35,595 results · Most relevant
Jed recommends switching the permission policy from 'always ask' to auto mode as the easiest, lowest-hanging fruit for Cloud Managed Agents integrations, letting Claude determine whether a single tool call is safe or destructive before acting.
Contested by David Pawlan on The a16z Show
Leash designs DEL libraries from 2–3 building blocks stitched together, considering shapes like forked (Y-shaped), linear, and macrocyclic geometries. Practical constraints shape the design: where the DNA tag goes and how that causes problems, commercial availability of building blocks (preferred over custom synthesis when speed matters, since custom synthesis takes longer), and the chemistry required to stitch them together. A chemical library built a certain way will be good for some protein target classes and bad for others, so plan which corners of protein space you're optimizing for. A known failure mode: fine-grained SAR mapping (e.g., moving a nitrogen around every position of a ring) is highly desirable but usually unavailable commercially. Quigley is not too worried about easy-asy
described as Leash's design practice
After basic concepts, Leash gives the model 10 binders and 10 close non-binders and asks it to explain why the binders bind. The model is then given a catalog of millions of molecules from which it picks 5-10 to test its binding hypothesis, gets answers back instantly like a coding sandbox, and updates its hypothesis. The goal is to find the best binder in as few steps as possible, like a medicinal chemist would.
Randy Hunt recommends building design teams where designers are highly technical and ship code in the production application — a 'software design team.' He says it's hard for him to imagine building a software product any other way, and he built such a team at Etsy in 2010, long before coding agents made it trendy.
One of Replit's actual shipped fixes: automatic separation of development and production databases so agents cannot touch production.
Paden Gayle recommends finding the consequential action, naming its owner, and deciding what happens when the system is wrong.
Paden Gayle notes that data owners ask who agreed to this use of documents, and warns that using client documents for AI processing without the data owner's agreement creates unauthorized data use.
Paden Gayle recommends finding a person who can tell you which review applies to the intended use, and warns that designing a system without confirming which reviews apply creates deployment risk. Applicable reviews depend on the institution and use case, so confirm them with the people who own those decisions.
Paden Gayle recommends confirming that existing controls really fit the intended use, and building the smallest useful version (MVP) with a credible path of approval; skipping the control check creates compliance gaps.
Paden Gayle warns that a review button only helps if the person can actually check the work; building a system where the analyst cannot verify the output makes the review button useless. A reviewer should be able to inspect the supporting record showing what was used to produce a result.
Paden Gayle, voicing the risk team's perspective, warns that if the model's output reaches the next system without a named reviewer, wrong exceptions flow downstream unreviewed.
Paden Gayle recommends asking who checks the content when the system drafts a response a customer will see, and warns that deploying a system that drafts customer-visible responses without a content reviewer creates unreviewed customer communication.
Paden Gayle recommends asking who owns the action, what happens if it sends something wrong, and how to stop and recover it when the system sends a response automatically; without these, uncontrolled operational risk results.
Paden Gayle recommends drawing the actual dependency chain — including the model provider, host arrangements, and your role in the operating service — and warns that deploying AI without doing so creates unmanaged third-party risk.
Paden Gayle recommends collecting evidence for the next authority decision by asking: are the reviews overriding the drafts, which failures keep recurring, and can people recover when something breaks?
Paden Gayle recommends asking which input was used, which configuration produced the output, what the reviewer saw, and who approved the action; failing to capture these records from the beginning makes it impossible to reconstruct consequential actions later.
Paden Gayle warns that when an AI system reads from a second system, a new data path exists and somebody needs to approve it; deploying without that approval creates unapproved data access.
Dr. Andy Galpin flags a common mistake: grossly under-appreciating that the legs are not a single muscle group — they contain a whole bunch of muscle groups. A classic flawed split is four days dedicated to upper body and only one day for 'legs,' which creates an imbalance over time with far more upper than lower body work. If doing body part splits, include the individual leg muscle groups rather than chunking everything in as 'legs' once a week. Training legs only once a week is acceptable, but that day has to be very challenging and include quite a bit of volume — and even then, you are almost surely not hitting the total weekly volume needed to optimize muscle growth.
Jason Ma recommends feeding a robotics world model three input classes: camera video/images from the past several seconds, the robot's own joint encoder angles, and a language instruction specifying the task. The model should output future video (or a representation of it) before outputting actions, using the predicted future to guide robot actions — analogous to chain-of-thought reasoning in visual/physics space. This is the key difference between world action model outputs and other robot model outputs.
Mohamed Sharique's graph construction: the transaction is the central node with typed edges out to its user and its merchant, plus a 5-hop follow chain back to its own recent history. GraphSAGE embeds a fixed-length vector over two hops.
Choose sources → to get a library of your own and a daily brief.