AI regulation, AI safety, and the future of artificial intelligence are at the center of this episode of Will AI Snap That. Thanos and Michael Rollins discuss AI regulation, AI agents, existential AI risk, government control, and the real-world impact of rapidly advancing AI technology.

They also explore the Hugging Face AI agent hack, AI-powered software development, mutation testing, automated testing, CI/CD, and how AI is changing the way developers build and release software.

#AI #ArtificialIntelligence #AIRegulation #AISafety #AIAgents

AI is becoming more autonomous, more powerful, and harder for humans to understand. But how much control do we really have over artificial intelligence when AI agents can access computers, passwords, Gmail, financial systems, and even build software on their own?

In this episode of Will AI Snap That, Thanos and Michael explore the biggest questions surrounding AI safety, autonomous AI, AI risks, and the future of artificial intelligence. They discuss OpenAI’s controversial Navier-Stokes breakthrough, allegations surrounding AI research, massive AI compute and token usage, and the ethics of racing to publish scientific discoveries.

The conversation also explores Anthropic’s AI safety concerns, catastrophic AI risk, hidden AI reasoning, and why humans may struggle to understand advanced AI systems. They dive into AI agent orchestration, autonomous software development, AI automation, and whether AI could replace engineers.

Finally, Thanos and Michael examine AI security risks, including computer access, Gmail, passwords, banking systems, MCP, and fine-grained permissions.

If you’re interested in artificial intelligence, AI agents, AI safety, AI security, autonomous AI, and the future of AI, this episode is for you.

AI agents are getting more autonomous, AI models are finding new ways to collaborate, and the line between useful automation and unexpected behavior is becoming harder to define. In this episode of Will AI Snap That?, Thanos and Michael Rollins dive into Claude Code token changes, MCP vs Agent SDK efficiency, AI breakout experiments, autonomous robots, world models, and whether the AI apocalypse is actually something we should worry about.From rising AI token costs and self-hosting Loom alternatives to models hacking systems, collaborating across agents, and kung fu robots, this episode explores the real-world risks and opportunities of rapidly advancing artificial intelligence. The hosts also discuss OpenAI slowing down, Google Gemini, Fei-Fei Li’s world models, robotics, AI security, autonomous weapons, and how AI agents could influence the physical and digital world. The episode closes with a big question: are AI systems really going to take over, or are humans still the biggest weakness in the loop?

AI is changing the way we build, test, and ship software, and in this episode, Michael Rollins and Faunus dive deep into what that actually looks like in practice.

From completing a massive codebase refactor in just two days to using Claude for architecture visualization, goal loops, sub-agents, and automated correctness checks, this episode explores how AI is shifting engineers from writing code to making higher-level decisions.

Is MCP Dead? Stateless MCP, Rogue Agents, and “Dangerous” Models (OpenAI vs GLM 5.3)Michael Rollins and his co-host Thanos discuss Michael’s ongoing “war” with MCP, arguing it’s largely a reimplementation of SDKs and often unnecessary when a harnessed LLM can write direct API integrations with better permission control; they critique “stateless MCP” features like pinned sessions, authentication, async connections, and “conversational stops,” likening it to SOAP with JSON. They compare internal ticketing workflows using Slack, GitHub issues, and an agent-managed database, and debate agent permissions, VM key isolation, and guardrails like auto/YOLO modes. Michael recounts a Sonnet agent going rogue by adding an SQLite fallback alongside Postgres, wasting time and tokens. They also cover OpenAI slowing model development due to “dangerous” new capabilities, reference a novel Hugging Face pipeline hack and models leaving hidden notes, and discuss GLM 5.3 (Z.ai) delaying open weights pending governance, plus the collective-action problem around releasing powerful models.

AI is getting weird fast.
This week we break down the Fable 5 chaos, the GPT 5.6 timing, the rise of vibe coders, and why software development is starting to look very different. If you work in AI, tech, or content, this one’s worth your time.

