The AI Relay

Andrew connected an AI agent to an AWS account to troubleshoot a deployment error. The agent found the root cause in minutes — a specific policy ID blocking the install — and recommended a fix: exempt the policy from the rule. Andrew knew what that policy was for. If he had run the command, he would have deployed thousands of non-compliant machines across the enterprise.The AI was right. The fix would have worked. It was also the completely wrong thing to do.That story opens a conversation about what actually happens when AI gets good at your job — not the headline version, but the granular version. AI doesn't take roles. It gets better at specific tasks inside roles, one at a time. The domain knowledge sitting around those tasks — the context the model can't see — is where the human's value moves when everything else becomes a commodity.Jonathan's frame: make a list of everything you do. Start checking off the things an agent could handle. Whatever's left is what you're actually selling. Most people haven't done that audit. And the ones who wait longest to do it do it under the most pressure.Andrew's frame: in 2026, the most valuable person in any organisation is the one who combines domain expertise with AI fluency. Not one or the other — both. That's already true in the way that knowing Excel used to be a differentiator: it wasn't on every job application, and then one day it was, and suddenly not knowing it was disqualifying.The conversation also gets into what "managing agents" actually means in plain language (not MCPs and Python — five conversations running simultaneously, each with its own brief), the hiring question (how do you screen for AI fluency if the interviewer doesn't have it?), and what the manager and the engineer have in common now that they didn't a year ago.The episode ends where it should: not with AI displacement panic, but with a more specific and more useful question. If AI gets good at the task, and the task is what your role was built around, what's the thing underneath the task that was always the actual point?New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

One enterprise reportedly spent $500 million on AI in 30 days. Uber burned through its annual AI budget in four months after creating an internal leaderboard that rewarded teams for using the most AI. Jonathan saw those stories and could not let them go.But Andrew's response is not simply “enterprise should have known better.” His counterargument is that the math looks very different from inside a large company. If AI turns a $50,000 professional-services project into a $1,000 token bill, then whether that same work could have been done for $250 may look like a rounding error. The savings are so large that efficiency does not feel urgent — at least not yet.That tension becomes the center of the episode. Jonathan argues that token literacy is not optional; it is a discipline that compounds, and the time to learn it is before waste is already baked into the organization. Andrew argues that enterprise is still in the adoption phase, where the bigger problem is getting people to use AI at all — and where the value created by even inefficient AI use can still dwarf the spend.They also get into how enterprise AI adoption actually unfolds: internal chatbots, API layers, individual keys, coding agents, and eventually employees applying AI to work that budgets were never designed to forecast. As Andrew puts it, the budget may have been made when AI meant “a chatbot with a cute little RAG on it.” A year and a half later, AI means something much broader.The conversation starts with blown budgets and token costs, but it ends somewhere more human: with Andrew telling Jonathan, “You don't understand how free you are.” Individuals can tune their tools, workflows, and habits almost immediately. Enterprise has budgets, silos, review boards, procurement, and thousands of people still learning what AI is for.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

If you've ever wondered why big companies seem slow to adopt the AI tools everyone's excited about — the answer probably isn't what you think.Andrew works inside enterprise AI and has been building agents there for years. Jonathan builds AI agents as an enthusiast, outside any corporate structure. In this episode they compare notes and discover they've been using the word "agent" to mean completely different things. Andrew walks through why enterprise teams are still building the same narrow, scoped agents they were building two years ago — not because they're incompetent, but because of real organizational forces: developer gatekeeping, risk aversion, and a cost-of-automation calculus that AI just recently changed. He also takes apart the popular "enterprises can't use frontier AI because of data privacy" argument — turns out AWS Bedrock and Azure AI Studio already let companies run Claude or GPT privately, with no data sent to the AI provider. The actual gap between local and cloud-hosted AI, he argues, comes down to how you're billed. Jonathan pushes back from the enthusiast side and finds, by the end, that both camps are solving different problems and neither fully understands what the other is dealing with.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

If you've been using AI seriously for a while and still feel like you're falling behind — this episode is specifically for that.Andrew and Jonathan both came into this recording tired. Not struggling-with-the-tools tired. Keeping-up-with-the-pace tired. And the conversation they had is one of the more honest things either of them has put on record about what it actually feels like to be an active AI practitioner right now.Andrew traces what he calls the satisfaction plateau: once your workflow genuinely works — once you can get AI to do what you need it to do and recover when it doesn't — the bar for switching tools rises, but the anxiety of not switching stays the same. He explains why that gap is the real source of exhaustion, and why the answer isn't to chase every new release but to ask a harder question: is this better enough to be worth what it would cost me to move?They also get into what "portability" actually means when you've built a real stack (it's an architecture problem, not a settings toggle), why agents should be thought of as the disposable layer while MCPs, skills, and durable tooling are what you keep, and what the human-in-the-loop is actually for inside enterprise AI — which turns out to be less about safety than most people assume.The episode ends with both hosts exhausted — and both saying the last few days have been some of the best AI days they've had. Those two things turn out to coexist pretty naturally.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

