Creator perspective · MoneyBy sharding, caching, and reducing compiler churn, Guitar brought incremental CI time from ~10 minutes to under a minute. They cached outputs aggressively using their Rust build system and added fixes to make compilation more cacheable and invalidated less often.
AI Engineer · Rebuilding CI for AI-Scale Commit Volume — Gautam Korlam, Sonar
Report · MoneyDintai Phong is Dyna Robotics' largest scaled customer, with Dyna robots folding thousands of napkins per day across many restaurant locations in the United States.
Latent.Space · Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Report · MoneyWhen Dyna wants to deploy their robots and platform to a new customer, they ask what technical problems they have to solve: generalization, whole body control, how to adapt quickly, how to make robots follow language, and dexterity.
Latent.Space · Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Report · MoneyJason Ma reports that for workflows where high-fidelity data can be collected today, there is no fundamental bottleneck — Dyna demonstrated a physical agent completing the entire laundry workflow as a human would. The laundry demo takes about one hour to finish the whole group, so the longest duration a robot can do a full job is already in the hour range.
Latent.Space · Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Principle · MoneyJason Ma recommends deploying robot models on-device (on-prem) because inference latency must be very short — networked calls have no latency guarantee, and dynamic tasks can't wait. He suggests training the largest model first, then distilling into a small on-device model, the same distillation approach used in language models that is converging in robotics. Dyna also builds its own inference stack for deployed models because latency is so critical that optimizing inference speed for the specific model is essential.
Latent.Space · Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Principle · MoneyJason Ma recommends buying mature components with clear input-output (e.g., LiDAR, some cameras) instead of building in-house, and reserving in-house building for components crucial to overall system performance. Specifically, Dyna is working to own and build its own actuators, because actuator quality and design determine how well you can model the robot in simulation.
Latent.Space · Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Report · MoneyDyna's model is trained on over a million hours of human data as a starting point, then fine-tuned with a very small number of demonstrations per task. With a strong foundation model, roughly 45 minutes of mapped data was enough to generalize to pulling drinks out of a fridge, and some tasks need only tens of minutes of collection. Data collection itself involves a person operating remote control doing the task for a bit, and Dyna collects a lot of data itself with employees and robots everywhere in the office.
Numeric figures (million hours, 45 minutes) are speaker-reported.
Latent.Space · Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Principle · MoneyJason Ma recommends using flow matching instead of vanilla diffusion for real-time robot control: it is a simpler-to-tune special case of diffusion that reduces 40-100 diffusion steps to about 4, enabling real-time re-evaluation. The host adds that flow matching is especially suited for action diffusion in robotics (unlike image generation) because the robot can re-evaluate in real time rather than committing to a one-shot output. Jason Ma also notes that tuning the diffusion scheduler is one of the hardest practical parts of using diffusion models, and flow matching removes most of this tuning burden.
Latent.Space · Dyna-2: General Purpose Physical Agents with Hour-Long Autonomy — Jason Ma, Dyna Robotics
Principle · MoneyKeep club qualifier announcements short and rapid-fire: one slide per winner with their stats, train executives to move through quickly, never let it run two hours. One speaker's closing night party spent ~30 minutes per qualifier and non-qualifiers became visibly done and embarrassed. At Carta, two senior revenue leaders alternated naming one person at a time — name, run up, cheer, get off — with each winner's numbers on a slide, moving ~40 qualifiers through quickly. Never hold rep-of-the-year style awards at the closing night party — they become a big time suck in front of everyone; instead run a dedicated awards night or announce them in team-specific breakouts.
Includes a cautionary anecdote and a positive example.
30 Minutes to President's Club | No-Nonsense Sales · #612: Sales Kickoff Masterclass: The Ultimate Planning Guide
Report · MoneyThe communication cadence is: six months out, two months out, six weeks, four weeks, and then weekly until one week out, and then every other day.
guest's stated cadence
30 Minutes to President's Club | No-Nonsense Sales · #612: Sales Kickoff Masterclass: The Ultimate Planning Guide
Principle · MoneyFor an investor/board panel at kickoff, have them discuss their vision, what they're feeling, and what they care about — it gives reps rare interaction with heavily involved but rarely seen stakeholders. To secure their participation, partner with your CEO and CRO (who can get them there) and their EAs, who can help prepare them across multiple steps. Put the CEO as the anchor speaker at the end of the kickoff day — even a tired rep will want to hear the CEO and board speak. A vision/inspiration speech from leadership makes employees feel they made a good decision and want to keep working there for years.
