Machine Dreams

When you ask an AI a question about your company's data, it looks something up before it answers. Most of the time that works. But often it joins the wrong tables and returns an answer that looks completely reasonable and is completely false, and nothing catches it. There is no error message. This week on Machine Dreams, Sam Adebayo and Leah Stern sit down with Weimo Liu, CEO and co-founder of PuppyGraph https://puppygraph.com​ Weimo previously worked on query systems at Google and spent time at TigerGraph, and he now helps write the international standard for how these systems get queried.PuppyGraph's approach is to check the question before the AI runs it. The product was designed for agents as the user, not humans, and a semantic layer sits over the enterprise data so the AI only needs to know that a teacher teaches a student, not which table holds which or how to join them. The engine handles the deterministic mapping and guarantees accuracy, so the AI reasons and something outside the model verifies. Weimo also covers the Sonnet 4.6 inflection, where early work with customers including AMD became reliable after proper context engineering, how Coinbase runs long-running compliance agents in natural language without writing a single query, and why chaotic enterprise data needs an interactive human-in-the-loop agent rather than automation.He closes with the warning that stayed with us. One AI agent writes a financial report, another AI agent reviews it, and no human looks at how it was generated, so nobody can know whether the answer is right. He contrasts this with code, where test cases can cover the gaps, and names the human factor plainly: with a tool this powerful, nobody wants to check every detail.

Is AI really conscious? This week on Machine Dreams, Sam Adebayo and Leah Stern sit down with Andy Thurai, founder and principal analyst at The Field CTO https://thefieldcto.com/ an independent technology advisory firm helping executives turn complex technology into clear business strategy. Andy brings more than 25 years of enterprise technology experience, with senior roles at IBM, Intel, and Oracle, and now advises business and technology leaders on AI, observability, and modernization.Andy's answer to the consciousness question is careful and grounded. Machines today have simulated awareness, not self-awareness. His sharpest illustration: throw hot water on a dog and it reacts immediately because it feels pain. Throw hot water on a robot and it only reacts if someone trained it to. That gap is the difference between real consciousness and simulation, and it is also where AI fails, when unknown unknowns arrive. He then offers his most provocative claim: the rush to prove AI is conscious may not be about philosophy at all. If a model can be shown to act on its own, its creators can step back from liability when it causes harm. He also draws a firm line on ethics. When an AI refuses an unethical prompt, it is not consulting a conscience. It is calculating which answer yields a higher reward. It is, in his words, an advanced calculator doing statistical ethics.He closes with practical advice for enterprise leaders. Work backwards from the problem instead of choosing a model and hunting for uses. Treat the LLM as a reasoning engine, and put safety, compliance, and guardrails outside the model in a deterministic control plane. Do not ask a probabilistic system to police itself, because it is architecturally impossible. His summary line: don't trust, always verify.If you are trying to separate the real AI risks from the noise, this conversation is a grounded place to start.

Everyone is worried AI will destroy us. This week on Machine Dreams, Sam Adebayo and Leah Stern sit down with Rob T. Lee, Chief AI Officer and Chief of Research at the SANS Institute: https://www.sans.org/. Rob is known as the godfather of digital forensics. He was a founding member of the US military's first information warfare unit, served with the NSA and the CIA, led threat intelligence at Mandiant, created the original SIFT workstation, and has trained more than 100,000 security professionals.Rob places the current panic in historical context. His parents practiced duck-and-cover drills for a nuclear war that never came. But his real argument is more practical: cybersecurity has been containing the worst malware ever written for 25 years, and none of it has escaped. The last real containment failure was the Morris Worm. His claim is that frontier AI labs are missing that know-how at the table, and that if they built containment the way security teams do, we would all be in better shape. He also addresses the learning gap. Setup is getting easier, but the awareness curve is the new problem. People are handing over bank accounts and private data to agents without asking what happens when the agent gets it wrong.He closes with a counterintuitive finding. He has been tracking entry-level AI jobs for six months, and the data shows flat or slightly increased numbers, not decline. The jobs will change, and the new role might be a junior analyst managing five SOC agents, but they are not disappearing. His hope is that AI, built with proper security involvement, makes the digital world safer. His worry is not apocalypse but culture: likeness rights, deepfakes, and schools that ban AI instead of teaching kids to use it well.If you are trying to figure out where the real AI threat actually lives, this conversation is a grounded place to start.

Salespeople spend most of their week on everything except selling. Research, CRM updates, meeting prep, follow-up emails. Only about a fifth of their time goes to actually talking to customers. This week on Machine Dreams, Sam Adebayo sits down with Johan Torssell, co-founder and CEO of Spiich: https://spiich.ai/ a Stockholm-based company building AI agents for B2B sales teams. Johan founded Spiich in June 2025 with CTO Dennis Hadzialic, and the company is backed by operators and founders from across the AI and enterprise software world, including investors behind Lovable and Neo4j. Before Spiich, he turned down offers from companies like Palantir and Meta because he wanted to build this now.Johan's position is simple and increasingly contrarian. Sales is intrinsically human. People buy from people. An agent can qualify a lead, research a company, draft a follow-up, and keep the CRM current, but it should not press send or close the deal, because that is where trust lives. He points to companies that tried to replace salespeople with AI and had to pull back and hire reps again. He also draws a clear line on autonomy: agents can act freely on anything reversible, but the moment a decision touches a relationship, the human should be in the loop.Looking ahead, he expects two kinds of salespeople to survive: those who build systems and those who build relationships. He also thinks the buyer side is changing, and that in five to ten years, agent-to-agent purchasing will shift sales from relationship and trust toward data.If you are a sales leader trying to figure out where AI actually belongs in your pipeline, or a founder wondering what to automate and what to protect, this conversation is a grounded place to start.

Cameras used to have one job: record what happened. Now they are starting to make sense of what they see while it happens, and that raises a bigger question. What does it mean for a camera to understand?This week on Machine Dreams, Sam Adebayo sits down with Prof. Ron Kimmel, Chief Scientific Officer at Lumana: https://www.lumana.ai/ and professor of computer science at the Technion. Ron has spent decades in computer vision, image processing, and 3D analysis, and he co-founded InVision, the company Intel acquired to form the foundation of its RealSense line. He is also the founder of the Technion's Geometric Image Processing Lab.Ron makes the case that AI has largely closed the classical open problems in computer vision, and that the remaining bottlenecks are often our own hardware choices. He points out that most cars today, including his Tesla, rely on a single forward-facing camera, when biological evolution settled on two eyes hundreds of millions of years ago. He also holds a firm line on what these systems can and cannot do. Cameras can flag patterns and correlations. They cannot find causality or predict the future, and he thinks we should stop asking them to. He also shares work from computational pathology, where AI reading tumor images has matched what was previously only possible through DNA analysis. The research was published in The Lancet Oncology, and it points to patients potentially avoiding radiation or chemical treatment.If you are a business leader wondering what your existing cameras could already tell you, or a builder trying to understand where machine vision actually stands, this conversation is a grounded place to start.