35,595 results · Most relevant
Dynamically discover the schema at runtime by sampling a bunch of documents in the document store and deriving a schema from that sample. Mongo is schema-less, but the same dynamic schema discovery principle can apply to non-schema-less databases. Discovered schema can include indexes that help improve the performance of queries. Inferring enums is a tricky part of dynamic schema discovery.
endorsed
A memory object should have an ID (like a primary key), text content, JSON, tags, and document data. It should have a vector embedding created over the content with a vector index for similarity search, and ideally a text index to facilitate full text keyword search that combines with vector search. Memory objects can have graph links to different edges so memories are related through an entity graph, providing additional context pulled back during the search stage. They should include temporal facts — when the memory is valid to and the last time it was used — and, super importantly, audit events: who created the memory, what it was used for, the log of how it's been used, and associated metadata.
Anders Swanson describes a hybrid search retrieval flow: apply tunable weights to the different query branches (vector distance, contains operator, graph table, temporal data); fuse results from multiple branches, apply thresholds, and re-rank to produce a context card with scored evidence; merge the branch results to an overall score and take the top K results.
endorsed
Anders Swanson walks through a memory formation flow: create the vector embedding of the memory content, then form the episode (slices of transcripts that created the memory) for provenance and additional context; apply any organization-specific policies during memory formation; then save down to the storage layer. To create episodes, take the agent transcripts, filter them, find candidates for episodes, save the memory record, and link it back using relations or graphs — he particularly likes graphs because they allow a broader search.
endorsed
To detect overfitting without a test set, split the training set into two parts: one for training and one held out from the training algorithm. The held-out portion is used to identify when the model starts to overfit. During training, both training loss and testing loss decrease until they start to converge, after which divergence continues for the rest of training. Identify the point where testing accuracy is maximized or testing loss is minimized, and use that model for deployment — this is early stopping, a regularization technique where training is halted at the point of best test-set performance. Models on the left side of the early-stopping line are underfitting, and models on the right side are overfit because divergence has occurred.
None stated.
To get the output of a trained perceptron, take the inputs, multiply them by the corresponding weights, add the bias, and apply a non-linearity. The lecturer emphasizes remembering how a perceptron works in three steps: dot product inputs with weights, add a bias, and apply non-linearity — this defines the entire perceptron forward propagation.
None stated.
Will describes a multi-step method for turning agent memory into skills: identify a scope (an entity's subgraph or a specific conversation), distill a workspace- or project-scoped context graph into a skill grounded in observed data with deterministic procedural steps, and keep most pipeline steps deterministic, using an LLM only for synthesizing claims and generating skill description text.
The team describes using density functional theory to estimate formation enthalpy — a measurement of the energy a material has. If that energy is higher than other materials the atoms could form into, the material is unlikely to be synthesizable; stability means lying on the convex hull of other materials and their energies. DFT is the best available method for predicting stability of new materials — not perfect, but better than alternatives they can think of.
Factory evaluates agent readiness across eight pillars: validation (linters/formatters), build system (documented CI commands), feedback loops (unit/integration tests), docs (readmes, agents.md), reproducible dev environments, code modularity with clear boundaries, observability (time to find root cause), and proactive security scanning.
Dr. Morin recommends using tempo sprinting at about 90% of max speed instead of maximum velocity running for high-speed exposure, because it is easy to understand, do, and verify. For young people, 100 to 150 meters per week can be an interesting dose (he notes he has no science to support this, it is based on opinion and experience). He prescribes 120 meters per week as a good starting dose, structured as two sessions of three reps of 20-meter tempo sprints: 20 meters to accelerate up to speed, 20 meters cruising at 90%, and 20 meters to decelerate. In a high-speed session, do not count the build-in distance as part of the 90%+ volume — only the flying portion counts toward the objective. He also prefers 20-meter flying sprints over 10-meter flying sprints to get more work, and notes that
Consequential creator advice requires independent verification before Today can promote it.