Presented as recommended design from experience.
30 Minutes to President's Club | No-Nonsense Sales · #612: Sales Kickoff Masterclass: The Ultimate Planning Guide
Principle · MoneyTell customer panelists to be upfront and not sugarcoat their experience — ask hard pointed questions like 'how could this experience have been better for you?' so the team learns from honest feedback. Structure panels around two question themes: their experience with your company (what's going well, what's not) and their day-to-day world (what they must conquer that no one understands). Customer panels also ground the team back in reality — it's easy to get caught up in your team's messaging and hear your product entirely differently than customers do.
Presented as practiced guidance.
30 Minutes to President's Club | No-Nonsense Sales · #612: Sales Kickoff Masterclass: The Ultimate Planning Guide
Creator perspective · MoneyRequire executives to have their decks done in draft form by the first week of December; executives never finish by the December 1st deadline regardless of the expectation set. Use a mid-December check-in telling executives the production company needs their deck because everything is locking down — which is negotiation theatrics, not a real deadline. Leave executives alone through the Christmas holiday, then resume contact on December 27th to close out deck completion.
30 Minutes to President's Club | No-Nonsense Sales · #612: Sales Kickoff Masterclass: The Ultimate Planning Guide
Creator perspective · MoneyPlanning kickoff early saves a lot of money because rushing in October means facing lead times of up to 12 weeks for items to ship (swag, notebooks, etc.); having the kickoff theme and materials decided early saves the company money.
30 Minutes to President's Club | No-Nonsense Sales · #612: Sales Kickoff Masterclass: The Ultimate Planning Guide
Principle · MoneyAlways walk the venue before signing a contract — one speaker signed twice without walking it and both were atrocious, including a 'missing second floor' where breakouts were unreachable except via one hidden elevator. Book your kickoff venue 9-12 months out — that's the sweet spot; you don't always need a full year and prices are often better closer to the event date.
The booking window is stated as learned guidance.
30 Minutes to President's Club | No-Nonsense Sales · #612: Sales Kickoff Masterclass: The Ultimate Planning Guide
Report · MoneyIn observational data, high AI adoption teams merged 98% more pull requests, but pull request review time increased by 91% and the average pull request became 154% larger.
The data is observational telemetry, not a controlled labor study, and does not prove that AI increased headcount.
AI Engineer · What Technical Due Diligence Finds When the Deck Says “70% Automated” — Soukaïna Abelhad
Creator perspective · MoneyFor physical point of sale they went from 40 raw attributes to 87 engineered features (TPR moved from 35.5 to 36.6 after selection and leakage checks); for online, nearly 200 engineered features picked up only about one point of TPR. Two to five times the feature surface yielding one point is a ceiling, not a tuning failure: every feature engineered from the row and the card's own history just projects information already had, because tabular features see one row at a time.
Figures are speaker-reported; some carry a possible numeric error flag.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Report · MoneyThe GNN-plus-XGBoost system achieved a 27–33% relative lift in true positive rate over a two-XGBoost baseline, using the same alert rate, evaluation window, and features.
publisher-reported result
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Creator perspective · MoneyOffline serving uses a memory-bound, GPU-sharded, layer-wise graph forward pass over the entire graph, batched by destination node — numerically identical to a full graph forward, just bounded to a few gigabytes at a time across multiple GPUs. Every user, merchant, and recent transaction gets its per-layer representation cast in a low-latency embedding store. At request time, a lightweight assembly step reconstructs neighbor states straight from the cache — no graph build, no sampling — with neighbor states re-sensitized by an exponential decay that reproduces training exactly. The real-time scoring service runs only the new transaction's owned-layer forward from cached states, plus the XGBoost head. Nothing that scales with graph size runs per request.
Described by the speakers as their implementation.
AI Engineer · Judge a Transaction by the Company It Keeps — Somya Rai & MD Sharique, EXL
Principle · MindThe host says dance is so important: get comfortable being weird with your body, including making noises. On the best mornings, he picks a song he likes, goes into a private room, and just be weird with his own body — moving in a way that feels good. He claims dancing is better than traditional weightlifting for feeling good and better than traditional stretching for actually loosening up your body, and that dancing to music you enjoy is one of the most joyful things you can do, opening the possibility of joy throughout the rest of your day and your life. He acknowledges that at first it can feel strange to let your body move to the music because of times you've been judged and the muscular armor you've developed.
Charisma on Command · I’m 39. Here’s What I Wish I Knew at 30