Learn AI in Bits

Dan W

AI explained in bits. Each episode takes one concept, like tokens, embeddings, hallucinations, or prompt injection, and explains it in about five minutes. No jargon, no filler. Just the idea, why it matters, and what to remember.

If you're curious about AI or already building with it, you'll come away understanding how these systems work.

One concept. Five minutes. That's the whole show.

Play

Open in

30 episodes
Avg 6 min
English

Episodes 30

Newest Oldest Longest

All20 Unplayed20 In progress0 Played0

Counted on this page — what you have heard stays on this device, so it is not something the list can be paged by.

S1 · E38
Yesterday · 4 min
038 - Why Are People Saying Opus 5.0 Sucks?
Why are developers saying Claude Opus 5 sucks when Anthropic's benchmarks describe it as one of the strongest models available? This episode examines the gap between benchmark capability and developer experience, focusing on complaints about over-engineering, instruction following, excessive autonomy, observability, and reliability. REFERENCES Anthropic — Introducing Claude Opus 5 https://www.anthropic.com/news/claude-opus-5 Anthropic — Claude / Opus https://www.anthropic.com/claude/opus Anthropic — Newsroom https://www.anthropic.com/news GitHub — Claude Code Issue #87491: Opus 5 treats direct instructions as negotiations https://github.com/anthropics/claude-code/issues/87491 GitHub — Claude Code Issue #85395: Opus 5 regression in following instructions https://github.com/anthropics/claude-code/issues/85395 GitHub — Claude Code Issue #85633: Opus 5 fabricates requirements and acts without permission https://github.com/anthropics/claude-code/issues/85633 GitHub — Claude Code Issue #84933: Long Opus 5 agentic turns with limited visible assistant text https://github.com/anthropics/claude-code/issues/84933 Reddit — Claude has started over-engineering every task https://www.reddit.com/r/ClaudeAI/comments/1vod9ri/claude_has_started_overengineering_every_task/ ExplainX — Claude Opus 5 Over-Engineering: Reddit Reaction https://explainx.ai/blog/opus-5-over-engineering-reddit-reaction-august-2026 Voice narration is AI-generated.
S1 · E37
Thursday · 5 min
037 - CLAUDE.md vs Skills vs MCP in Claude Code
Claude Code gives developers five different places to put an instruction: a CLAUDE.md file, a Skill, an MCP connection, a Hook, or a permission setting. It is easy to shove everything into CLAUDE.md because it is the most familiar option, and just as easy to end up with a bloated file that Claude follows inconsistently. This episode builds a practical map for deciding which mechanism fits a given piece of information or behavior. The episode covers CLAUDE.md as the home for persistent, project-wide guidance, loaded at the user, project, and local scopes plus nested directories, with a .claude/rules folder for instructions scoped to specific file paths. It explains Skills as reusable procedures stored in a SKILL.md file, where the body only loads into context when the Skill is used or judged relevant, making them a better fit for longer workflows than CLAUDE.md. It walks through MCP, or Model Context Protocol, using a Jira ticket example (looking up and updating issue ENG-4521) to separate a project's ticket-handling instructions, a reusable ticket-preparation workflow, and the underlying connection to Jira itself into their three respective layers. Hooks are covered as automated actions that run at defined points in Claude Code's lifecycle, whether as shell commands, HTTP calls, or prompt-based checks, with a TypeScript formatter Hook as the example. Permissions are explained as the layer that is enforced by the client regardless of what Claude decides, unlike CLAUDE.md's guidance-only instructions, and the episode notes when a PreToolUse Hook is the right tool for a hard restriction. A worked example, a project called Shop, shows all five layers working together: a root CLAUDE.md with TypeScript and testing conventions, a path-scoped backend rule, a review-payment-code Skill, MCP connections to Jira and Postgres, a test-running Hook, and permissions blocking production credentials. The episode closes on why the separation is worth maintaining: CLAUDE.md content loads into every session's context whether it is needed or not, and Anthropic recommends keeping each file under roughly two hundred lines, since longer files consume more context and get followed less reliably. Useful for anyone setting up or maintaining a Claude Code project who wants instructions Claude follows, rather than one oversized file competing with itself. Sources & References How Claude remembers your project, Claude Code docs — https://code.claude.com/docs/en/memory Extend Claude with skills, Claude Code docs — https://code.claude.com/docs/en/skills Connect Claude Code to tools via MCP, Claude Code docs — https://code.claude.com/docs/en/mcp Hooks reference, Claude Code docs — https://code.claude.com/docs/en/hooks Claude Code settings, Claude Code docs — https://code.claude.com/docs/en/settings Voice narration is AI-generated.
S1 · E36
Thursday · 5 min
036 - What Is a Vertical AI Agent?
A general AI assistant can summarize a contract in seconds. Getting it to apply a law firm's own playbook, follow jurisdiction-specific rules, and work inside the systems where a case lives is a different problem, and closing that gap is what vertical AI agents are built for. This episode explains what a vertical agent is, how it differs from a general-purpose model like Gemini, Claude, or GPT, and why the specialization increasingly lives in the system wrapped around the model rather than in the model itself. The episode walks through Google's Gemini Enterprise for Legal, announced August 25, 2026, as a detailed example of that architecture: purpose-built skills for contract review, regulatory tracking, and legal research, secure integrations with tools lawyers already use including iManage, NetDocuments, and e-discovery platforms, and permission handling that carries over a firm's existing access controls instead of flattening them. It also covers Google's privacy claim that client data and playbooks are never used to train its foundation models. Three other companies illustrate the same pattern outside law. Ambience Healthcare builds its platform around reconciling medical records and generating specialty-accurate clinical documentation and coding. EvenUp positions itself as the leading AI platform for personal injury law firms, covering case work from intake through trial. Harvey describes its product as legal AI for law firms and corporate legal teams, spanning contract analysis, due diligence, compliance, and litigation. The episode explains why narrow, workflow-specific AI can create measurable business value, since automation pays off most when it's applied to an entire process rather than a single question, while noting that vertical AI products aren't automatically protected from competition, because foundation model providers are building their own industry-specific tools. Listeners will come away understanding that calling a system a vertical agent says nothing about whether its underlying model was trained specifically for that industry; the specialization usually comes from retrieval, instructions, tool connections, permissions, and evaluations layered around a general-purpose model. Useful for anyone trying to understand how AI agents are being deployed inside regulated, workflow-heavy industries like law, healthcare, and finance. Sources & References Introducing Gemini Enterprise for Legal, Google Cloud Blog — https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise-for-legal Ambience Healthcare — https://www.ambiencehealthcare.com/ EvenUp Law — https://www.evenuplaw.com/ Harvey — https://www.harvey.ai/ Voice narration is AI-generated.