Liemandt advises making learning as easy as possible for struggling students as the first step. He warns not to put an eighth grader whose knowledge grade is third/fourth/fifth grade into eighth grade material — it's impossible and not doable. Instead, give struggling students individualized lessons at their actual knowledge grade level (e.g., third, fourth, or fifth grade lessons) so the work becomes doable. He notes matching lesson difficulty to the student's real level transforms the work from impossible to doable, and that in the Alpha system, for the first time students are in a lesson that's not too hard or not too easy — it's right for them.
Liemandt explains that everybody has a forgetting curve, and forgetting curves are different by person. With spaced repetition, remind the learner right before they're going to forget the material; if you do that over a few times, the material gets into long-term memory and will be there when you need it in the future.
endorsed method by the guest
Liemandt contrasts two ways to learn. One way is to give a math problem with no instruction and let the learner figure it out. The better, faster way is to give a worked example where you show the solution, then progressively remove the parts of the example so the learner progressively goes up until they can solve the problem with no scaffolding. He describes this as incremental learning with an example at the front end, and says it's a faster way to teach kids than just saying 'good luck' with no instruction.
endorsed method by the guest
For technical coordination movements (cleans, snatches, jerks, push presses), measure peak velocity — the highest instantaneous bar speed. For absolute strength movements (back squat, front squat, single leg squat, bench press, deadlift), measure average (mean) velocity. Use those numbers, tied to the individual athlete, to determine the next set and how the rest of the session flows; average velocity also builds a guide for future training by correlating speeds with load per athlete. If you don't have a VBT tool, train off percentages instead.
creator's endorsed method
Pick anchor lifts — the main priority movements the whole training system is based around. In this program, one box cleans anchored technical coordination work (their velocity trend reveals the athlete's state and how twitchy they are), the back squat anchored average velocity for lower-body strength (determining how heavy to go), the bench press anchored the upper body, and the flying 10 anchored max speed.
creator's endorsed method as applied in the program
Write a prompt that gets sent to the LLM to determine how a file should be tagged. Review the evidence showing where in the document the AI found support for a tag, review the confidence score returned by the model for each tag, and use thumbs up/down feedback on AI tags to provide few-shot learning examples that improve tagging accuracy over time.
Copy your use case description into the tool so tag suggestions are grounded in what you're building; leverage AI to auto-suggest which tags you should have (flat or multi-level taxonomy) and which classification values to use when you don't know them upfront. Upload existing taxonomy valid values via CSV if you already have them. Add conditional sub-tags so only relevant file categories (e.g., legal and contracts) pull out more specific data like contract type. Use human-in-the-loop review to choose which AI-suggested tags are relevant before finalizing.
Huberman lays out a visual focus training practice: pick a specific visual target (a wall, blank screen, or crosshatch on paper) and visually focus on that one location for 30 seconds to three minutes while breathing normally and staying relaxed. Refocus on the target whenever your gaze drifts — the refocusing is the key component of training the brain to focus better. Add five seconds to the duration each subsequent day; if you miss a day, simply repeat the same duration as your last session. He recommends a 30-second to three-minute visual focus period at the beginning of a 90-minute work or physical work bout, and using visual focus on a location as a ramp-up to direct focus back to weight training sets and reps, running, or cycling. For musical practice or math, focus on somethingother
Huberman recommends a visual focus practice: pick a location such as a spot on a wall (not your computer screen; blank the screen or use a crosshatch, or imagine a target), breathe normally, stay relaxed, allow blinking, and force yourself to refocus on the target whenever your gaze drifts. The refocusing component is the key part of training the brain to focus better.
Instead of just saying 'I work at a nonprofit,' say 'I love helping animals, so this place lets me do that, and honestly, I'm getting to live by the beach, and that's not such a bad perk either.' Answering with what you care about shares two things at once — in the example, helping animals and being by the water. Lacing enough values-based answers into conversation stops you wasting time on people you were never going to connect with and helps you find ones who match who you actually are.
Choose sources → to get a library of your own and a daily brief.