35,595 results · Most relevant
Jeff Cavaliere explains that the landing from the spinning roundhouse kick caused the ACL to blow out. The foot becomes fixed on the ground while the hips continue to spin, creating rotational torque. Ground reaction forces from the jump amplify the force. This torque, combined with anterior translation of the tibia, is a primary mechanism for ACL tearing.
Within the agent reflection step, the first mini-step is clustering: group failing rows (users and traces from production) by failing dimension using pure clustering without any LLM calls. The second mini-step is LLM-based root cause assignment: within each cluster, run the LLM to diagnose each row and assign a root cause plus a tag — e.g., whether a low dimension score is due to prompt, agent, harness, code, judge, or rubric issue. Because LLM-generated root causes are free tags, similar tags may describe the same thing without being identical, so the third mini-step is to merge (dedupe) the issues.
GEPA works by taking an AI pipeline written in any agentic framework or even raw LLM calls, running the system on a few examples, and collecting domain-specific feedback from whatever information the environment contains. It then runs reflection with an LLM or agent that reads the feedback and proposes a better prompt. GEPA keeps a Pareto pool where it keeps every single candidate that wins on even one training example, not just the top scorer.
Alignment is the process of annotating the examples added to your dataset, then running the eval and seeing if it matches the human feedback. Run the eval on annotated examples and check whether its output matches the human feedback.
TurboPuffer's vector index automatically clusters similar vectors into groups. When building permissions indexes, access is determined per cluster: a user has access to documents in certain clusters but not others. Clusters the user has no access to can be thrown out entirely and not considered at all during search. The search procedure is: first find all clusters the user has access to (i.e., clusters where they have access to at least one doc), then search the closest clusters, then scan sequentially inside those clusters. In TurboPuffer, clusters currently have about 200 documents each — a number being tuned and likely to increase over time to maximize throughput. This basic cluster-based strategy fundamentally makes search at scale with permissions possible.
Huberman advises that when thinking about supplementation for sleep, first ask how well (how deeply) and how much you are sleeping per night. If you sleep deeply and enough, feel rested, and need only a short nap or no nap, there is no discussion to be had about sleep supplementation. If not sleeping enough, first ask whether you are ingesting caffeine after 2 p.m., and if so, limit or eliminate it. Most people would also do well to avoid food within the two hours prior to bedtime, while not being so hungry that they can't fall asleep.
Tom Foley outlines an approach to pairing typefaces: first understand a typeface's calligraphic root — tracing it back in time to which calligraphic tradition its structural aspects come from. The next thing is to start to look at structural things and proportions, like x-height, which is a big one. As a concrete example, if you have a serif headline typeface in a modern or Bodoni style, those typefaces always pair better with grotesques because they have the same structure — the modern/Bodoni model of high contrast, shoulders blending into stems, and fine serifs eventually evolved into grotesque typefaces.
Test agent consistency by varying inputs that should not affect the output — supplier name, supplier address — and verifying the number of flags stays the same; reorder clauses to introduce randomness and verify results are unchanged; run the same contract repeatedly to confirm consistent output. All of these checks can be performed with code, avoiding LLM-as-a-judge.
To write well with models, first have an extremely long and detailed prompt — it cannot be long enough, and the more the better. Add every example you can think of what to do and how to do it, plus negative examples: things you don't like, things that don't sound right, and words you don't want. If you write newsletters, you can share your writing in full and it will just be better.
None stated.
Corey Gallon's review method comprises three levels of review, one addressing each of the three ways agents fail, and each level was built as an agent skill. Each level adds to the one before as the stakes go up.
A verification tool can include a doctor command that quickly checks if the database is running and there's network connection; drive commands to join the waitlist, reload the page, and reset all data; snapshot, network storage, console logs, local storage, and screenshot capture; a prove command checking for deterministic output; and a judge using a System 1 model for directional checks. It can also manage the health of the application — bring it up, take it down, run health checks.
Consequential creator advice requires independent verification before Today can promote it.