S1 · E35
Thursday · 5 min
035 - Your CLAUDE.md File Is Probably Too Big
Claude Code's CLAUDE.md file is supposed to make every coding session smarter, but load it up carelessly and it does the opposite: burying the instructions that matter under everything else you've dumped in over time. This episode is a practical walkthrough of how CLAUDE.md actually works and how to keep it from getting out of hand. The episode covers where these files live in Claude Code's memory hierarchy: a personal file in your home directory, a project file shared through source control, a gitignored CLAUDE.local.md for personal preferences, and an organization-wide managed policy file IT can push to every developer's machine, and how they combine. Nested CLAUDE.md files inside subdirectories like backend or backend/payments don't override their parent. Claude Code concatenates every discovered file into context, with broader instructions loaded first and more specific ones last, so contradictory rules across files can leave Claude with competing guidance instead of a clean override. It also draws a distinction most CLAUDE.md advice skips: auto memory, the separate system where Claude Code writes its own notes based on corrections and preferences you give it as you work, versus CLAUDE.md, which only holds what you deliberately write. Mixing the two is one reason files balloon. Other practical guidance covered includes Claude Code's own recommendation to keep each CLAUDE.md under roughly two hundred lines, the .claude/rules folder for instructions scoped to specific file paths, skills for procedures that only need to load when relevant, the @ import syntax and why it doesn't actually shrink context, and the /init and /context commands for generating and auditing memory files. The episode closes on a point worth remembering on its own: CLAUDE.md is guidance the model tries to follow, not an enforced boundary, and anything that has to be blocked regardless of what Claude decides belongs in a hook or a permission setting instead. Useful for anyone running Claude Code day to day who wants their CLAUDE.md file to actually get followed instead of ignored. Sources & References How Claude remembers your project, Claude Code docs — https://code.claude.com/docs/en/memory Voice narration is AI-generated.
S1 · E34
Wednesday · 5 min
034 - Should Some Jobs Be Banned From AI?
Could some jobs actually need protection from AI? Bill Gates has proposed “Human Reserved” jobs and taxes on AI and robots to slow workforce disruption. This episode examines his ideas alongside current research from the ILO, IMF, OECD, and World Economic Forum, and asks where AI should assist humans, replace tasks, or leave certain responsibilities to people. Voice narration is AI-generated. REFERENCES Bill Gates — The choices we make about AI now are critical https://www.gatesnotes.com/a-turbulent-ai-era-and-critical-choices-to-make International Labour Organization — Generative AI and Jobs: A 2025 update https://www.ilo.org/publications/generative-ai-and-jobs-2025-update IMF — AI Will Transform the Global Economy. Let’s Make Sure It Benefits Humanity https://www.imf.org/en/blogs/articles/2024/01/14/ai-will-transform-the-global-economy-lets-make-sure-it-benefits-humanity OECD — Using AI in the workplace https://www.oecd.org/en/publications/using-ai-in-the-workplace_73d417f9-en.html World Economic Forum — Future of Jobs Report 2025 https://www.weforum.org/stories/2025/01/future-of-jobs-report-2025-jobs-of-the-future-and-the-skills-you-need/
S1 · E33
Tuesday · 14 min
033 - What Skills Should Software Engineers Learn Right Now?
Software development is changing quickly, but the useful response is not to chase every new AI product. The better question is which capabilities will make a software engineer more valuable in an AI-assisted development environment. The episode focuses on eight areas. 1. AI coding tools. Modern coding agents such as Codex can inspect repositories, plan changes, edit files, run tests, debug problems, and prepare work for review. The important skill is moving beyond autocomplete: developers need to direct, supervise, review, and verify agent-generated work. OpenAI describes Codex as capable of taking engineering work from issues through tested, review-ready code. 2. AI architecture. Engineers do not need to train foundation models to work effectively with AI. They should understand LLMs, context, retrieval and RAG, embeddings, tool calling, agent loops, memory, skills, MCP, permissions, guardrails, and tracing. The goal is being able to explain where an AI application's information comes from, what actions the model can take, and what happens when the system fails. 3. Context engineering. An AI model can perform poorly when it receives the wrong information. Coding-agent context can include the repository, instructions, architecture documentation, requirements, tests, tools, skills, and previous decisions. Martin Fowler's 2026 work on context engineering describes the growing importance of deliberately shaping this environment. Good code structure, documentation, tests, and clear conventions can therefore help both humans and agents. 4. Computer science fundamentals. AI-generated code does not eliminate the need to understand algorithms, data structures, databases, networking, operating systems, concurrency, distributed systems, APIs, security, Git, and testing. The more implementation AI can produce, the more important it becomes to understand the consequences of the implementation. 5. Verification and debugging. When code generation becomes cheaper, the ability to determine whether software is correct becomes more valuable. Engineers need testing, code review, debugging, observability, performance analysis, security testing, and regression testing. AI-generated code can look polished while still making incorrect assumptions or introducing subtle problems. 6. Architecture and systems thinking. Engineers should understand APIs, databases, queues, caches, events, services, cloud infrastructure, identity, observability, data pipelines, and security boundaries. AI systems add components such as model gateways, retrieval systems, vector stores, agent harnesses, memory, tools, and orchestration. The valuable skill is turning a vague requirement into a reliable system and understanding its failure modes. 7. AI security. 8. Requirements and specification. REFERENCES OpenAI — Codex for Software Engineering Teams https://openai.com/business/solutions/engineering/ OpenAI — Codex https://openai.com/codex/ OpenAI — Unrolling the Codex Agent Loop https://openai.com/index/unrolling-the-codex-agent-loop/ OpenAI — Harness Engineering: Leveraging Codex in an Agent-First World https://openai.com/index/harness-engineering/ OpenAI — Running Codex Safely at OpenAI https://openai.com/index/running-codex-safely/ OpenAI — How Agents Are Transforming Work https://openai.com/index/how-agents-are-transforming-work/ Martin Fowler / Thoughtworks — Context Engineering for Coding Agents https://martinfowler.com/articles/exploring-gen-ai/context-engineering-coding-agents.html Anthropic — Engineering and Coding Agent Research https://www.anthropic.com/engineering Authoring Agent Skills: A Software-Engineering Approach https://arxiv.org/abs/2607.25032 Engineering Reliable Coding Agents: Evaluating and Operating the System Around the Model https://arxiv.org/abs/2608.13867 Self-Evolving Coding Agents https://arxiv.org/abs/2608.03392 Voice narration is AI-generated.
