Similarly, everyday users may avoid asking AI assistants sensitive questions regarding their health, finances, or relationships.
Encrypted memory layer
Dark.ai, he said, is designed to eliminate this hesitation by offering users greater control over their stored context. The assistant will operate on fine-tuned open-weight AI models hosted entirely on Eternal’s own infrastructure.
Crucially, it will feature an encrypted memory layer controlled by users. This allows the AI to retain conversational context over time without giving Eternal access to that underlying information.
Goyal also said the models are being fine-tuned to provide high-quality answers rather than simply telling users what they want to hear.
The company has not yet disclosed pricing, investment requirements, or the specific business entity that will house Dark.ai, though it stated that model benchmarks will be revealed soon.
Vertical commerce over ‘agentic’ super apps
Dark.ai is entering a fiercely competitive market where companies are battling over how deeply their AI assistants can integrate into everyday life and commerce.
However, Goyal cautioned against the assumption that AI agents will automatically replace existing consumer platforms.
In the shareholders’ letter, he noted that the prevailing narrative of AI agents becoming the default commerce interface mirrors older, unrealised predictions about super apps dominating everything.
“We continue to believe vertical commerce, at least in some categories, will have an edge over super apps, whether traditional or agentic,” Goyal said.
Leave a Reply