Voss structures the CI eval file like an ordinary test file because it is one: a describe block with a test for each query in the regression set (Coca-Cola, Apple versus Microsoft, Rivian, and the fake ZZZQ). Each test runs the agent, logs the report as the output, and then runs the same evals (ticker check and actionability judge) written earlier — no new evals are needed for CI. Running npm test runs the suite like any test suite, printing out a scorecard at the end. When a build fails, you don't just get a red cross — you get an explanation of whether things are good or bad and why they went wrong.
Laurie Voss recommends writing notes first and categories second: read all traces and write down in human language what you see (open coding, a term borrowed from ML), then group the notes into categories (axial coding). Guessing categories in advance is an anti-pattern because you don't know your failure modes ahead of time — categories end up too narrow, too broad, or wrong.
Harish Gaggar describes the MCP gateway request flow: on the first query the client sends a one-time authentication request to the MCP gateway, which saves session state in Redis and returns a JWE access token; the client then submits a tool/call request on a given namespace/MCP with the auth bearer JWE token; the gateway looks up the namespace (MCP server), unwraps the JWE token, attaches the user token, and forwards the request; the MCP server makes the API calls to the service endpoint, exchanges the ID token, and returns the response based on the user's authorized scope; the server unwraps the data, makes a payload compatible for the gateway, and the final payload is sent back to the client.
Uber built Summit Queue, a merge-queue system that tests all likely combinations of incoming commits in parallel, predicts which PRs might break, removes breaking PRs so authors can fix them, and only lets a PR merge to trunk if it passes with all combinations — keeping the trunk green. It is still running at Uber today.
Follow the 3-to-5 concept for speed and power training: 3 to 5 days per week, 3 to 5 exercises, 3 to 5 repetitions per set, 3 to 5 sets, and 3 to 5 minutes rest between sets. Execute every chosen exercise with high intent — this part is critical to the concept working. The framework has been tested extensively in coaching and scientific realms, is productive, easy to follow and grasp, runs true across speed, power, and related training qualities, and scales from a very low volume option (3 days a week, 3 exercises, 3 sets of 3) up to a high volume option (5 sets of 5 of 5 exercises, 5 days a week).
The communication cadence for kickoff is six months out, two months out, six weeks, four weeks, then weekly until one week out, then every other day. Six months out: only dates, location, and a calendar-hold request. Two months out: high-level program outline (arrival day, general session, breakout sessions, departure day). Six weeks out: reveal a few exciting details like the closing night party theme. Four weeks out: QBR deck requirement, prep expectations, and a more in-depth agenda. From four weeks out, run a weekly drip answering the questions everyone will ask: dress code, weather, agenda details, airport transportation, ride-sharing encouragement, and a Slack channel for questions and excitement. The day before kickoff, send a checklist of things attendees need to do, registration,
Somya Rai walks through a GNN layer: first gather each neighbor's hidden states across typed edges (transaction-to-user, transaction-to-merchant); then aggregate the neighbor set into a fixed-width vector with a permutation-invariant function — a choice that mattered more than expected; then update by concatenating the aggregate with the node's own state and pushing it through learned weights; finally stack the layer twice so every transaction knows about its neighbors and neighbors' neighbors before the classifier sees it.
The gold standard for assessing movement skill is a highly qualified physical therapist or movement specialist who can identify all your movement patterns. To self-assess, go joint by joint through the major joints (shoulder, elbow, low back, hip, knee, ankle), do a representative movement for each exercise you actually do, and perform an upper body press, an upper body pull, a lower body press, and a lower body pull — for example a push-up, a pull-up or bent row, a squat, and a deadlift.
endorsed method with a gold-standard alternative
The gold standard for VO2 max assessment is going into a laboratory for a test with a mask collecting all your gases while you're run to maximum; a true maximal test has a very specific completion protocol any exercise scientist will know. Without lab access, estimate VO2 max with the 12-minute Cooper test: run for 12 minutes as far as you can, record the distance, and enter it into an online calculator. A gentler alternative is the one-mile walk test (a submaximal Rockport test): walk a mile, record the time and your heart rate at the end, and enter those in to estimate VO2 max. The walk test suits people who can't or don't want to run 12 minutes maximally, or who have knee or back pain when running — it's used a lot in the executive program and is pretty accurate if done correctly. The
estimates rather than direct measurement
Choose sources → to get a library of your own and a daily brief.