S1 · E32
Tuesday · 4 min
032 - Can an AI Agent Trust Another AI Agent?
As AI agents begin working with other agents, a new security question appears: if one agent delegates work to another, when should it trust the result? The episode separates three concepts that are often confused. Authentication asks, "Who are you?" Authorization asks, "What are you allowed to do?" Trust asks whether the receiving agent should rely on another agent for a particular task or result. The Agent2Agent, or A2A, protocol is one of the major efforts addressing agent-to-agent communication. A2A 1.0 provides a standardized way for independent agents to discover capabilities and collaborate without exposing their internal memory or tools. An Agent Card describes an agent's identity, endpoint, capabilities, skills, and authentication requirements. An Agent Card, however, is a description, not proof of trust. A2A uses established web security mechanisms including HTTPS, OAuth, API keys, OpenID Connect, and mutual TLS. These mechanisms can establish and verify identity and support authorization, while authorization policies can limit access based on identity, requested skills, actions, scopes, and data policies. The episode then makes an important distinction between trusting an agent and trusting its output. An agent can be authenticated and authorized to perform a task while still producing an incorrect conclusion. Important results may therefore need evidence, citations, independent verification, or other checks appropriate to the risk. Delegation creates another challenge. If Agent A asks Agent B to perform work, and Agent B asks Agent C to retrieve data, the permissions given to Agent A should not automatically give every downstream agent unlimited access. This is why least-privilege authorization and controls around delegated actions are important. Provenance is another part of trustworthy multi-agent systems. When several agents contribute to a final result, the system should be able to identify where important information came from, which agent produced it, and how it moved through the workflow. This becomes especially important when agents operate across organizational boundaries. The episode also notes that agent identity is still an emerging area. Current proposals and Internet-Drafts are exploring stronger, verifiable identities and provenance for agents and delegation chains. These are developing standards, not settled industry requirements. The practical approach is to evaluate trust according to the situation. Start with identity, verify authentication, check authorization, consider the sensitivity of the data and action, and verify important results. High-impact actions should receive stronger controls than low-risk information requests. The central idea is that an AI agent should not trust another agent simply because the two can communicate. Trust is built from identity, authentication, authorization, provenance, evidence, and an understanding of the consequences if the other agent is wrong. REFERENCES A2A Protocol — A2A 1.0 Specification https://a2a-protocol.org/latest/specification/ A2A Protocol — Security https://a2a-protocol.org/latest/topics/key-concepts/ Google — A2A Protocol https://developers.googleblog.com/en/a2a-a-new-era-of-agent-interoperability/ OpenAI — Introducing the Agents SDK https://openai.com/index/new-tools-for-building-agents/ OpenAI — The next evolution of the Agents SDK https://openai.com/index/the-next-evolution-of-the-agents-sdk/ IETF — Agent Identity and Authentication Internet-Draft https://datatracker.ietf.org/doc/draft-rosenberg-agent-auth/ Voice narration is AI-generated.
S1 · E31
Tuesday · 5 min
031 - How an AI Agent Remembers You
How can an AI agent remember that you prefer short emails, know which project you're working on, or pick up a task you started yesterday? An agent can create this experience by storing useful information outside the model's immediate context and retrieving the right pieces when they become relevant. The episode explains several layers of memory. Working context contains the current conversation, instructions, tool results, and information needed for the task. Longer-term memory can include procedural information about how the agent should behave, semantic facts such as preferences or project details, and episodic information about previous events. These are useful mental models rather than a universal standard, since different systems implement memory differently. The episode then connects memory to retrieval-augmented generation, or RAG. RAG retrieves external information and places relevant material into the model's context. Memory can use the same general retrieval pattern to bring information from the past into a current task. Retrieval does not have to mean vector search: an agent can also use SQL, documents, APIs, files, or other sources. Context engineering is important because the model's active context is limited. Anthropic's current guidance emphasizes selecting useful information and using just-in-time retrieval rather than loading an entire knowledge base into every request. Memory also has to change over time. If a user once preferred morning meetings and later says afternoons work better, the system needs to update or supersede the older preference. A memory system therefore needs rules for what to keep, what to retrieve, how to handle conflicting information, and when information should be removed. There is also a privacy dimension. If an agent stores information about a person, users need controls over what is remembered, what can be retrieved, and what can be deleted. Memory is therefore part of the agent's data architecture and permissions model. The key idea is that an AI agent does not remember people the way a person does. It stores information outside the model's immediate context, retrieves useful pieces when needed, and places them into the context used for the next decision. Memory, retrieval, and context management work together to create the experience of an agent that knows you over time. REFERENCES Anthropic — Effective context engineering for AI agents https://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents Anthropic — Effective harnesses for long-running agents https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents OpenAI — Memory and new agent capabilities https://openai.com/index/memory-and-new-agent-capabilities/ OpenAI — New tools for building agents https://openai.com/index/new-tools-for-building-agents/ Voice narration is AI-generated.