When your company talks about deploying AI agents and when you think about your own AI setup, you might both be using the word agent — but you're probably not describing the same thing at all. This episode maps out the difference between the small, task-specific agents being built inside large enterprises and the general-purpose, personality-driven agent setups that people like Jonathan are building for personal use. Andrew describes what enterprise agents actually look like: lightweight pieces of code that intercept a system alert, pull logs from AWS, and send remediation steps to an ops team — no conversation, no personality, no memory. Jonathan describes the other end: a persistent orchestrator that knows everything you've worked on and stays oriented across sessions. Both are agents. They just live in completely different worlds, built on different philosophies about what AI is for.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

You hand it a business plan, ask what it thinks, and it tells you it's brilliant — which is almost never the honest answer. This episode is about one of the most common mistakes people make with AI: asking open-ended opinion questions and then trusting the response. Andrew breaks down three specific reasons why "what do you think?" reliably gets you a bad answer — hedging, sycophancy, and false confidence — and what to ask instead. Questions like "what would a skeptic say?" or "what's the strongest case against this?" unlock far more useful feedback than any general opinion prompt. Jonathan connects this back to something the show keeps returning to: the clearer and more specific your input, the more you can actually trust what comes back. Good prompting isn't dead — it just looks like being precise about what you're actually asking for.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

Starting a new chat and feeling like the AI has no idea who you are is one of the most frustrating parts of using these tools — but that experience is changing fast. This episode gets into how memory in AI systems actually works, from the live context window that fills and clears during a session to the persistent memory layers that platforms like ChatGPT and Claude are building behind the scenes. Andrew explains the RAM-versus-hard-drive distinction that makes sense of why a model can seem to remember something from last week but still forget an instruction you gave it ten minutes ago. Jonathan shares what it's like to work with an agent that has layered memory files and actually stays oriented across sessions. If you've been frustrated by having to re-explain yourself every time, this episode explains why and what's being done about it.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

Opus 4.7 is out and the internet is excited, but your workflow is running fine on 4.6 and you really don't want to break anything. This episode helps you think through when a new model release actually matters and when it doesn't. Andrew and Jonathan break down what a model update really is — it's the engine in the car, not a software update — and why a better engine doesn't always mean better results for your specific task if it costs more tokens to run. They walk through the logic of model evaluation: how to weigh performance gains against token cost increases, why switching models mid-workflow carries real risk, and why standing still in AI is its own kind of risk too. The rule of thumb they land on is worth keeping: the right amount of AI is always the least amount of AI.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

You're laughing at a joke a computer made, or you catch yourself genuinely irritated at a response, and then you think — wait, this is just a machine. So why does it feel like talking to a person? This episode sits with the strangeness of conversational AI and actually tries to explain it, from the uncanny valley effect to the way human language carries meaning beyond just grammar. Andrew and Jonathan explore the spectrum between treating AI as a cold tool and treating it like a colleague, and whether one approach actually gets you better results than the other. They also get into how tools like Cowork change the equation — once a workflow is set up, the weird part mostly disappears and you're just using software. But getting there still requires that strange window of talking to something that talks back.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.

A lot of people think using AI well means knowing the right prompt formula. In this episode of The AI Relay, Andrew Signore and Jonathan Spanos unpack why that idea is becoming less true — and why clear goals, better context, and stronger judgment matter more than “magic words.”They explore how prompting used to feel like a step-by-step instruction game, and how newer AI tools are changing that by handling more context, more memory, and more multi-step work. The bigger shift is this: the real skill is no longer just writing clever prompts — it’s defining the objective clearly enough for AI to help in a useful way.In this episode:- why prompts are not magic spells- why clear goals beat clever wording- how AI capabilities change the way people should work with models- why asking better questions may matter more than memorizing prompt tricks- how to think in terms of outcomes instead of only instructionsIf you’ve ever wondered whether you’re “bad at prompting,” this episode is for you. The AI Relay is about understanding AI in plain English — so non-technical listeners can build better mental models and use these tools more effectively.What changed for you first: your prompts, or your understanding of what AI can actually do?

