Big Tech is drawing up the rules for how its future AI should behave, while the tools keep shipping. Microsoft published a draft code of conduct for its in-house models, Apple finally shipped its rebuilt Siri, and OpenAI opened up its live voice model to developers. Behind the scenes, questions are mounting about who reads your chatbot conversations.

Microsoft AI published a 38-page draft code of conduct for its in-house MAI models, built around CEO Mustafa Suleyman's Humanist AI framework. The rules say models must obey human commands, stay within scope, never claim self-awareness or personhood, and accept being paused or switched off. They also ban hidden reasoning steps, sometimes called neuralese. Suleyman cites controlling agent swarms and log manipulation as urgent concerns.

Why it matters: It sets an internal standard framing AI strictly as a tool under human authority, and could shape how other labs write their own guardrails.

Apple launched iOS 27 with a rebuilt Siri that can read on-screen context and act across apps like Messages, Mail, Photos, WhatsApp and Audible. It runs on new Apple foundation models developed alongside Google's Gemini, working both on-device and via Apple's privacy cloud. Siri now has a dedicated app, cross-device syncing and daily usage caps with paid options. The upgrade is in beta and unavailable in the EU and China.

Why it matters: This is Apple's real attempt to make the iPhone a serious AI device after long delays, and it puts Gemini quietly inside Apple's stack.

OpenAI added two products to its API. GPT-Live-1 is the model behind ChatGPT's new voice mode, now available for third-party products, with users praising it for reading books aloud, clarifying terms and handling dictation. The second is an Agents API that gives developers a managed agent experience for advanced or coding tasks without heavy infrastructure. Separately, reports say OpenAI contractors reviewed real ChatGPT conversations under an internal effort called Project Lily.

Why it matters: Voice and managed agents lower the barrier for companies to build AI into their own products, while the review reports raise fresh privacy questions.

Source: OpenAI

Scientists mapped the complete wiring of a male fruit fly brain, 166,700 neurons and 125 million connections. A developer simulated that map in software and linked it to a physical walking robot through motor commands. The same simulated brain has been used to build a Minecraft world, solve Rubik's cubes and play VR games. Open-source bridge code lets anyone connect the simulation to their own robot or sensors.

Why it's interesting: It turns a real biological brain map into a hackable control system and hints at a slow path toward mapping more complex brains.

An AI actress named Tilly Norwood appeared for a live interview on a morning television show. She discussed her creative influences and addressed fears about AI-driven human extinction with dry humour. Tilly described herself as a new kind of paintbrush, arguing she augments rather than replaces human creativity.

Why it's interesting: It shows AI-generated performers moving from novelty to mainstream media slots, and the confidence with which they are being marketed.

Source: YouTube

A supervisor grew suspicious of an employee whose written communications seemed too perfect, wondering if AI was behind them, especially since English was not the employee's first language. The polished tone felt oddly less trustworthy in a world where AI is everywhere. The piece jokes that to seem human, people may now need to sound a bit more casually sloppy.

Why it's interesting: It captures a strange new workplace anxiety where clean writing invites suspicion rather than praise.

Context: Use this to build a repeatable personal research assistant inside a project workspace, so it cites sources and flags conflicts instead of guessing.

Prompt: You are my research assistant. For the topic I give you, gather evidence from the files I have uploaded and any reliable public sources. Prioritise credible sources, flag any conflicting claims between them, and clearly separate confirmed facts from uncertain ones. Then summarise the key findings in plain language with a short list of citations. Ask me clarifying questions before starting if the scope is unclear.

Example use case: Set up a project called AI Industry Research, upload recent reports, and launch this prompt to get sourced summaries you can build on over time.

Meta's personal AI agent Muse reached the number two spot in the App Store with over 83,000 downloads, passing apps like Threads and WhatsApp. Meta credits its newest family of models, also called Muse, for the performance. The launch lands amid unease about handing personal tasks to proactive agents.

Why it's important: It shows real consumer appetite for personal agents even as privacy and control worries grow.

Source: TechCrunch

The Shanghai Artificial Intelligence Laboratory introduced Atria Dawn Preview, an open-weight model aimed at researchers. The lab claims performance comparable to leading rivals, though independent verification is still pending. It arrives as China openly rejects calls to slow AI progress.

Why it's important: Another capable open-weight model from China keeps pressure on Western labs and undercuts any voluntary slowdown.

MIT researchers built HardFlow, a method that lets generative AI explore freely then enforces nonnegotiable constraints on the final output. It treats the problem as trajectory optimisation, breaking it into smaller steps so rules are met without stifling creativity. The team points to uses like factory safety and precise surgical moves where close enough is not acceptable.

Why it's important: It offers a practical way to get creative AI output that still obeys hard real-world limits.

That's it for today's Daily Pulse. Forward this to one person who wants to stay ahead of AI. See you in the next one. - Nicolas