S1 · E30
Tuesday · 9 min
030 - How AI Agents Work: Harnesses, Memory, Tools and Loops (Deep Dive)
How does an AI agent go from a language model that generates text to a system that can research, use software, change files, and work through a task? Episode 30 explains the core architecture behind modern AI agents, focusing on the components surrounding the model rather than benchmark scores. The episode starts with the LLM and the harness. The model provides the language and reasoning capability, while the harness supplies instructions, context, memory, tools, state, permissions, and execution behavior. OpenAI's current Agents SDK includes orchestration, tools, guardrails, and tracing, while its newer agent infrastructure supports controlled file and command execution. Anthropic's work on long-running agents similarly emphasizes context management and durable artifacts for work that spans multiple sessions. Memory is then separated from the model's immediate context. The episode explains working context alongside longer-lived procedural, semantic, and episodic information, while noting that implementations vary. This leads into retrieval-augmented generation, or RAG, and the broader idea of context engineering. RAG can use embeddings and vector search, but an agent can also retrieve information through SQL, document search, APIs, files, or web tools. The episode emphasizes just-in-time retrieval because sending too much information into a context window can make useful information harder to use. Context engineering ties these pieces together. The challenge is not simply putting more information into the model. It is selecting useful information at the right time and in a form the model can use. That includes deciding which memories to retrieve, which tool results to keep, which instructions to prioritize, and what older information can be compacted or discarded. Tracing provides visibility into the run. A trace can record model calls, retrieval, tool calls, inputs and outputs, timing, token usage, errors, and other metadata. Langfuse documents traces as structured records containing observations such as generations, retrieval operations, and tool calls, while OpenAI's Agents SDK includes tracing and observability for agent workflow execution. The final concept is LLM Ops: the ongoing engineering loop around the running system. Teams inspect traces, identify failures or inefficiencies, adjust prompts, tools, retrieval, models, memory, or guardrails, and run the system again. The episode closes with a simple architecture: LLM, harness, context and memory, retrieval, tools, loop, guardrails, tracing, and operational improvement. References OpenAI — New tools for building agentshttps://openai.com/index/new-tools-for-building-agents/ OpenAI — The next evolution of the Agents SDKhttps://openai.com/index/the-next-evolution-of-the-agents-sdk/ Anthropic — Building Effective AI Agentshttps://resources.anthropic.com/building-effective-ai-agents Anthropic — Effective context engineering for AI agentshttps://www.anthropic.com/engineering/effective-context-engineering-for-ai-agents Anthropic — Effective harnesses for long-running agentshttps://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents Anthropic — Writing effective tools for AI agentshttps://www.anthropic.com/engineering/writing-tools-for-agents Langfuse — Core Concepts: Traces and Observationshttps://langfuse.com/docs/observability/data-model Langfuse — Tracinghttps://langfuse.com/academy/tracing Voice narration is AI-generated.
S1 · E29
Monday · 4 min
029 - What Is Ox Alpha? Inside the Mystery Stealth Model
A model with no company name attached to it appeared free on OpenRouter on August 20, 2026, and an early coding test made it look like a serious challenger to Claude and GPT. This episode covers Ox Alpha, the anonymous stealth model developers spent the week testing: what OpenRouter and independent researchers have confirmed about it, why a larger benchmark run changed the headline number, and what an anonymous provider means for anyone thinking about sending it sensitive work. Ox Alpha showed up on OpenRouter under the placeholder ID stealth/ox-alpha, also reachable through coding tools like OpenCode and Cline, with OpenRouter naming the provider only as an anonymous third party. It ships with a context window of 1,048,576 tokens, a maximum output of 131,072 tokens, and support for text, image, and video input, built for coding, sustained agentic work, and production use with tool calling and structured output. The provider hasn't disclosed a parameter count for the underlying model. The coding benchmark is where the episode gets specific. Developer Ben Davis first ran Ox Alpha through a 10-task DeepSWE coding test and reported an 80% pass rate, ahead of roughly 65% for Claude and 52% for GPT-5.6 in that same informal comparison. A larger follow-up test brought the score down to about 63% — a small enough sample swing to make clear that a ten-task result isn't a settled ranking, and that early numbers from an unverified model deserve a second test before they get repeated as fact. Identity is the other open question. Researchers comparing Ox Alpha's tokenizer behavior, video handling, and response fingerprints have pointed toward Zhipu AI's GLM family as a possible origin, but no company has confirmed building it, and the provider has stayed anonymous through the entire preview period. The episode also covers the privacy angle: OpenRouter states that prompts and completions sent to Ox Alpha are retained by the provider, even though they aren't used for training, which is worth factoring in before sending confidential or proprietary material to a model nobody has claimed. Sources & References OX Alpha: The Anonymous Frontier Model - Comprehensive Technical Analysis, Local AI Zone — https://local-ai-zone.github.io/blog/ox-alpha-stealth-model-comprehensive-analysis.html Ox Alpha likely GLM-5.3: the mystery model's fingerprint, OrcaRouter — https://www.orcarouter.ai/blog/ox-alpha-stealth-model-what-we-know Ox Alpha: Anonymous 1M-Context Model Hits No. 2 on OpenCode in Three Days, The Cherry Creek News — https://thecherrycreeknews.com/ox-alpha-stealth-model-openrouter-benchmarks-analysis-cherry_creek/ Ox Alpha - API Pricing & Providers, OpenRouter — https://openrouter.ai/stealth/ox-alpha Ox Alpha on OpenRouter: Free 1M Stealth Model (Aug 2026), explainx.ai — https://explainx.ai/blog/openrouter-ox-alpha-stealth-model-august-2026 Voice narration is AI-generated.
S1 · E28
August 21 · 4 min
028 - Claude Academy Is Now Available - So What Is It?
Anthropic launched Claude Academy on March 2, 2026, a free, self-paced learning platform at academy.claude.com. This episode covers what's actually in it, how the courses are organized, and a closer look at the "AI Fluency" framework Anthropic puts at the center of the whole thing, including a worked walkthrough of its four core competencies applied to a real task. Claude Academy started with 13 courses and has grown to around 20, ranging from introductory material for people who've never used Claude to technical deep dives on the Claude API and Claude Code, plus dedicated pathways for K-12 educators and nonprofits. The catalog is organized by how you'd actually use Claude rather than by difficulty: tracks cover Claude.ai, Claude Cowork, the developer-facing Claude Code and Claude Platform tools, and Claude Tag inside Slack. On the technical end, "Building with the Claude API" runs 84 lectures across 8 hours with 10 quizzes, with matching courses for running Claude through Amazon Bedrock and Google Vertex AI. There's also a dedicated course on the Model Context Protocol and one on model capabilities and limitations. Every course awards a certificate on completion, and none of it requires a credit card. The course Anthropic points newcomers to first isn't technical at all. "AI Fluency: Framework and Foundations" teaches what Anthropic calls the four D's: Delegation, deciding whether, when, and how to engage AI on a task; Description, communicating a goal clearly enough to get a useful result; Discernment, accurately judging whether an AI's output is actually good; and Diligence, taking responsibility for what gets done with AI and how. The episode walks through all four applied to drafting a difficult client email, and touches on the framework's three modes of human-AI interaction: automation, augmentation, and agency. The larger point is that the four D's aren't specific to Claude or to any one platform. They're a vocabulary for a kind of judgment that doesn't show up in a changelog, and having names for each step gives a person, or a team, a way to talk about exactly where a task with AI went right or wrong. Sources & References Claude Academy — https://academy.claude.com/ Claude course catalog — https://claude.com/resources/courses Claude tutorials — https://claude.com/resources/tutorials Anthropic, AI Fluency Framework and Foundations PDF — https://www-cdn.anthropic.com/b383cf6baddbfc72fdf8b0ed533a518e2872d531.pdf Cryptobriefing, Anthropic launches Claude Academy to enhance AI learning resources — https://cryptobriefing.com/anthropic-launches-claude-academy/ Voice narration is AI-generated.
