Principle · MoneyTrimming a genuine user's transactional history to hit a class ratio breaks the connectivity their neighbors depend on. Downsampling the easy way keeps the neighbors but loses the neighborhood that makes them poor examples, and every negative drawn in graph sampling takes its edges with it — unlike a tabular dataset where rows are independent. Instead, handle class imbalance downstream with balanced batch sampling inside the trainer, and keep a fraud user's own genuine transactions in the graph as hard negatives — the most useful examples for the decision boundary.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Principle · MoneyMichele Grant insists that core math and reconciliation rules must run in deterministic finance-approved code, not generated as conversational prose from an LLM. Treat the model's written summary as a helpful explanation, not proof — the real proof lives in the telemetry: what records were pulled, what deterministic tools ran, and who signed the review. If reported savings do not match the general ledger, the agent does not touch the number; it flags the discrepancy and routes it up.
AI Engineer · The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale
Principle · MoneyMichele Grant warns to pay close attention to stage four: once humans only look at exceptions, your biggest vulnerability is what the agent failed to flag. She advises tracking missed material issues through independent sampling, monitoring error severity, and watching the override rate carefully.
AI Engineer · The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale
Report · BodyDr. Cleather's most critical principle when teaching someone to lift is to take their brain out of it. He uses a 'ready, jump' cue where beginners move on the coach's command, manipulating space and time so the lifter cannot overthink.
Garage Strength · Why Most People Train Explosiveness Wrong w/ Exercise Scientist Dr. Dan Cleather
Report · BodyJake Tuura describes Seth O'Neill's research standard: a single-leg seated calf raise performed in a decent level of dorsiflexion, with a cutoff of two times body weight — producing two times your body weight with a single leg in a seated position. Per O'Neill's idea, if you hit that standard, you're pretty good. Track and field sprinters and jumpers find two times body weight easy and produce far more, while other team sport athletes not granularly performance-oriented probably won't have as good a number.
The standard is attributed to Seth O'Neill's research; athlete comparisons are the guest's descriptions.
Garage Strength · How To Make Weak Tendons Strong w/ Tendon Expert Jake Tuura
Report · BodyJake Tuura describes the Berlin protocol, which has only been done specifically on the Achilles tendon with a straight leg, and measured changes in Achilles tendon stiffness. It was performed four times per week, taking about 10 to 12 minutes per session. You push for three seconds as hard as you possibly can at a mid-range position or a little bit of dorsiflexion, then switch and do the other leg for three seconds, alternating legs. It's four repetitions per set, followed by a two-minute rest, repeated for five rounds — almost like clusters, a really specific set-rep scheme that can be looked up by searching 'Berlin method'.
Frequency and duration are the guest's recollections ('I think it was four times a week').
Garage Strength · How To Make Weak Tendons Strong w/ Tendon Expert Jake Tuura
Principle · BodyThe more advanced a lifter is and the more compound the lift, the closer to failure they need to train to recruit all fibers; advanced lifters should probably train within three to five reps from failure, and as a rule of thumb within one to two reps from failure, especially when compensating with more volume or heavy loads. Isolation exercises allow more reps in reserve (leg extension vs. leg press, pec fly vs. bench press).
medium
Garage Strength · How To 2x Muscle Growth In Half The Time w/ Muscle Expert Borge Fagerli
Principle · MindBrooks distinguishes pleasure from enjoyment: the pursuit of pleasure alone is not the secret to happiness (jokingly, it's the secret to rehab). Enjoyment is created by taking pleasures and adding people and memory, delivering them to an executive experience. He advises learning the difference between the two and consciously converting pleasure into enjoyment.
Andrew Huberman · The Art of True Happiness | Dr. Arthur Brooks
Principle · BodyWhen 12 or more weeks out from football season, work in the 82 to 95% intensity range for almost all training. Once inside the 12-week zone, work up to 90 to 95% intensity only once a week. If strength is maintained through the season based on speeds and percentages at around 90 percent of speed work, the athlete maintains that strength. Once outside the football season, rebuild and get ready for the next season.
Garage Strength · Why Everyone Should Power Clean for Football
Principle · SocialThe host ranks pointing out what you notice about her in the moment at S-tier and calls it one of the best things you can do. Breaking away from the current topic to comment on something about her right now is really powerful: because women really care about other people's perception of them, it instantly brings the conversation back to being about her and snaps her attention onto what you're talking about.
