PrincipleFor aircraft collision avoidance, layers of defense include strategic separation (different altitude layers by direction), air traffic control, and the collision avoidance system — you don't want a critical scenario to make it through all the layers of the cheese. Similarly, LLM safety needs all these layers of defense, including runtime monitoring.
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
Creator perspectiveClark was encouraged by recent frontier lab leaders saying that we need independent evaluators. One of the goals of Clark's company is to build systems that can serve as that independent evaluator for AI. An independent evaluator should not just say 'we're independent, take our word for it' but should produce independently checkable certificates: the evaluator checks the system, publishes the certificate, and others can go check it themselves and confirm the evaluator got it right. Independence in AI safety evaluation is really important, and it would be a real shame if all of the safety was left in the hands of the labs or even some friends of the labs who claim they're independent.
Y Combinator · AI Safety Fact vs. Fiction | YC Paper Club
Report · MoneySQL has around 15 different engines (e.g., Postgres, Oracle Database, VSQL) with specific hints and subtleties that make queries not work well across them. Pre-create proven SQL queries and let agents access those rather than having an agent generate SQL dynamically every time.
speaker's claim about ~15 engines; recommendation is his
AI Engineer · Memory-Aware Agent Harnesses for Anti-Money Laundering — Nacho Martinez, Oracle
Creator perspective · MoneyNacho Martinez categorizes agent memory into episodic long-term memory (conversation episodes, tool logs, previous conversations and their summaries), procedural memory (sequences of skills/workflows being run), and semantic memory (user identity, preferences, favorite frameworks).
AI Engineer · Memory-Aware Agent Harnesses for Anti-Money Laundering — Nacho Martinez, Oracle
Report · MoneyOliver Kharraz claims the provider-data problem is to a large extent solved, explaining that provider directories go stale when published without transactions, whereas the transaction itself enables self-validating data. ZocDoc's data accuracy compares capably with average insurance directories.
The a16z Show · Why AI Could Finally Fix Healthcare Access | Zocdoc CEO Oliver Kharraz
Creator perspective · MoneyDoctors have last-minute cancellations, no-shows, and reschedules that create last-minute capacity, which Zocdoc surfaces to patients and recaptures — this is how the company achieves its three-day average wait time. Kharraz draws an analogy: airplanes used to fly 60% full in the 1970s, but with the advent of the internet and booking platforms, load factors are now over 90%. In the wild, doctors have load factors around 70%. He adds that the best doctors tend to have the worst utilization because their busy patients cancel more often and they don't double book, and that Zocdoc manages its network so patients can have the three-day experience, adding providers where capacity is low in a specialty-zip code-insurance quadrant.
The a16z Show · Why AI Could Finally Fix Healthcare Access | Zocdoc CEO Oliver Kharraz
ReportEach agent in a flow is spawned into its own session thread with physically separate context windows in the database, and context windows do not leak via file systems or anything like that by default.
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
PrincipleStoring API keys as plain text files (e.g., myapikeys.md, myapikeys.txt) inside a sandbox VM risks credential leakage even if hidden or in secret directories; users can prompt an agent product to extract plain-text credentials stored in its sandbox — there's just always a risk that plain text somewhere could make it out.
described
AI Engineer · Why One AI Agent Isn’t Enough (And How to Run 1,000) - CJ Avilla & Harrison Stall, Anthropic
PrincipleFailing to push a merged code layer back to the branch or open a pull request prevents teammates from merging the improvement into the repository, so merged work should be pushed or submitted via pull request.
Figma · Explore ideas in code | Figma
Principle · MoneyThe brand ran no ads for its first eight months, waiting until the product was validated organically before turning on paid acquisition.
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Creator perspective · MoneyEvery hat since the first has 'if you know you know' embroidered on the side as a subtle nod between dads — a signal that you know what their life is like, because once you become a dad your life changes forever in a way only other dads understand ('hey, I know what you're going through').
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Principle · MoneyThe brand started with a small initial run of 100 hats, partly a manufacturer minimum, to test demand before scaling. After selling out, the next batch was 300 hats rather than a large jump. The founders also avoided print-on-demand in favor of a direct factory relationship reached through a mutual connection.
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Principle · MoneyThe founders commit to posting on Instagram every single day without missing a day, and when launching a product they accompany posts with the story of why they made it; when they had no product to sell, they kept posting content and spreading the brand message to maintain momentum.
Foundr · We Started Dad Gang with $750 and One Hat - and did $40M in 4 Years
Principle · MoneyThe speaker points to special ops teams as an example: they are formed through elimination, cutting a gigantic pool of candidates. He suggests treating the first 90 days of onboarding as a probationary elimination period, and encourages owners to be willing to make cuts — lubricating the organization so managers and leaders can let go of mediocre people — and to create rewards for people who let go of bad performers.
Alex Hormozi · My honest advice on how to raise your team’s bar
PrincipleWhen making lots of copies of DNA, subtle biases can make some sequences amplify faster than others, so this must be controlled for. The approach is to include a little bit of unique sequence so that when you have lots of copies, you can collapse them and know they all came from one event. Relatedly, negative controls — like sticking the bead in by itself — are needed to account for pan-protein stickiness signals that would pull out certain chemicals from the library.
None significant.
Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
Creator perspectiveQuigley describes work by several groups showing that current state-of-the-art co-folding models memorize pocket locations rather than understanding physics: when the amino acid sequence is changed to break a pocket (fill it, remove it, or enlarge it), the models still shove the small molecule into where the pocket used to be. He attributes this to a data gap — structural training sets lack good negatives because only intact protein-plus-ligand structures make it into them, and nobody makes crystal structures of pocket-breaking mutants, so models never learn what disrupted binding looks like.
described with attribution to multiple groups' findings
Latent.Space · Lab-in-the-Loop at Compute Speed — Ian Quigley, Leash Bio
Creator perspective · MoneyYou cannot rely on Anthropic or OpenAI to stop something like this from happening to you; they can make smarter models, but large language models have two qualities that make them impossible to fully trust. First, LLMs are non-deterministic — there is always some randomness or chance associated with what they do; you can ask an LLM to do the same thing 10 times and it could do it 10 different ways. Second, LLMs have no structural separation between instructions and context: if something in the context conflicts with your instructions, the model won't inherently prioritize your instructions, because from its point of view all the information is coming through the same door.
AI Engineer · Your AI Isn’t the Problem. Your Infrastructure Is. — Jiachen Jiang, Coder
Creator perspective · MoneyThe four decision labels — data, model risk, third-party risk, and audit — represent decisions your deployments depend on. For each label, find a person who can explain the decision, the evidence they need, and the next step.
AI Engineer · Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers
Creator perspective · MoneyA common failed optimization approach is handing a massive prompt to a smaller, cheaper model, running it against a few thousand historical production examples, and discarding it when it disagrees. OutRival's cheaper candidate disagreed with their expensive production baseline about 18% of the time, so they initially threw it out. But language models in production deal with inherently ambiguous real-life inputs, so disagreement is a composite metric that introduces a critical hidden variable: the baseline variance of the model itself. When the input is messy, your production prompt will inevitably disagree with its own past decisions.
AI Engineer · How to Actually Tokenmaxx and Save Millions — Alvin Pane, OutRival
Principle · MoneyIf you record properly why the failure occurred, and someone else was able to approve it because it was a known failure mode, the change passes your checks; without a record, blindly trusting the agent pushing code is dangerous because the test could have been low quality.
AI Engineer · Rebuilding CI for AI-Scale Commit Volume — Gautam Korlam, Sonar