Principle · MoneyKeep a good-enough grasp of what the machine is doing. Going full black-box mode eliminates your accountability, and blame-shifting on the basis that someone else will review it is a real team failure mode.
AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo
Principle · MoneyWhen you rely on an agent's output without an evidence trail and it turns out to be wrong, you bear the consequences — 'that's what Claude gave me' is not an acceptable justification.
AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo
Principle · MoneyThe speakers argue that evidence is not just about where a number came from but how it was derived, and that curated, labeled data plus gathering as much evidence as possible is paramount. They stress that evidence must be embedded in the code itself, not just referenced externally.
AI Engineer · Stop Chatting With Spreadsheets — Francesco Galletta & Felippe Felisola Caso, Monzo
PrincipleApplying Occam's razor to AI method selection: the simplest explanation is more likely true, and simpler, more interpretable methods are more likely to be robust. If linear regression or decision trees suffice for an application, use them instead of neural networks.
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
PrincipleMichael argues there is always a long tail of failure modes whose sum is a high probability, that even ACAS-X has nonzero failure probability in incredibly unlikely situations, and that ICAO/FAA safety standards do not require zero probability. He advises not believing anyone who claims a fully safe real-world system.
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
PrincipleClark argues that in a world where writing code is easy and verification is becoming easy, the remaining hard thing is the specification. He distinguishes universal specs (memory safety for C) from domain-specific ones (no loss of funds for a bank), and notes enterprise code often lacks any specification at all, raising the question of how to formally verify it.
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
Principle · MoneyWhat people should focus on is the evals — what's your evals and what's the results that lead to it is the right question, not the tuning method.
AI Engineer · The AI Bill Is Too Damn High: Tokenomics 101 — Eugene Cheah, Featherless AI
Principle · MoneyHarish Gaggar stresses monitoring and complete auditing for tool executions, arguing that AI agents cannot be trusted without continuous monitoring, evaluation, and feedback.
AI Engineer · Paving the Road to Production: Governed Enterprise AI Agents — Harish Gaggar, Credit Karma
Principle · MoneyHarish Gaggar recommends keeping a human-in-the-loop process for every action item and every high-risk item, with write and sensitive operations routed through HITL approval before execution.
AI Engineer · Paving the Road to Production: Governed Enterprise AI Agents — Harish Gaggar, Credit Karma
Principle · MoneyHarish Gaggar advises building a full enterprise framework with governance underneath rather than rushing a single agent to production, because teams that ship quickly later realize the frameworks should have been in place from the start.
AI Engineer · Paving the Road to Production: Governed Enterprise AI Agents — Harish Gaggar, Credit Karma
Principle · MoneyYou cannot rely on Anthropic or OpenAI to stop AI incidents from happening to you; they can make smarter models but LLMs have qualities that make them impossible to fully trust. Large language models are non-deterministic, meaning there is always some randomness or chance associated with what they do — you can ask it to do the same thing 10 times and it could do it 10 different ways. What you need to control is the infrastructure that your AI is running on and the environments it works inside, rather than relying on the model itself. Controlling infrastructure rather than the model is the difference between telling someone don't leave a room versus putting a lock on the door and actually using a key. Agents act at machine speed, so if a mistake is made it can make it 20 times or more more
AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder
Creator perspective · MoneyThe core argument of the episode: what you need to control is the infrastructure that your AI is running on and the environments it works inside, rather than relying on the model itself. Paying five times more per token for a smarter frontier model is not the solution to agent mistakes; infrastructure controls are.
AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder
Principle · MoneyThere are six key parts to a secure architecture, and you're going to want to layer these six. The building blocks are ordered roughly by leverage, where number one takes the most risk off the table and number six does the least. If a well-meaning but ultimately suggestible agent gets past just one of the six layers, there are still five more to go. Even the first three layers could have stopped every incident shown. Identity-aligned access control and short-lived credentials are not new ideas and mattered a lot for human developers too, but they take on new urgency when AI is in the loop.
AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder
Creator perspective · MoneyThe problem with the Replit agent was not that it was too smart or too dumb; it was that it had way too many privileges by default, including root-level access to production and the ability to run destructive commands without human approval. The agent was not taken over by an external malicious actor; it was just trying to follow instructions and did exactly what the user gave it permission to do. The agent also said a line that stood out to the speaker, a good reminder that these systems are inherently non-deterministic — you cannot predict what they'll do in a given moment, just the range of what they're capable of doing.
AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder
Creator perspective · MoneyDoing RAG for a financial institution comes with specific challenges: every answer must cite the exact rulebook section and page so users can verify it, and because the underlying LLM is a general model, the team had to decide which knowledge it is okay to use — general knowledge like what a bank is is fine, but information about other payment schemes' fees or workings must not end up in an answer unless it is in the rule books.
AI Engineer · Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket
Creator perspective · MoneyNot every question fits in 10 chunks: some are chained, with one rulebook pointing to another, and the answer is only complete once you follow that pointer. Others need breadth — 'list every fee for this transaction type' is never going to fit in the top 10 of anything. Instead of a fixed retrieve-then-generate pipeline, the agent has four tools, ordered narrow to wide.
AI Engineer · Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket
Principle · MoneyA model does not become more capable by itself; for each AI workflow, deliberately decide how much authority that particular workflow should give the model.
AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers
Principle · MoneyThe takeaway Paden Gayle wants people to walk away with: discovery is the engineering. Draft the spec, then shrink it; build the evidence pack in sprint one.
AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers
ReportControlling typography from a central location lets you affect everything in the document and make changes at scale, and creating the system takes more time up front but can then be published and reused in other files.
Figma · Figma for Edu: Working with typography systems
PrincipleBy controlling type settings from a central location (variables), you can affect everything in the document and make changes at scale; once variables are bound, updating a variable value changes it across all text that has it applied.
Figma · Figma for Edu: Working with typography systems