Coach Kyle · What Women Actually Want to Talk About (TIER LIST)
Principle · MoneyHormozi urges getting a booth at B2B conferences where target customers attend and securing a speaking slot — even just 15 minutes — to pitch services. He argues B2B event clients are worth doing whatever it takes to close because they front the cost of gathering 500-1000 of your ideal customers in one stadium; closing even 1% of a 1,500-person room (15 clients) would be a 50% increase in business. When your ICP is gathered, treat the event as a lead generation opportunity.
Alex Hormozi · How to Get Rich for Service-Based Businesses (Thank Me Later)
Principle · MoneyCredentials never entered the sandbox in the new architecture — calls needing tokens are routed through a proxy that injects the token before calling the external endpoint. This means prompt injection can never extract tokens because the model never sees them; storing credentials in the sandbox makes you more susceptible to prompt injection.
AI Engineer · https://www.youtube.com/watch?v=Uv0nAf_7tyU
Principle · MoneyYoni Michael argues that the important part is not that the agent is right the one time you tried it, but that when you try it 10, 15, 20 times in a row, it gets it right every time. So run evals many times over rather than once to measure consistency, and use pass@k and pass^k metrics to evaluate agent consistency across multiple runs.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyWhen attributing ARR by marketing channel, an agent should not pick a join path based on column names — it should check whether each join key is actually unique by profiling the actual tables; join recommendations should be made with actual reasoning and evidence grounded in the data itself.
AI Engineer · The Data Context Layer for Agents — Yoni Michael & Brandon Callender, Typedef
Creator perspective · MoneyWhen deciding whether to stick AI into a workflow, first ask: is it a bound system? Second: is it worth it? Third: does it have a verifiable output? If the answer is no to all three — it's not bound, not worth it, and has no verifiable output because you just want it to think and make decisions for you — the answer is no, and you should look at other workflows.
AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog
Creator perspective · MoneyWhitney Lee explains that if a user types their credit card information into a chatbot unprompted — say an order-your-burrito-here button scenario — that PII ends up in your traces and on your system, which is a liability. Input guards are great for scrubbing anything users try to tell the system; you don't want to know their social security number or credit card number. With input guards scrubbing PII, you have proof that the sensitive data never got into the system. Accepting PII from anybody puts you in compliance land (e.g., GDPR) with extra cost, even if you don't have anybody in Europe.
None stated.
AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog
Creator perspective · MoneyForrester says the most common question he's asked is when should we stick AI into our workflow/platform. If you talk to anybody in Silicon Valley, the answer is always yes, 100%, as much as you can. His rubric: (1) is it a bound system with known boundaries and edges, preferably with an API attached or at least a well-known set of interfaces — though he'd be careful applying this to Kubernetes and cloud systems; (2) is it worth it — noting NVIDIA's CFO said about 70-something percent of the jobs being done by AI today could be done by humans more cheaply and more reliably, with the CFO's argument being that he's trying to lower the cost of AI so people can buy more, which the speaker found very telling; (3) does it have a verifiable output — if the AI generates software artifacts like Tof
The NVIDIA CFO figure is a quoted paraphrase ('70-something percent').
AI Engineer · Build a Platform and Watch It Burn — Michael Forrester, Accenture & Whitney Lee, Datadog
Creator perspectiveOmar argues that scaling document parsing to a million documents before confirming extraction accuracy on a small sample leads to bad results at scale.
AI Engineer · Parsing the Epstein Files for JMail — Palak Agarwal & Omar Alhait, Reducto
Principle · MoneyUse uniform conditions — the same setup every time — because changing even one variable makes it basically a new benchmark. If evaluating multiple agents, decide in advance which agents you look at. Always do multiple runs: a single run is benchmark lottery, a random result, just a ballpark number. Wolfram does five runs per benchmark in WolfBench, but you could do more if you have the money — it's always a tradeoff.
stated
AI Engineer · From Zero to Leaderboard: Agent Evaluation — Wolfram Ravenwolf, Weights & Biases
Creator perspective · MoneyJosh Woodward's first tip for setting up a lab team is to create a space where weird things can grow and not nestle it under an existing business unit — it has to have its own independence. In some companies that means the lab reports to the CEO, or at least at enough of a distance.
endorsed with 0.9-0.95 confidence
Lenny's Podcast · Why every company now needs to think and operate like a lab team | Josh Woodward (VP Google Labs)