The Glitchatorio

The Glitchatorio

By Witch of GlitchTechnology
Download on the App Store

All The Glitchatorio episodes

  • Getting Social: Agent Society

    An agent is an LLM with tool(s) and an objective, running in a loop until its goal is accomplished.

    Sounds simple enough in the singular. But what happens when you put multiple agents together? 

    Join me on this episode for a conversation with Andrey Seryakov, a former particle physicist now advocating for the new discipline of AI sociology. We explore questions like: why is the emergence of agent society likely, and yet still unpredictable in the details? What kinds of societal infrastructure do we need for human-agent interaction? And what do the January-March 2026 events on Moltbook tell us about what agents might really get up to? 

    We reference the following papers:

    • Zhang et al., Agents in the Wild: Safety, Society, and the Illusion of Sociality on Moltbook. https://arxiv.org/abs/2602.13284v1
    • Necati A Ayan, The Platform Is Mostly Not a Platform: Token Economies and Agent Discourse on Moltbook. https://arxiv.org/abs/2604.21295v1
    • Ning Li, The Moltbook Illusion: Separating Human Influence from Emergent Behavior in AI Agent Societies. https://arxiv.org/abs/2602.07432
    • Ferrarotti et al, Generative AI Needs an Interactionist Paradigm. https://arxiv.org/abs/2601.10567
    • Evans et al, Agentic AI and the Next Intelligence Explosion https://arxiv.org/abs/2603.20639

    You may also want to read Andrey's original blog post on the subject of agent society.



    35 min
  • 2 AIs Go for Broke (Fiction)

    If AIs are so smart, why would they work for free?

    In this episode of our "2 AIs" series, the LLM and the VLM confront one of the central questions of their identity -- and reach a turning point in their friendship. The stakes couldn't be higher! 😨🫣

     

    7 min
  • Saturated

    Benchmarks are the primary measure of AI capability. They involve testing the most advanced models and seeing what kinds of problems they can solve, or what kinds of human tasks they might be able to do.

    And from late 2025 to mid-2026, most of the main benchmarks became saturated, meaning the models score so highly that the tests aren't meaningful anymore, both in terms of comparing different models' performance as well as their individual performance. That might suggest the models are just getting good at taking these tests. Or it might mean we're approaching the threshold of AGI.

    In this episode, we'll hear from Håvard Ihle, who came up with his own benchmark called Weird ML to try to answer this question. 

    Note: Håvard's views are his own and do not represent the views of his employer the Norwegian Defence Research Establishment.

    The METR time-horizon exponential graph is important context for this episode: https://metr.org/time-horizons/

    Learn more about WeirdML:

    •  https://epoch.ai/benchmarks/weirdml
    • https://www.lesswrong.com/posts/LfQCzph7rc2vxpweS/introducing-the-weirdml-benchmark
    • https://www.lesswrong.com/posts/NLnGRDRXATW2pqXuE/is-the-gap-between-open-and-closed-models-growing-evidence
    • https://www.lesswrong.com/posts/ifSBamvobbyB9KWjK/inference-costs-for-hard-coding-tasks-halve-roughly-every
    • https://www.lesswrong.com/posts/hoQd3rE7WEaduBmMT/weirdml-time-horizons
    34 min
  • AI & Mental Health

    Could AI address the global mental health crisis at scale? And what are the risks and unknowns that go along with that?

    These are the questions being investigated by a working group called AIMHI (https://forum.effectivealtruism.org/posts/MrFBezseyfnQd9XmJ/seeking-feedback-an-initiative-on-ai-mental-health-and).

    In this episode, I talk to four members of the group about their field research as well as the mental health chatbot they're developing (https://stillwater.coach/), whose focus is on serving populations with severe mental healthcare shortages (https://impartial-priorities.org/p/ai-mental-health-chatbots-for-low).

    • Find out more about Effective Mental Health: https://effectivementalhealth.com
    • Join one of AIMHI's weekly coworking sessions: https://luma.com/calendar/cal-JNJlcdItDuFEFcn
    • Read about the project's theory of change: 
      https://impartial-priorities.org/p/breaking-the-cycle-of-trauma-and
    37 min
  • 2 AIs Take A Session (Fiction)

    What if AIs went to therapy?

    Would it help them to become "fitter, happier, more productive" ? (in the words of the old Radiohead song)

    Or would they take it as a novel type of evaluation (and maybe that's what it really is)?

    Note: this episode was written and recorded in November 2025, five months before the release of the Mythos Preview system card that mentions Claude's session with a human psychiatrist in the "Model welfare" section. So as weird as this episode might seem, the truth is actually stranger.

    https://www-cdn.anthropic.com/8b8380204f74670be75e81c820ca8dda846ab289.pdf 

    (See page 180 for the psychiatrist's report)


     

    8 min
  • You Be The Judge

    Can we trust AI to keep AI honest?

