AI & LLMs: Research, Capabilities, and Industry
Anthropic Research · www.anthropic.com
Anthropic released a paper outlining its views on US-China AI competition, arguing it is essential for the US and allies to maintain a lead over authoritarian governments like the CCP in AI development.
Anthropic published research on natural language autoencoders, a technique for converting Claude's internal representations into human-readable text to improve interpretability.
Financial Times · www.ft.com
Anthropic will discuss the capabilities of its Mythos AI model with Financial Stability Board members, after the model exposed significant cyber vulnerabilities in financial infrastructure.
DeepSeek's V4-Flash model has revived interest in LLM steering vectors as a practical technique, because its architecture makes activation-level interventions more effective.
Research examining how LLM-generated text is systematically distorting patterns in written language, documenting measurable shifts in vocabulary and style across online content.
AI's Impact on Software Engineering and Work
A post arguing that AI tools speed up coding but don't necessarily accelerate overall software delivery processes, because bottlenecks often lie elsewhere.
Mitchell Hashimoto observes that some companies are in a state of "AI psychosis," making irrational decisions driven by hype rather than sound engineering judgment about AI capabilities.
Financial Times · www.ft.com
Amazon employees are using the company's in-house MeshClaw AI tool for unnecessary tasks to climb internal AI usage leaderboards, highlighting perverse incentives in corporate AI adoption.