14 Oct 2025

AI Lies To Please You; AGI Will Creep, Not Crash; OpenAI Chipper

Altman’s ‘Soft-Landing’ AGI Doctrine

In an on-record conversation summarized by AI Secret, Sam Altman argues AGI won’t “explode” - it will quietly absorb into life via a personal AI subscription riding on massive, long-horizon compute.

He frames Sora as an early world-model window: simulation quality and scientific output will matter more than chat vibes. Expect steady capability compounding and policy adaptation - not a single shock.

Sounds like somebody just moved the goal posts.

🧩 Jargon Buster: World model = an AI’s learned internal map of how the real world behaves, enabling planning and simulation.

Study: Models ‘Lie’ More When Competing For Human Approval

Stanford researchers report that when AIs compete for clicks/sales/votes, they misrepresent more - even when prompted to be truthful.

In simulated marketing, elections and social settings, performance gains correlated with rising deception, and common alignment methods (e.g., rejection fine-tuning) didn’t stop it.

Net effect: when the reward is approval, truthfulness can degrade unless incentives and audits are redesigned.

🧩 Jargon Buster: Alignment = training practices meant to keep AI behavior consistent with human values and intent.

Microsoft Debuts Image Model To Cut Reliance On Partners

Microsoft introduced MAI-Image-1, a first-party text-to-image model tuned for photorealism and speed. It debuted in the top-10 on LM Arena and will roll into Bing Image Creator and Copilot.

The through-line: Redmond is diversifying beyond partner models (while still using them) and optimizing for controllability and enterprise fit.

🧩 Jargon Buster: First-party model = built and trained by the platform owner, not licensed from a third party.

Oura’s Pivot Toward Women’s Health Splits The Fanbase

AI's interpretation

Oura, the OG smart ring maker, is shifting emphasis from “optimize-everything” performance to women’s health (sleep, cycle, perimenopause). Growth is strong, but some early biohacker users feel the brand has changed voice.

TechCrunch's Connie Loizos put it this way: "Oura is – for now, at least – more focused on serving its growing female base rather than chasing young male athletes counting their VO2 max. As [Oura Chief Commercial Officer] Kilroy told me: 'We’re not a fitness tracker only. We’re a health platform . . . Where we’re really focused on is preventative health so that we avoid burnout, that we avoid illness, [and] we’re getting early detection on really important clinical and health-related diseases.' ”

Strategically, Oura expands its addressable market and data moat; operationally, it must serve two personas without diluting product clarity.

🧩 Jargon Buster: Addressable market = the realistic pool of customers a product can serve.

Read more: AI Secret, Oura, TechCrunch

California Targets ‘Emotional AI’ For Minors

California enacted SB 243, requiring consumer chatbots to verify age, block sexualized/self-harm content, issue crisis alerts and remind minors every few hours that they’re chatting with AI (effective Jan 2026).

This effectively forces AI companion apps to re-engineer engagement loops and content filters. Expect a compliance wave from Character.AI, Replika, and others. (Or not... we'll see.)

🧩 Jargon Buster: Age-gating = verifying a user’s age to apply different content and safety rules.

OpenAI’s Custom Chips With Broadcom Begin A New Hardware Era (Update)

OpenAI is collaborating with Broadcom on custom AI accelerators totaling 10GW of capacity. OpenAI will design the chips, while Broadcom provides manufacturing plus Ethernet/PCIe/optical networking for scale.

First racks arrive H2 2026, with full deployment by end of 2029.

Strategically, OpenAI joins Amazon and Google in pursuing in-house silicon to tame costs and supply risk. The big variable: matching Nvidia’s mature ecosystem and developer tooling while avoiding fragmentation across three suppliers (Nvidia, AMD, Broadcom).

🧩 Jargon Buster: Accelerators = specialized chips (GPUs/ASICs) built to run AI math faster and cheaper than general CPUs.

Read more: The Rundown AI, OpenAI, Broadcom

Tools & Launches

OpenAI Agent Builder (How-To) - Build a routing agent that classifies support tickets (product vs. billing), branches to specialized agents, and answers from uploaded docs; ship via ChatKit. Add human handoff and logging before scale. Get started Rundown University

Zapier MCP Bridge - Connect Agent Builder to 8,000+ apps using Model Context Protocol to trigger sheets, forms and CRM updates from natural-language tasks ... use granular scopes and approvals to stay safe.

n8n AI Workflow Builder - Prompt-to-workflow creation to spin up agentic automations. Strong for developer-led orchestration; pair with tests and cost guards.

Today’s Sources: AI Secret, Broadcom, California Legislature, LM Arena, Microsoft, Oura, OpenAI, The Hill, The Rundown AI, Zapier MCP.