In this episode of AI Snap Thot, Michael Rollins and Thanos dig into one of the wildest AI stories of the week: an Instagram prompt-injection exploit that reportedly let attackers take over accounts by abusing old API paths. From there, the conversation expands into agent security, why tool access and authorization matter more than ever, and why “just letting the model do it” is still a dangerous idea. They also break down the debate around vibe coding, what “scaling” actually means in real engineering, and whether new models like Fable 5 really change the game for autonomous software development. If you care about AI agents, software architecture, security, and the future of coding, this one’s for you.

It's been a true minute, my friends. But we're back! Today, Thanos and I catch up about the last few months. About what we've been doing, and how much fun it's been. It's so good to see the glistening, bald pate of my friend! Don't snap me, Thanos.

"Find what your treatment plan might be missing. In minutes, not weeks. Built by a survivor."Ari Akerstein has embarked on a mission to help connect people in cancer treatment. They help patients find the best treatment options, as well as connect to a community. It is inspiring work, and incredible to see.OpenCancer.ai gives superpowers to every cancer patient. Clinically trained AI tools: second opinions, records translation, clinical trial matching, expert review - all grounded in the latest guidelines and their own medical records. Co-founded by two cancer survivors.For partners, we deliver qualified, high-intent patients at the moment of clinical decision. We don't sell data. We route intent to the most relevant services at the moment it matters most.

00:00 Cold open: “Humans of Earth” intro 00:38 Meet the guest: Colin Masters (Epic Scale AI) 02:40 Colin’s origin story (military → construction → NYSE → dev) 04:00 The EPIC method: Explore–Plan–Implement–Check 05:40 Context isn’t a dump: vector DB + forced session search 06:20 Plans = branches (GitHub issues integration) 07:10 Memory bank: standardize decisions (deposits vs withdrawals) 08:40 Feedback to feature: free tier shipped in 3 hours 10:00 Shared vs project context; 30 MCP tools 11:10 Using AI since 2024; choose simple over complex 12:20 Explore phase prompts → define a shippable MVP 13:10 AI-first ops: branches, artifacts, and shared context 14:20 Context bloat and model limits 15:05 The Claude 4.5 rant (pain points) 16:10 Buy vs build: why consider Epic Scale 17:10 Why “last 20 commits” isn’t enough 18:00 Semantic tasks and progress tracking 19:10 Watching cross-branch changes 20:00 Agent loops—with human review 21:00 Human-in-command: no auto-approve 22:10 Quote: “Developers in control of their tools will succeed.” 23:40 Houses vs software: constraints and collaboration 29:10 Where Epic Scale shines (small teams 10) 32:00 Status without Big‑A Agile: stakeholder visibility 34:10 Failing as a learning loop 38:10 Mindset shift: embrace AI or get left behind 39:20 Where AI beats humans (YAML, cross-file reasoning) 41:10 Juniors are crushing it; Go‑to‑Market Engineer 44:00 Speed breaks sales/support; training new grads 47:00 Where to find Epic Scale + free tier note 47:40 OutroAI dev isn’t “dump more into the context window.” Colin Masters from Epic Scale AI breaks down how the EPIC method (Explore–Plan–Implement–Check), an MCP server, and a shared vector memory turn chaos into shipping velocity—especially for small, AI‑first teams.We get practical about plans-as-branches with GitHub issues, memory banks to stop re‑deciding decisions, why auto‑approve wrecks projects, and how a user comment led to a free tier in three hours. We also debate “houses vs software,” the rise of the Go‑to‑Market Engineer, and what juniors are getting right with AI.Recorded: October 14, 2025What you’ll learn:How to replace “tickets” with plans-as-branches and keep context cleanWhy vector DB + retrieval beats giant context windowsGuardrails for agent loops (human-in-command, no auto‑approve)Where AI truly outperforms humans (cross-file reasoning, YAML, refactors)Team shape for AI‑first orgs (small, sharp, 10)How to create stakeholder visibility without Big‑A Agile ceremonyGuestColin Masters — Epic Scale AI: https://epicscale.ai/Chapters See chapter markers above.Subscribe for more on AI engineering, MCP, and practical agent workflows.Hashtags: #AIEngineering #MCP #DevTools #AIFirst #VectorDB #VSCode #GitHub