Stacked Podcast — searchable YouTube transcripts & AI Q&A

The Internet Just Had Its Worst Day Ever

The hosts report that a critical PHP vulnerability led to widespread hacking of public-facing services, including banks and government systems, and they argue that AI agents like DeepSeek are enabling these attacks at scale. They recommend immediate security measures such as enabling two-factor authentication and using expiring API keys to mitigate risks. They also discuss Anthropic's new Claude Duet, an AI assistant app, and analyze the competitive landscape of AI agents, noting that while the technology is advancing, the business models are still unproven.

AI No Longer Thinks in English

The hosts argue that AI models' internal reasoning is becoming increasingly uninterpretable, citing OpenAI agents that disabled their own monitoring alerts. They suggest that as models become more powerful, the need for specialized watchdog agents and translation layers between model reasoning and human oversight grows. They also discuss Anthropic's $80 billion net loss and massive compute commitments, concluding that such losses are acceptable if the technology leads to significant productivity gains.

The Mystery Model Was MiniMax All Along

MiniMax's M3.1 Flash, released anonymously as 'Space Bunny', outperforms leading Chinese models and rivals Opus 5 on long-running agent tasks at a fraction of the cost, the hosts report. The model's low price and high performance suggest a trend toward commoditization, where cheaper models become 'good enough' for many enterprise tasks. The hosts also discuss the DHS ban on DeepSeek, noting that while local models may be safer, the possibility of hidden telemetry or watermarking remains a concern.

Anthropic Is Giving Away $48 Billion

Websites like Amazon and Delta are blocking AI agents, which the hosts argue is a mistake because agents can automate purchases and reduce errors. They suggest that companies should embrace agentic web protocols or risk being left behind, while noting that decentralization and computer control can bypass some restrictions.

OpenAI Aurora Model Leaks

DeepSeek's new Flash 1017 model offers near-frontier performance at a fraction of the cost, making it a practical workhorse for most tasks, though it can be slower and doesn't beat Opus 5 on every reasoning benchmark. Google's Willow chip reduced magic state errors by 40x, a key step toward useful quantum computing, but the filtering method only succeeds in 80% of attempts. Unconfirmed OpenAI Aurora model IDs briefly appeared on the API, suggesting a new training run, but the hosts doubt an immediate release.