37,617 results · Most relevant
The speaker summarizes with five questions to ask when building AI features: are people using it or just saying they will use it; do you read real outputs or just look at the averages; does it still pay off after factoring in all the costs; is a win on the best model something you can afford; and are you clear legally on which markets you can ship in. The question is no longer just whether the feature won the A/B test.
hedged
Zach Lloyd says the factory should have a strict definition — workflows, skills, MCPs, agent definitions, automations, benchmarks, webhooks, and where agents run — stored in a code repo. Without this, you lose the ability to see the full factory definition at any point in time, version control its parts, test different versions, and let coding agents automate updates to it.
Ryan Carson describes his rule that humans never go into builder agent sessions, never talk to them, and never manage them directly — everything goes through the manager agent.
Erina Karati describes the optimization loop as a ratchet: try a change, score it, and keep it only if the scorecard improves and guardrails hold. This makes every accepted change a strict improvement under measurement.
The combination of deterministic storage and pointer-based retrieval ensures the retrieved information actually came from the source it claims to come from. Use hashes as source IDs in your own system to ensure you're pulling back facts that actually exist. Use systems like code that write individual pointers so the model can only read those pointers and grab information back if they exist in the database. Preprocess information and store it deterministically as a verification method.
Operate under a contract where citations are either verified, corrected at search time, or flat-out refused.
Yi Chen describes guided experience philosophy as providing curated steps for users so they just make choices and don't need to input manually a lot, and recommends breaking a long form into simpler focused steps, with each page having one focus and one detailed question.
Rob Snyder endorses the 'solve your own problem' heuristic, but with a key condition: it works when you found something you would have paid somebody to solve and nobody was solving it. It falls flat when the problem is something you merely think should exist but wouldn't have paid for yourself.
None stated.
For each reasoning step, the memo is attacked with a what-if test — e.g., management promised cost savings from the EBSIS deal, but what if the savings were zero? Would the call change, and does the memo explain this with clear reasoning about the moving parts? A memo whose call depends on unexplained management promises is indefensible under such tests.
When a company has recently acquired and divested businesses, the analyst must rebuild the company's financials as of the relevant date — including the acquired business for the full year and excluding the divested business — to get an accurate sales figure. Using raw reported sales yields a materially wrong figure (e.g., ~$7.5B rebuilt vs. reported figures). The rebuild also involves determining what the new business really earns, netting out what was sold, and rebuilding the debt.
minor — exact reported figures are uncertain in the transcript (possible numeric error)
Voss argues that a customer saying 'I need to keep track of my orders' is not a sufficient requirement — the real job is figuring out what that means for their specific business. The only way to extract a business owner's requirements is to talk to them directly. Requirements differ completely depending on whether you run a bakery, an auto parts store, or a giant manufacturing concern.
When serving an LLM prompt, you should not re-run pre-fill on tokens that are already pre-cached. Keep cached pre-fill results hot in DRAM, a Redis cache, or an SSD using an LRU cache, as a big opportunity for savings.
Ben Horowitz advises judging what a technology is going to be, not what it currently is — but don't extrapolate its improvement trajectory without checking the underlying mechanism (3D printers don't follow Moore's law, so they never improved fast enough to build a giant company on). Evaluate a new technology by whether it's unbelievably good at ONE thing, not whether it beats the incumbent at everything — the iPhone was a bad phone and a worse computer than a PC, but it put a camera and GPS in your pocket, which enabled Uber. Evaluate technologies the same way you evaluate entrepreneurs: ask how strong the strength is and how different the technology is.
An MCP server lacking hard tenant isolation allowed users to access any other user's data, and the flaw ran unnoticed for months. The speaker recommends hard tenant isolation at the AWS-account/GCP-project level, tightly scoped workload identity, provenance tagging, and audit trails.
Andrey Rybka states that if a system combines untrusted input, sensitive data, and external action, it needs human sign-off.
A sandbox blocked web access but forgot DNS, so a hidden prompt instructed the agent to smuggle data out through DNS lookups, which were readable in DNS logs. The speaker recommends default-deny egress including DNS and placing egress controls outside the sandbox.
Figure out who your avatar is: ideally a market that is growing, in pain, has a problem to solve, can afford your stuff, and is easy to find. The 'easy to find' criterion mattered less in the past because algorithms were worse — you had to put out a beacon saying 'this is for you.' Now AI knows who content is for and serves it directly.
There are two ways to win at building credibility: accomplish something notable, or do so much work that people want to bargain on that work. Credibility can come from volume of work or volume of accomplishments — either works. If you haven't accomplished anything, show the work: do a thousand outreach attempts or knock on 10,000 doors and document it.
Farza explains he doesn't care about view counts on launch videos — he cares about actual numbers like downloads and retention. After the Clicky video, 7,000 people downloaded the product, and about 100 of them kept using it every day for five days straight. That retention — not virality — inspired him to keep going.
Because many variables are involved, evaluate the closest configuration possible that reflects the actual work and environment you will be measuring. A law firm, for example, should not look at Terminal Bench results unless it is incorporating coding agents with its developers, because the benchmark configuration must reflect the actual environment.
Choose sources → to get a library of your own and a daily brief.