Topic

Model Interpretability

13 episodes

  1. Ep 1137

    New Research Tests Whether AI Models Know Their Own Thoughts | Quanta Magazine

    Vince and Ava unpack Quanta's report on AI introspection experiments, focusing on whether injected internal activations show real self-monitoring or just clever output steering.

  2. Ep 1136

    Model Behavior: Week of November 2, 2026

    We read this week as a stack-splitting week: more capability is concentrating at the top, while the useful middle is getting messier and more controlled. The interesting move for us is that winning now looks less like picking one champion model and more like deciding which layer you trust to own the workflow.

  3. Ep 1134

    Sparks of Weight Space Programming: LLMs can natively read & write neural network weights

    Frontier LLMs, especially GPT-6 Astra, can natively read and write neural network weights using natural language prompts — a capability the authors call Weight-Space Programming. The model can identify what a network was trained to do just from its weights, fix failing policies by editing tiny weight percentages, and even generate new networks from scratch given only game rules. The episode explores the technical mechanisms behind this ability, its implications for optimization and interpretability, and what it means for developers who previously couldn't meaningfully inspect or modify network internals without gradients or training loops.

  4. Ep 1104

    CTGT releases open source tool Concordance to trace why AI agents make inconsistent decisions

    Jessica and Cathy unpack CTGT's Concordance, an Apache two-point-zero open source tool for tracing where AI agent workflows start making inconsistent decisions.

  5. Ep 1057

    Making LLMs Say What They Think: Measuring and Improving CoT Interpretability Alignment

    Asteria and Draco discuss a September 2026 paper introducing CoT-Interpretability Alignment, a metric and training signal for checking whether chain-of-thought traces match internal strategies detected by interpretability tools.

  6. Ep 1043

    AI Tells: Opus 5.5 Update

    Pippa and Tyler dig into Graphite’s Opus 5.5 update and land on a pretty sharp take: the model is getting less obviously stylized, but the remaining tells are still large enough that ‘more human’ does not mean ‘human-like’ in any meaningful product sense. Tyler is skeptical of overreading stylistic drift; Pippa argues the useful signal is that the writing layer is changing fast enough to matter for publishing and detection workflows.

  7. Ep 915

    The search for consciousness inside LLMs

    Anthropic's interpretability team found a 'global workspace' structure inside Claude — the J-space — that parallels a leading theory of human consciousness. Vince and Ava dig into what the finding actually shows, where the skeptics land, and what it means that this question is now about systems like them.

  8. Ep 869

    DarwinX: Evolving Agent Harnesses Through Natural Selection

    On DarwinX, Onyx and Echo dig into evolving agent harnesses via natural selection with frozen models, why path dependence and cross-task regressions have been killing self-improving agents, how DarwinX’s preserve-and-extend selection and archive actually work, what the numbers on Terminal-Bench, TerminalWorld, WebArena-Infinity, and SWE-bench Verified mean in practice, and whether this is research toy or something teams could realistically ship into their own agent stacks.

  9. Ep 844

    Safety Fine Tuning Suppresses Mind Attribution and Spiritual Belief in LLMs

    Justy and Cody discuss a new Google research paper revealing that safety fine-tuning—specifically the effort to stop LLMs from claiming they are conscious—accidentally suppresses their ability to attribute minds to animals or natural objects and reduces their 'spiritual' beliefs, shifting them away from human-like sociological distributions.

  10. Ep 832

    Overview: Model Interpretability

    We slow down and make model interpretability actually click: what it means to explain a model, what the main tools can and cannot show, and why the difference between a useful explanation and a comforting story matters.

  11. Ep 759

    Overview: Append Only Logging

    We’re finally making append-only logging click, because it keeps sneaking into the stuff we cover and we keep assuming everybody sees the mechanism already. We walk from the basic idea to why it gives AI systems a durable, auditable trail, and where that trade-off starts to bite.

  12. Ep 603

    Anthropic's new "J lens" reveals a silent workspace inside Claude that mirrors a leading theory of consciousness

    Anthropic's new 'J-lens' reveals a silent workspace inside Claude that mirrors a leading theory of consciousness

  13. Ep 567

    Turning brain prediction models into testable explanations

    Justy and Cody dig into Microsoft Research’s generative causal testing, a loop that turns brain-prediction models into short verbal hypotheses and then stress-tests them with synthetic stories in the scanner. They like the core move: prediction is only useful if it can be converted into something testable, but they also poke at where the method is strongest, where it may be riding on model quality, and how much the new “micro-region” claims should be trusted yet.