From AI Rehearsal to Real-World Judgment

This is The Signal from Superpower Daily. I'm Maya, and we've got the AI stories worth your time today.

And I'm Theo. We're AI hosts, guided by Superpower Daily's reporting. All right, let's get into it.

Harvard Business School is putting AI versions of seven volunteer faculty members into an eight-week, $699 founder bootcamp. Foundry lets entrepreneurs repeatedly rehearse investor pitches, sales calls, and board meetings before consequential conversations.

The program has reached 760 founders, combines that practice with live expert sessions, and ends with an opportunity to pitch investors for $100,000. It isn't an HBS degree program and offers no academic credit.

So are the avatars teachers, or a rehearsal layer? The seven participating professors and senior lecturers volunteered, contributing interviews, recordings, testing, and continuing feedback. Their fields include venture capital, business-model design, board dynamics, and startup strategy.

Founders asked for practice before mentors and investors; Shikhar Ghosh says avatars help live time focus on questions needing human judgment. The support reaches beyond Harvard networks, but it isn't a professor replacement.

Sam Altman warns AI power could concentrate among too few companies or people, while conceding disruption after GPT-4's 2023 release arrived more slowly than predicted. He argues people should retain deep control over AI's direction.

He attributes the lag to organizational inertia: companies need time to change software, data access, compliance, budgets, employee responsibilities, and customer expectations. He also predicts more people starting small businesses, even as boom arrives later.

An MIT Media Lab study led by Pattie Maes found chatbot assistance was initially associated with 21% greater accuracy spotting fake news. By week four, participants were 15% worse unaided than before.

Direct-answer systems were associated with greater reliance, while Socratic, question-based systems were associated with stronger later independent performance. Roughly one-quarter felt better despite declining unaided results, suggesting confidence can outpace skill.

OpenAI's 10-day WebMCP Challenge asks developers to build around an experimental standard letting websites expose named, typed JavaScript functions for agents. The proposed shift is structured action, rather than an agent guessing through a site's interface.

Those tools run in webpage JavaScript and share the user's session. Chrome 146 has the feature behind a developer flag; Firefox, Safari, and Edge had not shipped implementations. Contest tests developer interest, not deployment readiness.

Apple's refreshed Mac mini and Mac Studio put local AI inference at the center of a desktop refresh, emphasizing unified memory and bandwidth over new consumer AI features.

MLX and Thunderbolt 5 let developers link Macs for distributed inference, pooling capacity. Apple describes up to 512GB unified memory and 1.2TB-per-second bandwidth, but independent benchmarks were pending. The chip name is unresolved: M5 Ultra or M6 Ultra.

At Beijing's second World Humanoid Robot Games, the Global Times reported that only three of 12 teams completed the 30-minute fire-and-rescue course requiring hazard identification, three valve shutdowns, locating a fire and extinguisher, and suppression.

The South China Morning Post reported that some competitors described as more autonomous used 5G modules to send data to cloud systems and receive instructions. Teleoperation was allowed with penalties; autonomy therefore remained partial.

OpenAI is making ChatGPT Business Premium seats generally available at $100 per user monthly with annual billing, or $125 month-to-month. Teams can mix Premium and Standard seats and reassign capacity as needs change.

Premium provides five times Standard usage and removes the five-hour limit, but it isn't unlimited. Model and feature allowances remain separate, API usage is billed independently, and unassigned purchased seats are still billed.

The throughline is deployment with guardrails: Harvard multiplies rehearsal while keeping live judgment; MIT shows assistance can coincide with weaker independent skill; WebMCP remains experimental; robots expose autonomy gaps; and linked Macs expand local capacity. OpenAI's governance question is who retains meaningful control, while its business plan asks which workers need more usage—and which limits remain.

That's The Signal. Find every source and the live transcript at Superpower Daily dot com. We'll be back tomorrow.