Cal Newport dissects recent Anthropic research that found a 'hidden space' where Claude processes concepts in a way resembling the global workspace theory of consciousness. He provides a grounding tutorial on how LLMs work, explains the annotation methods used, and argues that the findings are interesting but don't support claims of AI consciousness. He cautions against media hype and suggests we're still far from machines having private thoughts.
This summary was generated from show notes and public descriptions, not from a full transcript review. Details may contain inaccuracies.
Highlights
•
Anthropic discovers a 'hidden space' in Claude13:50
Anthropic's new research paper reveals a 'hidden space' where Claude maintains conceptual processing during a delay period, resembling the global workspace model of consciousness.
Cal provides a high-level tutorial on large language models, explaining that they are pattern-matching systems completing next-token predictions based on training data.
The study's methodology relies on extensive human annotation of model states, training probes to detect when specific abstract concepts are being 'held' internally.
Articles from Axios and MIT Technology Review framed the research as evidence that Claude might be conscious or having secret thoughts, a leap Cal critiques.