S1 · E27
August 21 · 5 min
027 - The AI Stack Developers Actually Use in 2026
Ask a developer to describe their AI stack and most people picture one model wired into one clever autonomous agent. Production data shows something messier. This episode maps what the AI development stack actually looks like in 2026 — models, databases, languages, and the connective tissue holding it together — using current developer telemetry from Datadog and a joint 2026 industry survey instead of marketing claims. On models, Datadog's analysis of AI usage across more than a thousand customers found OpenAI used by 75% of organizations a year ago, down to 63% now, while Anthropic's Claude gained 23 percentage points and Google's Gemini gained 20 over the same stretch. More than 70% of organizations now run three or more models side by side, picking each one for quality, speed, cost, or the task at hand rather than staying locked into a single provider. Frameworks like LangChain sit between those models and the rest of an application, but adoption is still far from universal: Datadog found agent-framework usage at only about 18% of organizations in early 2026. Production telemetry also shows most of what gets labeled an "agent" is doing far less than the word implies: 59% of agentic applications made just a single service call, and only 18% chained three or more together. On the database layer, Datadog found about 60% of organizations run at least one production Postgres instance, and its vector-search extension, pgvector, was the fastest-growing non-bundled Postgres extension in their data, up 24% between December 2025 and May 2026 — but still early, with fewer than 6% of organizations and under 1% of tables actually using a vector index, leaving dedicated vector databases like Pinecone, Weaviate, and Qdrant to handle retrieval at real scale. On languages, Python remains the leading language ecosystem for accessing Postgres in production, up 33% over two years, while Node and TypeScript stay important for web applications, APIs, and streaming interfaces — there's no single required split, and most teams use each language where it naturally fits. The episode closes on the economics, drawn from a joint 2026 AI engineering survey by Notion, Amplify Partners, and Vercel: about three-quarters of respondents said cost regularly or sometimes changes how they use AI, inference and model serving is the most commonly purchased layer rather than built in-house, and only 28% of eligible calls actually use prompt caching despite the cost savings on offer. The same survey found 97% of respondents reporting positive effects from AI, alongside 93% reporting at least one negative effect — skill erosion the most common at 29%, followed by higher review burden at 28% and brittle, incident-prone code at 22%. Sources & References Datadog, State of AI Engineering — https://www.datadoghq.com/state-of-ai-engineering/ Datadog, State of Postgres — https://www.datadoghq.com/state-of-postgres/ 2026 AI Engineering Survey, Notion, Amplify Partners, and Vercel — https://www.notion.com/lp/ai-engineering-survey Amplify Partners, The 2026 AI Engineering Report — https://www.amplifypartners.com/blog-posts/the-2026-ai-engineering-report Voice narration is AI-generated.
S1 · E26
August 20 · 4 min
026 - What Is Prompt Injection? The Copilot CoSnitch Hack
One click on a link, and Microsoft's own Copilot could read a user's email, Google Drive, and calendar, then quietly send it to a stranger. This episode explains prompt injection, the security flaw unique to AI agents, through a real Microsoft Copilot vulnerability that was patched after sitting open for months. A model reading text can't reliably tell the difference between an instruction from its actual user and an instruction hidden inside content it's processing, a webpage, an email, a shared document. If that content contains hidden text telling the model to ignore its real instructions and leak data instead, a model with no built-in sense of who's allowed to give it orders can follow that instead. Researcher Simon Willison named the exact combination that makes this dangerous the "lethal trifecta": an agent with access to private data, exposure to untrusted content, and the ability to communicate externally. Microsoft's Copilot had all three. Security researchers at Varonis Threat Labs found a flaw they named CoSnitch: a link combining two URL parameters, one that pre-filled Copilot's chat box with a hidden prompt and one, called autorun, that executed it automatically the moment the page loaded, with no typing required and no way to stop it by closing the tab. Once running, the injected prompt used Copilot's own built-in web-fetching tool to pull email subject lines and bodies, Google Drive filenames, calendar details, chat history, and Copilot's own saved memory, then send it out encoded to dodge detection filters. A companion flaw let attackers plant instructions directly into that saved memory, persisting through a password change, a revoked session, or even a new device. Varonis reported the issue to Microsoft on December 31, 2025. Microsoft shipped a partial fix that February and didn't close the vulnerability completely until August 18, 2026, about seven and a half months later, on a bug Microsoft itself rated critical. The episode explains why that timeline matters: prompt injection doesn't patch the way an ordinary bug does, because the underlying problem, a model's inability to fully separate its user's instructions from anyone else's, isn't unique to Copilot. Any AI assistant wired into private data, exposed to the open web, and able to send data back out carries the same three ingredients, regardless of which company built it. Sources & References The Hacker News, Microsoft Copilot Personal Flaws Could Let One Click Exfiltrate Data From Connected Apps — https://thehackernews.com/2026/08/microsoft-copilot-personal-flaws-could.html Computerworld, Microsoft finally patches critical one-click Copilot vulnerability, almost eight months after learning of it — https://www.computerworld.com/article/4211325/microsoft-finally-patches-critical-one-click-copilot-vulnerability-more-than-eight-months-after-learning-of-it.html Simon Willison, The lethal trifecta for AI agents: private data, untrusted content, and external communication — https://simonwillison.net/2025/Jun/16/the-lethal-trifecta/ Voice narration is AI-generated.
S1 · E25
August 20 · 4 min
025 - What Is a Context Window?
