Creator perspective · MoneySomorjai describes the calculations in Salesforce's M&A playbook: whether they can deliver something on their own much faster than the six to nine months a typical acquisition takes, and whether the product team has the internal expertise or can easily hire it — noting AI engineers are very hard to hire right now.
The a16z Show · Inside Salesforce’s $100 Billion M&A Playbook | John Somorjai
Creator perspective · MoneyPhilipp Schmid argues that a co-worker, whether human or AI, needs a brain to do work, a computer to save artifacts and iterate, a playbook because they are new to your environment and domain, tools to do their job, credentials, house rules on what they should and should not do and where they need approvals, and schedules for working hours.
hedged
AI Engineer · 7 Things Your AI Coworker Needs - Philipp Schmid, Google DeepMind
Creator perspectiveThe speaker lays out four outcomes when asking atom questions with and without the skill: (1) answered both ways — the skill has no impact, the knowledge was already in the model; (2) answered without the skill but degraded with it — the skill hurts that aspect; (3) unanswerable without the skill but answerable with it — the ideal case of teaching something new; (4) unanswerable either way — the skill may be written uncleanly even if the topic is relevant.
hedged
AI Engineer · Your SKILL.md Is 70% Stuff the Model Already Knows - Nikolay Petrov, Intel
Report · MoneyThe enablement capacity framework evaluates three things: cognitive load (how much of a rep's brain power a change takes), learning level, and capacity (how much total enablement time the org can absorb per quarter).
30 Minutes to President's Club | No-Nonsense Sales · #623: My Sales Enablement System that 3x'd My Team's Quota Attainment
Report · MoneyRule 1 of RDMs: the context stays outside the model, so the code can cut it into slices. Rule 2: the decision can choose to call another decision, up to a maximum depth (a recursive tree). Rule 3: it must work with no generative model at all — the speaker says this is very important to him, and that a recursive decision model must also compose with LLMs and with agents. In the Doom example, the full state of the game is outside the model, with each sub-harness (first-person shooting, map-exploring bird's-eye view) having its own specific subset view of that state.
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Principle · MoneyRaymond Weitekamp states three rules for recursive decision models: context stays outside the model so code can slice it; a decision can call another decision up to a maximum depth; and the system must work with no generative model at all.
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Creator perspective · MoneyFor a decision model and a language model to talk to each other they need a shared interface. LLMs in their raw form emit an open-ended stream of tokens with no structure, so the map provides that structure: a set of pointers, each with a name, a simple description, and a function underneath. An entry like 'TV on' may flip a boolean or hide a much more complicated function.
hedged
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Creator perspective · MoneyIn audio interaction, decisions are made before the speaker has even finished the sentence. The arbiter decides, given information across all channels, which decision to allow through. Each parallel lane may vote for a certain action, and the arbiter's job is to figure out where it has the confidence to allow that action to change the state. If two models making the same kind of decision both act, one may change the state before the second makes a different decision, causing collisions — so the arbiter is a gate for what can actually act on the system.
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Creator perspective · MoneyThe key insight from the TypeSafe smart home demo: decision models handle what they can, and anything that cannot be solved purely with a decision is escalated to an LLM to do generative work.
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Creator perspective · MoneyA decision-first agent uses a recursive decision model to decide when to trust a smaller supervised model that is learning on the fly, and once trust is built it farms decisions out to the fastest, smallest available model — the arbiter pattern.
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Creator perspective · MoneyThe decision model decides from the map which things are the right thing to do, while the code underneath implements the function. The LLM has a structured interface to add, edit, and modify the map, and can learn from new user behavior — e.g., learning that 'movie time' means turning on the TV and turning off the lights. The map itself can be multi-dimensional, essentially representing embeddings.
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Creator perspective · MoneyThe arbiter is a primitive included in recursive decision models to manage parallel decision execution. Because decision models are fast and cheap — runnable on a phone or a 10-year-old laptop — you can run many at once and speculatively run a bunch in parallel when uncertain. But if all of them act on the state directly you get many collisions and conflicts; like a CPU, many things can read in parallel, but you need a lock when writing to avoid colliding writers.
AI Engineer · Your Agent Doesn't Need an LLM for Every Decision - Raymond Weitekamp, OpenProse
Principle · MoneyVercel's model for all internal agent development is a tripod: a GTM engineer, a data scientist, and the best subject matter expert in the function, working shoulder to shoulder.
The Official SaaStr Podcast: SaaS | Founders | Investors · SaaStr 885: Going from 10 to 1 SDRs. How We Deployed AI Across our Sales Org with Vercel's COO
Report · MoneyJason Cohen describes the profit whale concept: compute profit per customer using gross margin (including delivery costs like servers, space, tokens, and tech support), line up all customers from most to least profitable, and add up the cumulative curve. The least profitable customers often lose money — e.g., heavy tech support users paying $5/month. The pattern is universal: a very small number of customers represent almost all your profit and almost all the destruction of your profit.
Startups For the Rest of Us · Episode 857 | Finding Hidden Multipliers in Your Business (with Jason Cohen)
Principle · MoneyCustomers are nervous about ingress but less nervous about egress, so the system is designed so that only the customer makes outward connections — Factory never makes an inbound connection. The agent sandbox machine establishes an open WebSocket connection to the relay in the control plane rather than the vendor connecting inward.
AI Engineer · Your AI Agents Shouldn't Care Where They Run - Nikhil Harithas, Factory
Creator perspective · MoneyIn the revenue velocity model, Katie looks at the number of opportunities coming in, ACV (how large opportunities are), win rate, and what the sales cycle looks like, including conversion rates in between stages. Baseline data is pulled over a six to eight quarter time frame to understand what's been going on.
30 Minutes to President's Club | No-Nonsense Sales · #622: Sales Enablement Playbook: How To Win Your First 90 Days
Creator perspective · MoneyKatie Givens uses the revenue velocity model — credited to Jonathan K ('Coach K'), an enablement leader, whose framework she adapted by mapping it to sales skills — to pair qualitative listening data with quantitative metrics: number of opportunities, ACV, win rate, sales cycle length, and conversion rates between stages.
30 Minutes to President's Club | No-Nonsense Sales · #622: Sales Enablement Playbook: How To Win Your First 90 Days
Creator perspective · MoneyMark Kosiglo, CRO at DeCebo who led Outreach from zero to 250 million, taught Katie that to understand a metric you need to understand two things: context and trend. Context means knowing what the metric was and what the benchmark or goal is — 10 booked meetings is great for one rep but terrible for 30 reps with a goal of 90. Trend means comparing the metric over time — beating a goal of 5 with 10 looks good until you see the trend was 20, 30, 26 meetings over the last 12 weeks, meaning performance is declining. You need both to know if a metric is good or bad.
30 Minutes to President's Club | No-Nonsense Sales · #622: Sales Enablement Playbook: How To Win Your First 90 Days
Creator perspectiveAlex Cooper clarifies that normal negative feelings about motherhood are distinct from pre or postpartum depression, and that pre and postpartum depression is not something to feel shame over or be afraid to discuss publicly.
Call Her Daddy · The Unknowns of Motherhood
Principle · MoneyAgent traces alone miss side effects, so the capturing mechanism needed to include process creation events, file system events, and network packet capturing — because an npm install post-installation script can invoke a new process that reads and writes to a file and connects to an endpoint, which is exactly how infostealers work.
AI Engineer · You Can't Tell What an Agent Skill Will Do by Reading It - Roey Ben Chaim, Zenity