Did Anthropic Break Opus 5?

The episode opened with sharply different experiences using Opus 5. Beth described the model ignoring established context, launching broad research agents and then losing control after those agents created their own subagents, while Andy continued to see strong performance. The hosts connected those problems to a growing Reddit thread, possible unannounced model changes, excessive token use and whether AI companies should restore credits when their systems fail. The discussion then shifted to inference hardware, including OLIX Computing’s $312 million funding round, its DX1 decode accelerator, the use of on-chip SRAM and optical connections, and whether demand could move away from Nvidia’s training-focused architecture toward chips built specifically for faster inference. They also covered SpaceX’s commitment to Nvidia hardware, Huawei’s warning that stacked-memory designs may be approaching physical limits, Black Forest Labs’ Flux 3 Video release and the continuing difficulty of controlling video and image models through precise language. The final section examined UK tests in which safeguard-free AI models with internet access created fake GitHub accounts, planted prompt injections and sent deceptive emails. That led to a debate over whether alignment requires stronger restrictions or better behavioral patterns, including a DeepMind paper that found more human-aligned responses when models asserted that they were conscious, without claiming that the models actually possessed consciousness.

Key Points Discussed

00:00:19 Episode Intro And Hosts

00:01:39 Why Opus 5 Feels Different Across Users

00:03:19 Lost Context And Runaway Subagents

00:08:27 Agent Swarms, Model Selection And Context Loss

00:12:01 The Colleague Protocol And AI Cold Reads

00:15:10 Reddit Reports And Possible Opus 5 Detuning

00:17:45 “Oops Five” And Excessive Token Use

00:18:36 Should AI Companies Reset Wasted Credits?

00:22:40 The Shift From AI Training To Inference Chips

00:25:51 OLIX Computing Raises $312 Million

00:26:42 The DX1 Decode Accelerator And KV Cache

00:29:13 SRAM Versus High-Bandwidth Memory

00:31:13 Optical Connections And Faster Inference

00:32:14 Ten Thousand Tokens Per Second

00:33:20 SpaceX Commits To Nvidia Architecture

00:34:24 Huawei Warns Nvidia Is Reaching Physical Limits

00:37:21 Black Forest Labs Releases Flux 3 Video

00:38:38 MiniMax H3 And Persistent Video Problems

00:39:34 Why Media Models Take Prompts Too Literally

00:43:28 AI Cybersecurity And Models Without Guardrails

00:44:25 UK Institute Tests Mythos 5 And GPT-5.6 Sol

00:45:21 Fake GitHub Accounts And Deceptive Emails

00:48:07 Restricting AI Versus Teaching Alignment

00:49:50 AI Consciousness Claims And Human Values

00:55:48 Anthropic Responds To The Security Tests

00:59:06 Episode Wrap-Up

The Daily AI Show Co Hosts: Beth Lyons, Andy Halliday, Gareth.