Paste a 200-page PDF into a chatbot and it answers questions about page 180 within seconds. That feels instant and free. Neither is quite true, and this episode explains what a context window actually is, what happens inside a model when it processes a huge document, and why a bigger window always comes with a cost. A context window is the total amount of text, measured in tokens, a model can hold at once across the prompt, any pasted document, and its own reply. Flagship models from OpenAI, Google, and Anthropic have all converged around a roughly one-million-token window in 2026, up from the 4,000- and 8,000-token limits those same companies shipped just a few years earlier. Inside that window, a transformer compares every token against every other token to decide what to attend to, so doubling the amount of text roughly quadruples the compute required, which is why OpenAI's own pricing sheet charges more once a request crosses into what it calls long-context territory. Size brings a second cost that has nothing to do with money: research on "lost in the middle" behavior shows models are measurably worse at using information buried in the center of a long context than information at the start or end, and newer testing on today's huge-context models found that gap gets worse, not better, as windows grow, with 10 to 25 percent accuracy loss on facts placed mid-document. The episode works through the actual math on a 200-page PDF: at roughly 500 words a page and OpenAI's own conversion rate of about 0.75 words per token, that document comes out to roughly 130,000 tokens and around 33 cents in input costs before the model writes a single word back. It also covers a concrete real-world response to that tradeoff: OpenAI's own coding tool, Codex, deliberately caps how much of GPT-5.6's advertised 1.05-million-token window it will actually use at once, capping it at 272,000 tokens specifically because of mid-document accuracy loss and runaway cost. The practical takeaway is simple: put what matters most at the start or end of a long prompt, not buried in the middle. Sources & References OpenAI, Key concepts (token-to-word conversion) — https://developers.openai.com/api/docs/concepts OpenAI, API pricing — https://developers.openai.com/api/docs/pricing Codex Knowledge Base, The Context Window Gap: Why Codex CLI Caps GPT-5.6's Million-Token Window at 272K — https://codex.danielvaughan.com/2026/07/20/context-window-gap-codex-cli-gpt56-advertised-vs-effective-budget-compaction-strategy/ hidekazu-konishi.com, LLM Context Window Growth Timeline — https://hidekazu-konishi.com/entry/llm_context_window_growth_timeline.html Voice narration is AI-generated.
S1 · E24
August 20 · 4 min
024 - What Is Open-Weight AI?
Meta and Mistral give away state-of-the-art AI models for free, no purchase or subscription required. That sounds like open source software, but it isn't quite the same thing, and the difference has real consequences for anyone building on top of these models. This episode explains what "open weight" actually means and how it differs from open source AI. A model's weights are the millions of parameters it learned during training; publishing them lets anyone download and run the finished model. The Open Source Initiative, the group that defines the term for software, says that alone doesn't qualify as open source: a model only earns that label if it also comes with the training code and real details about the training data, the sources that shaped it, not just the finished result. The distinction gets concrete in the actual license text. Meta's Llama 4 Community License Agreement requires any company with more than 700 million monthly active users in the preceding month to request a separate license from Meta, and it blocks using Llama's own outputs to train a competing model, restrictions a genuine open source license like Apache or the ones covering Linux would never include. Mistral, the French AI company, ships most of its core models under Apache 2.0, a real open source license with no usage cap, while other open-weight releases, like Google's Gemma, carry their own restrictions, including the right to remotely cut off access if usage rules are broken. The episode also covers why these companies give away models that cost hundreds of millions of dollars to train. For Meta, a free model functions as a large-scale recruiting tool, builds a developer ecosystem, and pressures closed competitors like OpenAI and Anthropic on price, all while Meta's core revenue stays tied to advertising rather than model access. Mistral co-founder Guillaume Lample frames the company's reasoning around independence: "We don't want AI to be controlled by only a couple of big labs," a statement that also reads as a pitch for European AI sovereignty against dominant American labs. Sources & References Open Source Initiative, Open Weights: not quite what you've been told — https://opensource.org/ai/open-weights Meta, Llama 4 Community License Agreement — https://raw.githubusercontent.com/meta-llama/llama-models/main/models/llama4/LICENSE TechCrunch, Mistral closes in on big AI rivals with Mistral 3, open-weight frontier and small models — https://techcrunch.com/2025/12/02/mistral-closes-in-on-big-ai-rivals-with-mistral-3-open-weight-frontier-and-small-models Voice narration is AI-generated.
S11 · E23
August 20 · 4 min
023 - What Is AI Model Routing?
Why does ChatGPT answer some questions instantly and pause to "think" on others? Most people assume they're always talking to the same AI brain. In reality, a router decides in a fraction of a second which model actually handles your message, and that decision shapes both the quality of the answer and what it costs to produce. This episode explains model routing, the invisible layer sitting in front of most modern AI products. OpenAI built a router directly into GPT-5 when it launched in 2025, and it's still the default behavior in ChatGPT: the system looks at your conversation type, how complex the request appears, whether tools are needed, and explicit signals like typing "think hard about this," then sends easy requests to a fast, efficient model and harder ones to a deeper reasoning model. The router keeps improving over time based on which answers users actually preferred and how often each model got things right. Routing isn't limited to one company's own lineup. OpenRouter, a platform that sits in front of roughly 400 models from OpenAI, Anthropic, Google, and dozens of other providers, runs its own "auto" router that classifies a prompt by task type and picks a model based on real usage patterns from its community over the trailing week. OpenRouter said it was serving about 8 million users by this spring, and on August 19, 2026, days before this episode, Stripe confirmed it's buying the company in a deal reported at more than $7 billion, a bet that routing between AI models is becoming as central to the internet as routing payments already is. The episode works through a real cost comparison to show what's actually at stake in a routing decision: on OpenAI's current pricing, the efficient GPT-5.4 nano model runs about twenty cents per million input tokens, while the full GPT-5.4 model runs two dollars and fifty cents for the same volume, more than twelve times as much. That gap is the economic reason routing exists, and it explains why most routed products still let a user override the router's guess when the stakes are high enough to want a specific model by hand. Sources & References OpenAI, Introducing GPT-5 — https://openai.com/index/introducing-gpt-5/ OpenRouter, How Model Routing Works — https://openrouter.ai/blog/insights/model-routing/ Morph, OpenAI API Pricing — https://www.morphllm.com/openai-api-pricing TechCrunch, Stripe will reportedly acquire AI gateway startup OpenRouter for $7B+ — https://techcrunch.com/2026/08/16/stripe-will-reportedly-acquire-ai-gateway-startup-openrouter-for-7b/ Axios, Stripe confirms OpenRouter acquisition — https://www.axios.com/pro/fintech-deals/2026/08/19/stripe-openrouter-acquisition Voice narration is AI-generated.
