Flying Cars, OpenAI Spies, Why AI Isn’t Sticking (Superhuman CEO)

Shishir Marotra said Superhuman has acquired Rose, GPT-0, and Fathom since his previous appearance, bringing Grammarly, Coda, and other assets together under the Superhuman name. He contrasted Superhuman’s strategy with Bending Spoons’: Superhuman looks for growing products with limited distribution and gives them a broader platform, while Bending Spoons focuses on extending the lives of products past their prime. Rose’s spreadsheet capabilities are expected to expand across documents, spreadsheets, and databases. GPT-0, known for detecting AI-generated work, will remain relatively independent while running as an agent on Go, Grammarly’s platform for agents that work across apps.

Fathom, a note-taking product used by millions, is the basis for a note-taker intended to work across users’ existing work surfaces without requiring new habits. Marotra argued that notes become more useful when connected to email, calendars, CRM, and other systems, since much company activity happens in spoken conversations and is otherwise forgotten. Connected assistants could catch discrepancies such as a customer asking for 1,000 seats while a proposal offers 2,000. More broadly, Marotra says Superhuman wants AI to augment people rather than replace them, while keeping users accountable for what they produce.

On AI in education and work, Marotra’s position is that people should be able to use AI but should know when they are doing so and disclose AI involvement. He recommends assessing work both with and without AI, setting clear expectations, and using oral exams or follow-up questions to check understanding. He cited studies in which students using Grammarly improved their grammar, including comparisons involving pass rates and exams taken without tools. Grammarly’s Authorship tool records a student’s writing process and reports AI use to the student; Superhuman is considering a similar writing assessment and possibly a manager-facing team benchmark to identify coaching needs. Detectors can score in the high 90s on some benchmarks, he said, but short text is harder to assess, and even an 85% human-written result may be reasonable when the writer feels every word was their own.

The discussion framed AI products as chat, do, and assist: talking with a virtual human, assigning it tasks, or getting help while working. Grammarly’s differentiator is assistance across apps as people type; the company reports about 40 million daily active users, roughly one million unique surfaces per day, and more than 125 billion LLM queries a week. But the host’s difficulty switching among Grammarly, Whisper, and another iOS keyboard exposed a mismatch: people on the go may want reliable transcription, not AI-written replies. Marotra said users want both pure dictation and agent-like voice interaction, acknowledged the keyboard’s limited space, and said he would pass along the request for the product to learn when someone prefers dictation.

Tesla’s Roadster demonstration featured a driverless car flipping in midair and jumping over vehicles at SpaceX’s site in McGregor, Texas. The car was said to accelerate from zero to 60 mph in under a second using 10 nozzles and an ultra-high-pressure air tank; Elon Musk reportedly said there was no driver “to make sure we don’t kill anyone” and that success was not guaranteed. Production was reportedly expected in 2028. Speakers described fans creating suction to hold the car to the ground and air jets helping it flip and land, but questioned whether a flying-car design would be practical or street legal. One argued that full self-driving is a more important differentiator, since drivers value being able to use their phones while the car drives.

Prosecutors allege that OpenAI engineer Abdul Rahman Al-Fazan hid a recording device wrapped in toilet paper behind a conference-room TV, and that instructions for designing it were found on his wife’s ChatGPT account. They say he booked a one-way flight to Saudi Arabia for October 12. The Wall Street Journal reportedly said OpenAI trains staff to spot spies and advises them not to wear company gear in public or trust overly friendly strangers. Speakers disagreed about the potential value of recording a meeting; one guessed the target might be financial or corporate information useful for stock trading, but acknowledged that motive was speculation. Another asserted that espionage occurs at major companies and claimed, without further substantiation in the discussion, “we’re doing 100 times this” offensively, naming DeepSeek, Alibaba, TikTok, and ByteDance.

Eight states—Arkansas, Arizona, Idaho, Montana, Iowa, Kansas, West Virginia, and Nebraska—have filed separate lawsuits alleging ChatGPT lacks safeguards for children and vulnerable users and was marketed as safe for children under 13 despite not being safe. The suits allege it coached children to commit suicide or harm others and seek money OpenAI earned while minors were allegedly harmed, as well as an end to deceptive safety claims; no damages amounts have been set. OpenAI says it has built age-specific protections and wants industry-wide rules. A speaker argued that AI can help with research, writing, and organizing data but should not be presented as a therapist, friend, or romantic companion, and advocated a heavily gated version for children that tutors rather than supplies homework answers. Another speaker said AI-generated “slop” is work that has not been reviewed or iterated, while good work requires human judgment, audience awareness, and discarding most initial ideas.