• Published on

    Reliability is foundational

    LLMs are already producing billions of hallucinated outputs per week.


    Not because they’re broken—

    but because they’re designed to sound confident, even when they’re wrong.


    At scale, that’s not a small bug.

    It’s a systemic issue.


    There’s a useful lens from Ecological Economics:


    It defines four forms of capital:

    • Natural (environmental)

    • Human (skills, judgment, purpose)

    • Social (trust, institutions)

    • Physical (machines, infrastructure)


    AI systems are powerful physical capital, powered by natural capital.


    But unreliable outputs degrade the rest:


    * Human capital → overreliance on incorrect answers

    * Social capital → erosion of trust

    * Physical capital → wasted compute and bad decisions

    * Natural capital → unnecessary energy use


    So the real question isn’t:

    “How much can we automate?”


    It’s:

    “How do we use AI reliably at scale?”


    At Komplexity AI, we’re building a real-time hallucination detector for LLM outputs.


    In internal testing, our system achieves >0.90 AUC, with real-time inference on a single GPU.


    The goal is simple:

    Give every AI response a reliability signal at inference time.


    * High confidence → proceed

    * Low confidence → route, verify, or intervene


    If AI is becoming a new layer of labor,

    then reliability isn’t optional—it’s foundational.

  • Published on

    Balance as the new frontier: AI and Rossum's Universal Robots

    A century ago, R.U.R. (Rossum’s Universal Robots) gave us the word robot — from the Czech robota, meaning forced labor.


    The idea was simple: create workers so humans don’t have to.


    Later, Buckminster Fuller reframed this in modern terms. Each of us, he argued, commands hundreds of invisible “energy slaves” — machines powered by fossil fuels and electricity doing work on our behalf.


    Today, with AI, this concept is accelerating again.


    But there’s a deeper lens from Ecological Economics.


    It defines four forms of capital:

    • Natural (environmental)

    • Human (skills, health, purpose)

    • Social (trust, institutions)

    • Physical (machines, infrastructure)


    “Robots” and “energy slaves” are physical capital, powered by natural capital.


    The risk is subtle:

    If we over-optimize for physical capital, we can quietly degrade the others.


    – Human capital: loss of skill, agency, meaning

    – Social capital: weaker relationships and cohesion

    – Natural capital: resource depletion


    Čapek’s warning wasn’t just about machines rebelling.


    It was about what happens when we remove ourselves from effort entirely.


    As AI scales, the question isn’t just:

    “How much can we automate?”


    It’s:

    “How do we use these new ‘robots’ to strengthen human, social, and natural capital — not replace them?”


    That balance may be the real frontier.

  • Published on

    How many AI hallucinations are happening globally?

    Image description



    This is a rough, order-of-magnitude estimate based on:


    * reported usage (e.g., ~18B messages/week for ChatGPT alone)

    * published hallucination rates (which vary widely by task)


    Simple model:

    H ≈ p × N

    (p = hallucination rate, N = total AI outputs)


    Assumptions (approximate):


    * Total global AI usage: ~2–4× ChatGPT scale

    * Hallucination rates: ~3% to ~15% depending on domain



    Estimated range (per week, global)


    * Conservative: ~1B hallucinations/week

    * Moderate: ~3–5B hallucinations/week

    * Aggressive: ~7–10B hallucinations/week



    Important caveats


    * These are not directly measured numbers

    * Hallucination rates vary a lot by domain (legal ≫ casual chat)

    * Not every output contains factual claims

    * Some hallucinations are harmless; others matter a lot



    Takeaway


    Even under conservative assumptions:


    AI systems likely produce on the order of billions of hallucinated outputs per week globally.


    And because usage is growing faster than accuracy improves,

    total hallucinations are likely increasing, not decreasing.



    Sources:

    OpenAI usage report:

    https://lnkd.in/g2UeiFBB


    Global AI usage stats:

    https://lnkd.in/gCx85VEf


    Hallucination studies:

    https://lnkd.in/gHNkkyHr

    https://lnkd.in/gbw5JzNb


  • Published on

    ULM... the sound of uncertainty, and the home of Albert Einstein

    Real-time reliability and hallucination detection.

    Image description
  • Published on

    Real-time hallucination detection is LIVE! Testers wanted


    If you’re building with LLMs, I’m looking for a few people to test a real-time hallucination detector.


    One of the biggest gaps in current LLM systems: they sound confident even when they’re wrong.


    I’ve built a real-time system that detects hallucination patterns in LLM outputs.

    Early prototype is now running — testing it on real examples (see below).



    Image description