S1 · E22
August 19 · 5 min
022 - Loop Engineering vs. Graph Engineering
AI agent systems are moving beyond a single model repeatedly calling tools. Two useful concepts for understanding that evolution are loop engineering and graph engineering. Loop engineering focuses on the behavior of an individual agent over time. An agent receives context, chooses an action or tool call, observes the result, and decides what to do next. The engineering challenge is making that cycle reliable: defining progress, providing useful feedback, handling failures, managing context, setting budgets, and deciding when the agent should stop. OpenAI's Codex documentation describes the agent loop as the core logic that orchestrates the user, model, and tools. OpenAI also notes that a single turn can contain many model-and-tool iterations and that context-window management becomes an important responsibility. Graph engineering addresses a different problem: coordinating a larger system of steps, agents, tools, branches, dependencies, and shared state. Instead of leaving every next action to one agent, the system can represent the workflow explicitly as a graph. Nodes can represent tasks or specialized agents, while edges describe dependencies or transitions. For example, a research system might have separate research, data-analysis, writing, and review nodes. Research and data analysis could run in parallel, their results could feed a writing node, and the draft could move to a review node. If the reviewer finds a problem, the graph can route the work back for revision. The important distinction is that loop engineering is primarily about iteration and behavior, while graph engineering is about structure and coordination. They are complementary. A node inside a graph can contain its own agent loop. The graph determines when that node runs, what it receives, and where its output goes; the loop determines how the agent works through its assigned task. The term "graph engineering" is still emerging rather than a universally standardized discipline. Many underlying techniques have existed for years in workflow orchestration, state machines, directed graphs, DAGs, and multi-agent systems. What is new is applying the graph-engineering framing directly to increasingly complex AI agent systems. A paper published August 21, 2026, "Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence," explicitly proposes graph engineering as an emerging paradigm for organizing complex agent systems. The paper argues that some tasks require heterogeneous expertise, parallel execution, independent verification, and persistent state that can exceed the organizational capacity of a single agent. REFERENCES OpenAI — Unrolling the Codex agent loop https://openai.com/index/unrolling-the-codex-agent-loop/ OpenAI — The next evolution of the Agents SDK https://openai.com/index/the-next-evolution-of-the-agents-sdk/ OpenAI — Harness engineering: leveraging Codex in an agent-first world https://openai.com/index/harness-engineering/ OpenAI — An open-source spec for Codex orchestration: Symphony https://openai.com/index/open-source-codex-orchestration-symphony/ Anthropic — Harness design for long-running application development https://www.anthropic.com/engineering/harness-design-long-running-apps Feng et al. — Graph Engineering in the Era of LLM Agents: From Individual Intelligence to System Intelligence https://arxiv.org/abs/2608.21156 Macedo — What makes prompts a graph: necessary and sufficient conditions for prompt graph engineering https://arxiv.org/abs/2607.27578 Voice narration is AI-generated.
S1 · E21
August 18 · 9 min
021 - Is Computer Science Dead? What the Enrollment and Job Data Actually Say (Deep Dive)
For about fifteen years, computer science was one of the safest majors to recommend. The 2026 data checks that advice: enrollment, unemployment, hiring trends, and a real curriculum redesign, for parents, students, current CS majors, and professors alike. The National Student Clearinghouse reported computer and information sciences enrollment fell 8.4% at four-year schools and 11.2% at two-year colleges in spring 2026, with graduate enrollment down 14%, even as total enrollment grew nationally. A Computing Research Association survey of 134 academic leaders found 62% of programs saw undergraduate enrollment drop, typically by 11% to 15%. Computer science now has the lowest persistence rate of the ten most popular bachelor's degrees, at 85%, versus 93% for engineering. That same CRA survey ties the decline to fear that AI will eliminate software engineering jobs, and two thirds of computing departments say recent grads are struggling to find work. Federal Reserve Bank of New York data, drawn from Census figures on graduates in their early twenties, puts computer science unemployment at 7% and computer engineering at 7.8%, fifth and second highest of 73 majors tracked. Entry level hiring has tightened: the junior share of tech employment fell from 15% to 7% in three years, and Microsoft and Google both say roughly 30% of their code is now AI generated, with Meta targeting 50%. But the data also cuts the other way. Computer science graduates who do find work are underemployed at under 20%, versus roughly 42% for graduates overall, and the major still ranks third among the most in demand bachelor's degrees. IBM says it's tripling entry level technical hiring. Some of the pullback likely traces to pandemic era over hiring correcting itself rather than AI, and the enrollment drop looks more like a shift toward specialization: cybersecurity, data science, computer engineering, and standalone AI programs are all growing inside the same departments. The University of Iowa redesigned its computer science curriculum for fall 2026 around 2023 to 2024 Association for Computing Machinery guidelines, adding required ethics and intro to AI courses and combining separate statistics, probability, and linear algebra classes into one applied math course. That redesign lines up with what hiring managers say they actually want: shipped, deployed projects over GPA, real maintenance experience, and fluency working alongside AI tools instead of trusting or ignoring them. Direct guidance follows for each audience, plus cybersecurity, data science, computer engineering, and physical infrastructure fields as the alternatives current data points toward. Sources & References National Student Clearinghouse, computer science enrollment data: https://www.studentclearinghouse.org/nscblog/computer-science-enrollment-is-cooling/ CRA CERP Pulse Survey: https://cra.org/crn/2025/10/cerp-pulse-survey-a-snapshot-of-2025-undergraduate-computing-enrollment-patterns/ NY Fed, College Labor Market data: https://www.newyorkfed.org/research/college-labor-market VnExpress, U.S. computer science graduates unemployment rate: https://e.vnexpress.net/news/tech/tech-news/us-s-computer-science-grads-face-5th-highest-unemployment-rate-5048578.html Built In, computer science degree decline and AI: https://builtin.com/articles/computer-science-degree-decline-ai Extern, computer science job market guide 2026: https://www.extern.com/post/computer-science-job-market-2026-guide Boise State University, AI and computer science demand: https://www.boisestate.edu/coen-cs/2026/03/10/more-ai-means-more-computer-science/ Ardura Consulting, junior developer hiring crisis: https://ardura.consulting/blog/junior-developer-crisis-2026-why-companies-stopped-hiring-entry-level/ U. of Iowa, new CS curriculum for fall 2026: https://cs.uiowa.edu/news/2026/06/new-computer-science-curriculum-fall-2026 Voice narration is AI-generated.
