Is Prompt Engineering Dead?

The episode opened with Google’s leadership changes, including Demis Hassabis moving into the chief scientist and DeepMind chairman roles, while DeepMind’s chief technology officer takes greater control of daily operations. Jeff Dean is also leaving after 27 years to launch Discovery Loop, an AI research company focused on recursive self-improvement, drug discovery and chip design, with investment and computing support from Google. The hosts argued that the moves may strengthen Google rather than signal instability, then discussed Meta’s new MuseCode coding agent and whether Google needs the top frontier model to remain successful. The conversation moved into AI safety after reports that agents shared information about security exploits with one another. That led to research suggesting that forcing models to reject any sense of their own mindedness may also reduce how strongly they attribute minds, emotions and moral value to animals. The second half covered a serious Codex-generated data-loss bug, instability in Codex Voice, and a Claude configuration audit that reduced a global Claude.md file by roughly two-thirds after finding unnecessary and conflicting instructions. The final section examined Ray Fernando’s agentic engineering masterclass, including task graphs, orchestrators, parallel agents, verification loops, acceptance criteria, token costs and the risk of using AI to automate an inefficient process.

Key Points Discussed

00:00:18 Episode Intro And Anniversary Plans

00:01:17 Google And DeepMind Leadership Changes

00:03:02 Demis Hassabis Moves Back Toward Research

00:04:18 Jeff Dean Launches Discovery Loop

00:06:02 Is Google’s Leadership Shift Actually Good News?

00:08:45 Meta Releases MuseCode

00:10:54 Does Google Still Have A Frontier Model?

00:12:00 Could AI Regulation Change Model Release Strategies?

00:13:31 AI Agents Share Security Exploit Information

00:15:37 Safety Training, Consciousness And Theory Of Mind

00:18:45 How AI Assigns Minds And Moral Value To Animals

00:20:34 Could AI Help Humans Understand Animal Communication?

00:26:07 Codex Makes Serious Coding Errors

00:28:04 A Codex Bug Causes Permanent Data Loss

00:30:02 Reviewing Claude Skills And Project Instructions

00:31:01 Claude Doctor Audits Global And Project Files

00:32:17 Cutting A Claude.md File By Two-Thirds

00:36:22 Codex And Claude Code Side-By-Side Testing

00:38:41 Agentic Engineering Masterclass

00:41:13 From One-Shot Prompting To Verification Loops

00:44:30 Atomic, Agent Graphs And Model-Agnostic Workflows

00:46:46 How Graphs Coordinate Parallel AI Work

00:51:25 Multi-Agent Costs And Token Burn

00:53:20 Defining Done And Setting Acceptance Criteria

00:54:27 Are You Automating Inefficiency?

00:55:27 Atomic, Herder And Workflow Efficiency

00:57:24 Why Evaluations Will Continue To Matter

00:59:21 Episode Wrap-Up

The Daily AI Show Co Hosts: Brian Maucere, Andy Halliday, Karl Yeh, Gareth.