35,682 results · Most relevant
Bigger ARR is better for valuation. A $10M ARR business is worth more than a $2M ARR business with the same metrics. The $1-2M ARR threshold is critical for attracting institutional capital and private equity. $5M and $10M ARR thresholds attract larger private equity funds, typically increasing valuation. If close to $2M, aim to cross it before selling.
At 20% growth, the buyer pool changes; growth-focused private equity tends to step back below that threshold. If growth is stagnating and approaching 20%, consider selling now rather than waiting to stall into 15% growth.
Tara Seshan recommends working with agents as persistent teammates or coworkers. Let the agent do more and more independently, come in to provide steering at higher levels of abstraction, then let the agent continue to cook. Sync up at different cadences, look at each other's in-progress work, and provide iterative feedback.
Tara Seshan advises that building for where models are now or where you think they will be in a year leads to failure. The only correct timeframe is two to three months ahead, as both extremes are equally wrong.
When selling an AI agent product, be careful with what you promise it can actually do, because if it only delivers 80-90% of the value (e.g., replacing an SDR), customers will churn. Promise only what it can actually do, not what you wish it could do. Price an AI agent based on the value it replaces (e.g., a portion of the human's salary), not on a per-seat SaaS model. Use the rule of thumb: if you save someone $1M/year, charge $100K/year (one tenth). If an AI agent replaces an SDR making $80K-$100K/year, charge $8K-$10K/year as a starting point. If you have existing SaaS with a seat and an operator doing a task, build an agent to replace 80-90% of that role and charge 10-20x more per seat. Apply the SaaS playbook for pricing based on value and having a value metric to AI agents. When anAI
Jeff Cavaliere recommends performing three biceps workouts and three triceps workouts per week, alternating them, and taking one rest day after the six-day training block, then repeating the schedule for week two. He advises that if the audience is already doing another workout program, they should remove any exercises specifically dedicated to biceps and triceps from that program, but not remove exercises that use biceps or triceps as secondary muscles, such as pull-ups, underhand rows, bench press, or overhead press. He also recommends performing the arm workouts at the very end of the main training session.
Yi Chen argues that on the redesigned item detail page, image and price are the most important things to prioritize: put the image at the top, automatically trimmed to show the item at its best, put the price in the biggest font size, and highlight who's selling and how trustworthy they are.
Yi Chen recommends rewarding users with XP for activity so they earn reputation, making it easier for their posts to be featured or prioritized. XP-based reputation signals that a user is real and actively uses the app to finish personal tasks, so they can be flagged as high-quality.
Yi Chen recommends reducing the amount of information users see on the same page by separating the category hub into its own page, and showing only one category's subcategories at a time so the more detailed subcategories make search results more accurate.
Yi Chen recommends replacing an account page with a profile page that represents the user's identity on the platform, adding levels with XP, a streak, and badges to keep users coming back.
The event stream provided at-least-once delivery so subscribers receive all ledger events; the team carried idempotency concepts from the producer side forward through event streaming and made subscriber operations idempotency-safe so failed batch jobs can retry without creating duplicate ledgers.
The creator argues that all the operational questions people ask about live shopping are answerable by AI because all the information is a commodity. He advises using AI to answer detailed setup questions like how to set up an Amazon affiliate or a TikTok affiliate, how to ship products in packaging so they don't break, and what Whatnot's percentage on a sale is, including specific operational questions like how to ship sharp scissors via FedEx and whether Whatnot allows it.
Grubbs says the 1080 has been huge for return to play from lower body injuries (hamstrings, groins, calves) — common in baseball, football, and track. It lets you manipulate the force-velocity spectrum to accommodate the specific injury, loading musculature to the speed the athlete can handle, e.g., slowing a hamstring-injury athlete down to keep movement safe and progressively lightening load as comfort grows. She emphasizes working with ATs and PTs through the process.
her experience-based account
Leash gives reasoning agents a hypothesis-testing loop: show it binders and non-binders, let it form a hypothesis about why the binders bind, then let it pick 5–10 molecules from a catalog of millions to test the hypothesis, return results instantly, and update the hypothesis. Agents get a small seed set (e.g., 10 binders) and a budget (e.g., 100 molecules) to find the best binders from a library of 6.5 million — and they succeed within budget, 'so it seems like it's working.' The ultimate goal is to get the best binder in as few steps as possible, very much like a medicinal chemist would. Hard negatives (structurally close non-binders) are included to force the model to articulate why the binders bind rather than learning shallow correlations.
described as Leash's setup; numeric details flagged as possible transcription errors
The clinic's x-ray diagnosis taking an hour and 15 minutes is terrible and causes a 15 to 30 day booking delay for diagnoses. Being booked out 15 to 30 days for diagnoses is losing the business money, because patients in pain will go elsewhere if someone can see them sooner. Reducing diagnosis time so patients are seen sooner (moving from ~70% sure rates to 85-90%) would give roughly a 20% lift in revenue and return on ad spend from that one change.
projections are estimates, not measured results
Josh Woodward explains that Labs plans on a rolling basis with roughly a six-month horizon — thinking about where the product can move meaningfully forward in about 100 days — because every lab is cranking out new pre-training models a couple of times a year. When asked for a two-year roadmap, he answered that the team has a six-month roadmap plus a multi-year vision, since a thing seen two days ago can become a 48-hour project you cannot plan for, only recognize in the moment.
endorsed/described with 0.7-0.85 confidence
Tara Sesha recommends aiming for two to three months of where the models will be rather than being overly anchored in the present or the far future.
Varun describes pillar 4 as governed actions, which depend on each team, organization, access levels, restrictions, and industry. He advises defining the specific guardrails AI needs to operate under: read, write, underwrite, and the specific conditions for write.
Varun describes pillar 1, model orchestration, as having two layers: layer one is keeping up with the model treadmill with the latest models, and layer two is matching the best model to the task. He poses examples — Gemini for image-specific reasoning, OpenAI for deterministic steps, Claude for open-ended investigations — and says you need to constantly evaluate what the best possible model for the task is at that point in time. Resolve AI built its model-orchestration engine up front, very early.
Specific bar speed targets govern loading: don't go below 0.5 m/s on absolute strength lifts; stay above 1.8-2 m/s on technical coordination lifts; keep back squats at 0.46-0.48 m/s or faster (around RPE 7); operate the bench press between 0.55 and 0.7 m/s average velocity; target 2 m/s peak velocity on one box power cleans on day one; keep single leg squats between 0.7 and 0.9 m/s average velocity; keep power cleans at 1.65 m/s or higher for football players, and 1.65-2 m/s consistently for non-football athletes; hit high-speed power cleans at 2.5 m/s on speed day.
Use saved training trajectories and a controlled intervention to explain a failure, separating equilibrium theory from observed empirical stability.
Demonstrate all 3 repaired bugs and defend why a passed gradient check is necessary but not sufficient for generalization.
Use the tractable fixture and posterior predictive results to separate approximation, sampling and model errors; justify one revised modeling assumption.
Choose sources → to get a library of your own and a daily brief.
Map each implemented component to the original paper and explain 2 reduced-model failure modes without claiming full-scale reproduction.
Use calibration and shift measurements to justify a monitoring or abstention threshold, and give an explicit confounding counterexample to causal feature-attribution claims.
Defend a causal claim or refusal using the DAG, numerical estimates and a counterexample showing why predictive model fit cannot establish identification.