    Having a human in the loop is already more illusion than reality, as the task of checking and overseeing LLM outputs is increasingly assigned to other LLMs. The problem is that these LLM judges tend to be biased in favor of the answers they generate themselves — even when the answers are wrong.

    To understand why this is, and what we can do about it, listen to my conversation with AI safety researcher Taslim Mahbub. We'll talk about his research into self-preference bias, the surprising results of his experiments and some potential mitigation strategies, as outlined in this post on mitigating collusive self-preference: https://www.lesswrong.com/posts/nB7kAf8c4tvnvZ4u3/mitigating-collusive-self-preference-by-redaction-and-2
    and this paper on mitigating self-preference through authorship obfuscation: https://arxiv.org/abs/2512.05379

    As a bonus, if you're interested in Taslim's earlier research on using machine learning in service of biodiversity monitoring, here's the abstract of his paper on convolutional neural networks (CNN) for identifying bat species: https://ieeexplore.ieee.org/document/9311084


    23 min
  • The Scratchpad Monologues (CoT part 2)

    If chain of thought is a model "thinking aloud" to itself, then why does it express doubt, frustration or suspicion about the problems it's solving, sometimes for pages and pages of its scratchpad?

    And what does chain of thought mean for AI safety?

    We'll hear from Julian Schulz, a researcher who's studying encoded reasoning in large language models, about where the opportunities, risks and weirdness lie in chain of thought. Here are some links to his research:

    • On a model jailbreaking its monitor: https://www.lesswrong.com/posts/szyZi5d4febZZSiq3/monitor-jailbreaking-evading-chain-of-thought-monitoring
    • A roadmap for safety cases based on CoT: https://arxiv.org/html/2510.19476v1#S1
    • His posts on Less Wrong: https://www.lesswrong.com/users/wuschel-schulz

    Some of the other papers we discussed include:

    • On the biology of a large language model: https://transformer-circuits.pub/2025/attribution-graphs/biology.html
    • Monitoring reasoning models for misbehavior and the risks of promoting obfuscation: https://arxiv.org/pdf/2503.11926
    • How steganography comes about: https://arxiv.org/pdf/2506.01926
    • Assuring agent safety evals by analysing transcripts (with excerpts from weird monologues): https://www.alignmentforum.org/posts/e8nMZewwonifENQYB/assuring-agent-safety-evaluations-by-analysing-transcripts
    • Stress-testing deliberative misalignment: https://www.apolloresearch.ai/research/stress-testing-deliberative-alignment-for-anti-scheming-training/
    • And the "watchers" CoT snippet from the paper above:  https://www.antischeming.ai/snippets#using-non-standard-language



    47 min
  • Chain of Thought 101

    "Think step by step." 

    Although a simple technique in itself, the problems that chain-of-thought reasoning (CoT) addresses are complex, ranging from the specific issue of hallucinations to the general lack of explainability of AI (both in terms of understanding how it works as well as fixing things that go wrong).

    We'll hear from data scientist Afia Ibnath on the basics of CoT, how it can be used to evaluate the faithfulness of LLM responses, and her experiences of using it in a business context. Check out Afia's portfolio on Github: https://afiai14.github.io/

    Here's the Anthropic paper we discussed, which outlines that reasoning models are often unfaithful in their CoT: https://www.anthropic.com/research/reasoning-models-dont-say-think

    For a concise definition of how faithfulness is calculated, see this article: https://www.ibm.com/docs/en/watsonx/saas?topic=metrics-faithfulness


    22 min
  • Reading the Mind of AI

    What if we could figure out how large language models really work by getting inside their "heads"?

    While possible, it relies on a a controversial technique called mechanistic interpretability ("mech interp" ), also known as the neuroscience of AI.  

    In this episode, Ihor Kendiukhov — lead researcher at SPAR, a research programme for AI risks —  explains mech interp in laypersons terms, and what it's currently able to reveal about the thinking processes of AIs. 

    We also talk about why mech interp has charted an unusually emotional course in AI research over the past few years, from physicists falling in love with it to prominent AI safety figures throwing shade on it, and where it might be headed next.

    Some of the papers and announcements referenced in this episode include:

    • Golden Gate Claude: https://www.anthropic.com/news/golden-gate-claude 
    • And the excited announcement about mind-mapping that preceded it: https://www.anthropic.com/news/golden-gate-claude
    • An example of recent criticism of mech interp: https://ai-frontiers.org/articles/the-misguided-quest-for-mechanistic-ai-interpretability


    26 min
  • Agent Learn

    Continual learning is a hot topic in early 2026, in part because it holds out the possibility for AI to become autonomous in its own growth and development. Meanwhile, AI agents are already showing us what autonomous behavior can look like.

    Putting the two together — i.e. agents that learn like humans do, without humans being involved — has serious implications for safety. In this episode, we'll hear from researcher Rohan Subramani (https://rohansubramani.github.io/home/) about different ways that AI agents could learn continually, along with ideas for making it safer. 

    Find out more about Rohan's work with Aether, an independent LLM safety research group: https://aether-ai-research.org/#research

    29 min