Crazy Wisdom
Episode #483: The Limits of Logic: Probabilistic Minds in a Messy World
- AI agents
- LLMs as brains with tools
- agentic process automation
- reliability and repeatability
- non-deterministic versus deterministic processes
- probabilistic reasoning
- human attention
- nonlinear thought
This video isn't available here. Try YouTube, or listen to the audio below.
Watch on YouTubeAbout this episode
Timestamps
- 00:00Derek Osgood explains what AI agents are, the challenge of reliability and repeatability, and the difference between chat-based and process-based agents.
- 05:00Conversation shifts to probabilistic vs deterministic systems, with examples of agents handling messy data like LinkedIn profiles.
- 10:00Stewart Alsop and Derek discuss how humans reason compared to LLMs, token vs word prediction, and how language shapes action.
- 15:00They question whether chat interfaces are the right UX for AI, weighing structure, consistency, and the persistence of buttons in knowledge work.
- 20:00Voice interaction comes up, its sci-fi allure, and why unstructured speech makes it hard without stronger memory and higher-level reasoning.
- 25:00Derek unpacks OpenAI’s approach to memory as active context retrieval, context engineering, and why vector databases aren’t the full answer.
- 30:00They examine talent wars in AI, credentialism, signaling, and the difference between PhD-level model work and product design for agents.
- 35:00Leisure and creativity surface, linking downtime, fantasy, and imagination to better lateral thinking in knowledge work.
- 40:00Discussion of asynchronous AI reasoning, longer time horizons, and why extending “thinking time” could change agent behavior.
- 45:00Derek shares how Double O orchestrates knowledge work with natural language workflows, making agents act like teammates.
- 50:00They close with reflections on re-skilling, learning to work with LLMs, BS detection, and the future of critical thinking with AI.
- One of the biggest challenges in building AI agents is not just creating them but ensuring their reliability, accuracy, and repeatability. It’s easy to build a demo, but the “last mile” of making an agent perform consistently in the messy, unstructured real world is where the hard problems live.
- The shift from deterministic software to probabilistic agents reflects the complexity of real-world data and processes. Deterministic systems work only when inputs and outputs are cleanly defined, whereas agents can handle ambiguity, search for missing context, and adapt to different forms of information.
- Humans and LLMs share similarities in reasoning—both operate like predictive engines—but the difference lies in agency and lateral thinking. Humans can proactively choose what to do without direction and make wild connections across unrelated experiences, something current LLMs still struggle to replicate.
- Chat interfaces may not be the long-term solution for interacting with AI. While chat offers flexibility, it is too unstructured for many use cases. Derek argues for a hybrid model where structured UI/UX supports repeatable workflows, while chat remains useful as one tool within a broader system.
- Voice interaction carries promise but faces obstacles. The unstructured nature of spoken input makes it difficult for agents to act reliably without stronger memory, better context retrieval, and a more abstract understanding of goals. True voice-first systems may require progress toward AGI.
- Much of the magic in AI comes not from the models themselves but from context engineering. Effective systems don’t just rely on vector databases and embeddings—they combine full context, partial context, and memory retrieval to create a more holistic understanding of user goals and history.
- Beyond the technical, the episode highlights cultural themes: credentialism, hidden talent, and the role of leisure in creativity. Derek critiques Silicon Valley’s obsession with credentials and signaling, noting that true innovation often comes from hidden gem hires and from giving the brain downtime to make unexpected lateral connections that drive creative breakthroughs.
Episode transcript
Welcome to the Crazy Wisdom Podcast. This podcast is for you. If you have an insane drive to find the truth of things, it's not the good answers that we seek, but the good questions. I interview a range of different guests from many different fields, all with the intention to uncover the simple truths that are hidden in plain sight. Most people don't want to go there. I go there, my guests go there, and you benefit. Please let me know if you enjoy these episodes and as always, subscribe on itunes, Spotify or wherever you listen to the podcasts. Welcome to the Crazy Wisdom Podcast. I've got Derek Osgood and He is the CEO of DoubleO AI and they are an AI agent platform helping non technical teams turn any manual process into fully automated systems using teams of AI agents simply by describing the process in plain English. Welcome to the show.
Oh man, what's not the hardest part? It's really easy. Like, I think the hairy thing is not so much like, can you create, you know, basic agents? It's more, it's always, you know, kind of the last mile of this. And like, how do you get, how do you create an agent that's able to run with high degrees of reliability, accuracy, repeatability? I think everybody is. It's very easy to go from zero to demo. And I think when you kind of break down what an agent is at the most primitive level, ultimately an agent's just like an LLM as a brain with some kind of prompt and then access to some tools that allow it to take action kind of in the universe. And so it can kind of decide what to do and, you know, go do that thing and whatever tools it's got access to. And I think it's very easy to like, have, you know, AI at this point, like write a prompt and, you know, have it write the kind of like decisioning process that the agent's going to follow. The tricky thing is then, you know, ensuring that it always follows the same process and making sure that it actually does the things you want it to do in the right way. And I think that's the hard part when you talk about like building, you know, quote unquote, agentic process automation especially, but also just like agents in general, because there's also different form factors here you have chat based agents which are, you know, not necessarily designed to like run a repeatable process. And then you also have, you know, process agents where it's more of a workflow and you're trying to like run something consistently repeatedly using the non deterministic, you know, power sleeping stack.
Your first full transcript is free. After that, an email opens every transcript in the index — a list of readers we can write to, not a guest book.