Browse Comments — Relevant (AI ∩ value)

Ruben Hassid This 7-day checklist shifts from passive learning to active integration. Day 3 is key: creating .md files with voice samples and banned words externalizes negative constraints, an advanced pruning technique. Without it, the model can't distinguish noise from signal. Connectors and scheduled tasks transform Claude from chatbot to executor. Week one builds it, week two maintains it. Actionable system. Next step: weekly audit of which scheduled tasks still add value.

Nice anatomy. But bodies have one thing this model skips: an immune system. Who decides what the agent is NOT allowed to do? Who stops RAG from surfacing the confidential board deck to the intern's chatbot? The four layers are the easy part. Layer five — governance — is where most implementations quietly die.

Good breakdown. One pushback on layer 1: calling the LLM a “brain”makes it sound like it reasons. It predicts tokens. That distinction isn’t pedantic. It changes how you design the other three layers. If the first layer thinks, you trust its output and bolt tools onto it. If the first layer only pattern-matches, you build guardrails around it: grounding, verification, business-logic checkpoints. Different mental model, different architecture, different risk profile. The metaphor isn’t just wrong. It’s expensive.

What makes this historically important is not simply the scale of AI adoption. It is that governance itself is starting to become executable infrastructure. Once ministries shift from direct operators to supervisors of autonomous systems, the state begins transforming from a bureaucratic institution into a runtime coordination architecture. That changes the meaning of: authority, accountability, oversight, and even sovereignty itself.

Excellent post pro From a security perspective: LLM: Protect against prompt injection, jailbreaks & unsafe outputs RAG: Enforce data access control, sanitize retrieved content Agent: Apply least privilege + human-in-the-loop for sensitive/irreversible actions MCP: Zero-trust between services, strong authentication & encrypted communication

I would not be surprised that there is Strong oversight on the process that AI agents implement. While AI agents can handle data, humans are needed to handle situations. Leaving everything to AI agents decisions isn't wise.

Honest take: most AI video tools generate decent first drafts, but they fall apart when you need to iterate on a specific hook or scene. That's actually why I built GridVid — you can swap individual nodes without redoing the whole video. Curious what tools you've tested so far and what specifically isn't working for you.

Biggest mistake I see people make... they test Claude with fake work. Use it on real work from D1 or you'll never trust it enough to actually build something worth it

Luís Rodrigues Great breakdown. One thing I would add: in enterprise AI, the weakest layer is often not the model. It is the system around it. LLMs, RAG, Agents, and MCP only create value when data quality, permissions, governance, ownership, and decision flows are designed intentionally. Agentic AI is powerful, but without the right operating model, it can simply automate confusion faster. The real work is designing the whole system.

This is the responsible thing to do, and other frontier labs and politicians should be lining up behind them. The climate change parallel is sobering. In both cases, the people closest to the science sounded the alarm while political and economic systems moved far too slowly. The difference with AI is the timescale. Climate change unfolded over decades, and AI disruption could compress that into years. Olah's candour is valuable precisely because it comes from inside. But candour without structural change is just confession.

Ruben Hassid Artificial intelligence becomes substantially more impactful when it is connected to contextual information, persistent memory, and live business processes. In that environment, it no longer resembles just a chat interface—it operates more like essential infrastructure.

From my understanding, AI should not replace people, rather it should free up people to do what they do best: solving the world's unique problems and exercising the freedom to innovate. I appreciate Olah's honesty. However, in my opinion, handing over AI governance to governments have the potential to do more harm than good. If any government turns against its own people, this technology can destroy humanity resulting in the manipulation of information and distortion of facts in global proportions. For me, the best option is to let the markets handle it. Like Healthcare systems who have Compliance and Cybersecurity governance to set up frameworks to safeguard patient information, enterprises who operate AI labs should have similar governance standards that keep innovation within bounds. It's not perfect but I prefer this option over trusting governments with governing information to this degree. Olah's right, people shouldn't be trusted. Much more so with governments, just saying.

Great to see Anthropic taking a lead here and consulting sources of wisdom to guide AI adoption across a broad ideological spectrum. The real question is, is it all fluff or are they going to back it up with serious business strategy and values?

The scary part about AI isn’t just job replacement. It’s that humans get meaning, identity, routine, and social connection from work too. Society is not psychologically prepared for that conversation yet.

Luís Rodrigues This is a useful way to frame the AI stack. LLMs, RAG, agents, and MCP - each solve a different problem, but enterprise value only appears when they are designed together around workflows, controls, and business outcomes. Deepesh Khandelwal, MBA, PMP, SA

The real question is not what humans will do when AI does the work. It is what humans will do when the work they currently do loses its social meaning. Work provides structure, status, and community. If we automate the output without replacing those three things, we get efficiency without dignity. Pascal naming redesign of participation as the unresolved problem is the hardest truth in the entire AI transition.

This is the part that matters most for enterprise governance. The issue is not whether a particular lab is well-intentioned. Some clearly are. The issue is that good intentions do not neutralize the incentive structure. If frontier AI labs operate under commercial pressure, capital pressure, geopolitical pressure, and speed-to-market pressure, then enterprise buyers should not treat vendor assurances as a complete governance basis. That does not mean AI should not be adopted. It means adoption has to be authorized against present capability, disclosed limitations, model-change risk, output-shaping controls, and the actual conditions under which reliance is justified. Self-governance is not enough when the market is also selling the reliance. I wrote about this from the adoption-integrity side here: The market is selling the projection. The law has to make them show the machinery.

this is where the AI debate becomes larger than technology policy. Human dignity is not protected only by asking whether an AI system performs well after deployment. It is also protected by asking what roles AI is being sold into before adoption occurs. If AI is marketed as labour replacement, decision support, emotional support, professional assistance, or autonomous execution, then organizations need a clear account of what the system can presently do, what it cannot do, what remains unproven, and what human responsibility must remain intact. That is where ethics, governance, and regulation meet. The question is not simply whether AI can be useful. It can be. The question is whether institutions are being pressured to rely on AI before they understand the limits, assumptions, and human costs of that reliance. I wrote about this from the adoption-integrity side here: The market is not only selling software. It is selling reliance.

The real work task on day one is the instruction that separates this from every other AI tutorial. Most people start with experiments and toy prompts and wonder why AI never feels useful enough to stick with. Doing something that actually matters on the first day changes the relationship immediately because the value is real rather than theoretical. The voice file on day three is the other one worth highlighting. Most people never build this and then spend months frustrated that outputs do not sound like them. Giving Claude your actual writing samples and your banned word list is the setup that makes everything downstream faster and more useful. Seven days is the right frame because it creates enough repetition to start forming a habit before the novelty wears off.

Ramon Portillo, Ph.D. Maybe in some places, but I happen to be sitting in front of over 120 papers and have read them and graded them myself. There was no AI used to evaluate the papers. There wasn't even a TA. You might want to think about who you are throwing under the bus here with this assumptive statement about "academia". While different schools have different climates towards research vs. teaching, I think it is fair to say that many of us take our jobs as educators and experts in our fields seriously.