Monday, September 28, 2026 · Run 2 · 4:12 PM ET

AI Labs Sprint Toward Self-Improving Models Even as Their Own CEOs Call for a Slowdown

2 stories from 1 newsletter

TL;DR

The jump from GPT-5 to GPT-6 took 13 months versus roughly three years for GPT-3 to GPT-4, and Anthropic says Claude now leads 26% of its own research work with about 30,000 agents running research and engineering at any time.
A new Claude tutorial shows how to turn a job description into a structured interview kit, competencies, questions, and scoring guide, with a reusable prompt.
Today's other big stories, Microsoft's Copilot overhaul, the industry-wide AI-agent security incident count, and Anthropic's Pentagon appeal loss, were already covered in this morning's run; no repeat here.

AI Tips1

Build a structured interview kit from a job description with Claude

Open Claude and upload the job description you want to build the interview around.
Ask Claude to extract the role's core competencies, responsibilities, technical skills, and behavioral requirements.
Tell Claude which interview stage you're designing (screening, hiring-manager, technical, or final-round), then run the prompt below.
Review the output, remove anything duplicated or generic, and ask Claude to build a one-page interviewer scorecard from the same criteria.

Prompt

Using this job description, create a structured interview kit for a [role]. Identify 6–8 job-relevant competencies and write 2 behavioral questions for each. Add useful follow-up probes, clear scoring criteria from 1–5, and examples of strong, average, and weak evidence. Keep all criteria directly tied to the role. Organize the kit into Competency, Main Question, Follow-Ups, Evidence to Look For, and Scoring Guide.

Source: Superhuman AI

AI News1

AI labs are racing toward self-improving systems even as their own leaders publicly call for a slowdown

OpenAI's jump from GPT-3 to GPT-4 took nearly three years; GPT-5 to GPT-6 took 13 months. Anthropic says Claude now leads 26% of its research work with roughly 30,000 agents working on research and engineering at any given time, OpenAI claims an AI-powered "research intern" working under human direction, and DeepMind revealed Dream RSI, a technique that uses simulated dreaming to improve performance while cutting compute costs.

Why it mattersThe gap between AI-safety rhetoric and shipping speed is a live talking point for any AI-adoption or governance conversation. When a client asks whether to slow their own AI rollout, the industry's own leaders preaching caution while accelerating is worth surfacing.