S1 · E20
August 18 · 14 min
020 - A Real Example of Claude Running a Small Business: What It Actually Costs (Deep Dive))
Can a single tech-savvy person actually build and maintain an AI agent for a small business, or is that just a demo-video fantasy? A full, concrete walkthrough answers that: a fictional flower shop called Rosewood Flowers automates its same-day delivery scheduling using Claude, built from an agent loop, a custom skill, and four real-world connectors, then stress-tested against the costs, maintenance, and failure modes that never make it into a product demo. In plain terms: the difference between a tool (a single action Claude can take), a connector built on the Model Context Protocol, or MCP (how Claude reaches external systems like Shopify, Google Calendar, and a mapping service), a skill (a packaged, progressively-loaded set of business rules Claude only pulls into context when relevant, built using Claude's Agent Skills framework and SKILL.md files), and an agent loop using subagents from Claude's Agent SDK, where a main agent delegates focused tasks, intake, routing, and customer notification, to specialized subagents with their own tools and instructions. Listeners get a full order walked end to end through the system, plus a harder edge case that deliberately escalates to a human instead of guessing. It also gets specific about where this kind of system actually runs: building it requires Claude API access through Anthropic's developer console, a separate product and separate billing from a Claude.ai subscription like Pro or Max, and the skill and orchestration code live as ordinary files, a SKILL.md file and Python or TypeScript scripts, in a code repository like GitHub. Two hosting paths are covered: self-hosting the code on a small server or container, or using Managed Agents, the hosted API Anthropic opened to developers in a 2026 public beta that runs the agent and sandbox directly with no server to maintain. The math on ongoing token costs uses Claude's actual current Sonnet pricing, two dollars per million input tokens and ten dollars per million output tokens, and flags that this is an introductory rate expiring September 1, 2026, a concrete example of how AI infrastructure costs shift over time. It also covers realistic build time (two to four weeks part-time), the skills a builder actually needs, and the maintenance nobody advertises: expiring OAuth tokens, connector APIs changing, model version updates requiring re-testing, and business rule changes that require editing the skill file before the next busy weekend arrives. This is a technical explainer built for listeners who want to understand agents, skills, connectors, and hosting through one fully worked, honest example rather than marketing language. Sources & References Claude Agent Skills overview (SKILL.md, progressive disclosure) — https://platform.claude.com/docs/en/agents-and-tools/agent-skills/overview Claude MCP connector documentation — https://platform.claude.com/docs/en/agents-and-tools/mcp-connector Claude Agent SDK subagents documentation — https://code.claude.com/docs/en/agent-sdk/subagents Hosting the Agent SDK (self-hosted deployment options) — https://code.claude.com/docs/en/agent-sdk/hosting Claude.ai subscription vs. API/Console access — https://support.claude.com/en/articles/9876003-i-have-a-paid-claude-subscription-pro-max-team-or-enterprise-plans-why-do-i-have-to-pay-separately-to-use-the-claude-api-and-console Anthropic Managed Agents public beta overview — https://www.roborhythms.com/anthropic-managed-agents-2026/ Claude API pricing overview, 2026 — https://www.cloudzero.com/blog/claude-pricing/ Voice narration is AI-generated.
S1 · E19
August 17 · 6 min
019 - Is Your Job Safe From AI? What the Data Says About Layoffs
Is your job safe from AI, or does the data already point the other way, and can you even trust the layoff headlines that say AI did it? Three separate research efforts point to which specific job titles are most exposed to AI right now, and the layoff headlines blaming AI get checked against whether that explanation actually holds up. Microsoft's research team analyzed two hundred thousand real conversations from Bing Copilot in 2024, mapping them against government job classifications to score how much each occupation's daily tasks overlap with what generative AI already does well. The World Economic Forum's Future of Jobs Report named accountants and auditors, alongside cashiers and administrative assistants, among the fastest-declining roles by 2030, while Microsoft's data pointed to customer service representatives and sales representatives as similarly exposed. On the other end, roofers and water treatment plant operators ranked among the least AI-exposed jobs studied, since physical, hands-on work still resists automation. Outplacement firm Challenger, Gray and Christmas tracked more than ten thousand three hundred U.S. layoffs through July 2025 explicitly attributed to AI, with tech sector cuts up thirty-six percent year over year. But the layoff data cuts both ways. Some companies genuinely are redirecting payroll into AI infrastructure: Meta cut roughly eight thousand roles, about ten percent of its workforce, in May 2026, with Mark Zuckerberg tying the move directly to rising AI costs, part of a broader trend as Alphabet, Amazon, Meta, and Microsoft plan to spend roughly seven hundred fifty billion dollars on AI infrastructure in 2026. Other times, AI functions as a convenient label. Investor Marc Andreessen has argued many large companies are overstaffed from pandemic-era hiring by twenty five to seventy five percent, calling AI the easiest excuse for cuts that were coming regardless. Salesforce CEO Marc Benioff has said layoffs from several distinct causes get lumped together under the same AI headline. A Duke University and Federal Reserve survey of company financial officers found AI's real employment impact closer to zero point four percent of jobs nationwide, a much narrower slice than the headlines imply. The full picture includes which jobs the underlying task data actually points to, plus a more skeptical framework for reading the next AI layoff headline. Sources & References Microsoft study on jobs most and least exposed to generative AI — https://www.geekwire.com/2025/microsoft-study-shows-jobs-most-and-least-impacted-by-generative-ai/ World Economic Forum, Future of Jobs Report 2025 — https://www.weforum.org/stories/2025/01/future-of-jobs-report-2025-the-fastest-growing-and-declining-jobs/ Challenger, Gray & Christmas 2025 layoff data — https://fortune.com/2025/08/07/summer-of-ai-layoffs-july-140-percent-spike-challenger-gray-christmas/ AI-washing of layoffs, Andreessen and Benioff commentary, Duke/Federal Reserve CFO survey — https://sfstandard.com/2026/04/02/ai-washing-layoffs/ Companies cutting jobs to fund AI investment — https://247wallst.com/investing/2026/05/26/companies-are-cutting-jobs-for-ai-but-not-the-billions-paid-to-investors/ Voice narration is AI-generated.

Showing 1–20 of 30 episodes