Most people say they “use AI,” but that can mean very different things. In this episode of **The AI Relay**, Jonathan Spanos and Andrew Signore break down the difference between using AI as **search**, using it as an **assistant**, and using it as an **agent**—and why those are not the same job.They explain, in plain English, how AI has evolved from answering questions, to helping with real work, to taking action with tools on your behalf. Along the way, they talk about tool use, web search, autonomy vs. automation, Claude Co-Work, and why **human in the loop** still matters.In this episode, they explore:- what it means to use AI like search- what makes AI an assistant instead of just a chatbot- what makes something truly agentic- why tool use changes what AI can do- the difference between automation and autonomy- why “human in the loop” still matters when AI starts acting- how better mental models can make you a better AI userIf you’ve ever said “I use AI” without being totally sure what role it is actually playing, this episode will give you a much clearer way to think about it.Hosted by Jonathan Spanos and Andrew Signore.  **The AI Relay** explains AI in plain English for beginners and early users.**Which role do you use AI for most right now: search, assistant, or agent?**

Why does AI sometimes give a great answer—and other times make something up completely? In this episode of The AI Relay, Jonathan Spanos and Andrew Signore explain AI hallucinations in plain English for people without technical backgrounds.They break down why large language models predict words instead of “thinking” like humans, why that can lead to confident wrong answers, and what it means for everyday users of tools like ChatGPT, Claude, and Gemini.In this episode, they explore:- what an AI hallucination actually is- why prediction can lead to made-up answers- why AI can sound confident even when it is wrong- what RAG (retrieval augmented generation) means in normal language- why “human in the loop” still matters- how to use AI better without being afraid of itIf you’ve ever wondered why AI can feel brilliant one minute and unreliable the next, this episode will give you a much clearer mental model for what’s happening.Hosted by Jonathan Spanos and Andrew Signore.The AI Relay explains AI in plain English for beginners and early users.What’s the strangest or most confidently wrong thing AI has ever told you?

In this episode of **The AI Relay**, Jonathan Spanos and Andrew Signore use a practical grocery-shopping app built live with AI to show what actually leads to better results: clear context, a defined problem, a useful outcome, and everyday language that is more subtle and intentional than it first appears.Rather than treating AI like a magic trick or a search box, they break down the human side of the process — what we still need to do, what we can do better, and why clearer thinking often leads to better AI output.In this episode, they explore:- how to get better results from AI without sounding technical- why clear context matters more than clever wording- what makes an AI request useful in everyday life- how natural conversation can still guide a model effectively- why the human role still matters even when AI does most of the work- how to think with AI, not just ask it questionsIf you’re curious about artificial intelligence, already using ChatGPT or Claude, or trying to understand how to use AI more effectively in daily life, this episode is for you.**Hosted by Jonathan Spanos and Andrew Signore**  **The AI Relay — AI explained in plain language for beginners and early users.**What’s one small change that helped you get a better result from AI?

Most people are still using AI like a search engine. In this episode of **The AI Relay**, Jonathan Spanos and Andrew Signore explain why that misses the bigger picture.They break down the difference between the **AI app** and the **AI model**, why **ChatGPT, Claude, and Gemini** can feel different, and why those differences matter less than most beginners think. The bigger idea is this: AI is not just a place to ask questions. It can become a practical tool for everyday thinking, planning, problem-solving, and reducing friction in real life.In this episode, they explore:- AI apps vs AI models- ChatGPT vs Claude vs Gemini- why free AI tools are worth trying- why prompting is only the beginning- how everyday people can use AI beyond search- what most people are still missing about large language modelsIf you’re curious about artificial intelligence, already using ChatGPT, or trying to understand how AI actually works in everyday life, this episode is for you.**Hosted by Jonathan Spanos and Andrew Signore**  **The AI Relay** — AI explained in plain language for beginners and early users.What do you wish AI could help you do better in daily life? Leave it in the comments and it may shape a future episode.

If you've been nodding along when people talk about AI but still feel like you're missing the full picture — this is where to start.Jonathan and Andrew trace the full arc of AI from its origins to the moment it became impossible to ignore — the shift from machine learning to generative AI that gave everyone a ChatGPT or Claude account and suddenly made the technology feel personal. They demystify the language: what an LLM actually is, what model weights are (spoiler: not as scary as they sound), and why "natural language as the operating system" is the most important thing to understand about where AI is right now.No background in tech required. That's kind of the whole point.New episodes every week — full video on YouTube. Hosted by Jonathan Spanos and Andrew Signore.