35,595 results · Most relevant
Isabelle Williams describes creating a blueprint that acts as a ceiling of access for each agent type in your application. The agent token's scopes are set to the intersection of the user's permissions and the agent blueprint's ceiling scopes, and the subject of the agent token is the agent ID so actions are attributable to the actual agent acting on behalf of the user.
AuthMD is an open protocol (a markdown file you host with your application) that describes how agents can register for accounts with applications. It is meant for agents to read, not humans, and is where an agent can discover how to get started with your application, including mechanisms for getting access, creating an account, and calling your API. Your application is the one enforcing auth at your API, not AuthMD.
According to Freddy Lecue, the recipe for the transaction foundation model lies in how tokenization is done and how post-training is performed; a too-huge vocabulary makes training messy, so significant effort goes into deciding what to encode as a token, e.g. regrouping zip codes by distribution rather than one token per zip code.
When your data is much more specialized in its representation than language, build a transaction foundation model where transactions, not language, are the base. In a transaction foundation model, the model predicts the next transaction, or the next token in the transaction, using all available channels.
The missing piece in an XGBoost fraud model is the relationship between the transaction being classified and all past transactions. The relationship between a transaction and past transactions is lost when using an XGBoost model unless encoded in a very hard way, which is not easy. Transformer architectures give attention from transactions to past transactions, which is what helped push the fraud detection rate much higher.
Huberman's core methodological principle: individual ingredients are the way to figure out what works best for you; explore stimulant-based and neuromodulator-based supplements separately before combining them into a single formula, because only you can determine your best protocol.
Design a materiality scale to understand which AI use cases are riskier than others and prioritize them. Materiality directly rank-orders with respect to the harm associated with AI use cases, so more resources can be allocated to higher-materiality use cases. The scale can be company-specific, and risk ratings can change over time.
Route AI system usage based on reliability and materiality, ensuring high-risk rated systems are only used in regions where you are confident of performance. Only make the deployment decision once you are confident about the region; otherwise it may be too expensive due to downstream failure costs.
Evaluate the full AI system based on risk, not just one component. A narrow evaluation can miss failures that matter in the real world. When extending evaluation to the whole deployment system, test whether retrieval brings back the right information, what happens when a tool fails in the agentic system, whether human review catches mistakes, and whether the system can finally recover from failures.
Tanmay Sah argues that AI failures have structure and that some sub-dimensions are more critical than others: human mediation (human-in-the-loop effectiveness) and harm detection (monitoring) are strongly associated with harm, while scale and system design appear frequently but are not strongly associated with harm. He recommends prioritizing sub-dimensions by harm potential and allocating investigation resources accordingly rather than treating all equally. Deployment controls rank most critical, followed by business impact, then technical failure; securing deployment controls for high-risk systems gives assurance failures can be handled in advance, and deployment controls are far more important than other sub-dimensions.
associations as reported by the guest from their incident analysis
First step in pairing: trace each face back to its calligraphic root (broad nib vs pointed nib / expansion) and pair faces sharing the same structural tradition. High-contrast modern serifs (Bodoni style) pair well with grotesques because modern serifs structurally evolved into grotesques.
After structural tradition, compare proportions — x-height is a major factor, along with ascenders and cap height — then match stem weights and overall weight range/color/texture. This three-step filter (calligraphic root, vertical metrics, weight) narrows pairings quickly, especially when pairing functional with expressive typefaces.
The team tried to reapply the same mechanism learned on RAG to the contract project: get the answer first, then have the agent find it back (invert the problem). When you have a well-known technique for one sub-task (like document retrieval from RAG), reapply that same mechanism to a new domain-specific project. When no public benchmark exists for a domain-specific agent task, build a custom test set by inverting the problem: create the answer first, then test whether the agent can find it back. When building an agent for a task with no benchmark, identify the sub-steps of the agent and determine which ones already have known techniques versus which ones need new evaluation approaches.
Adit Abraham's top advice for evaluating document-processing tools: the number one thing to care about is the distribution of data — test on data representative of your own documents rather than trusting generic benchmarks. He also advises validating that your eval dataset actually holds before running any model-vs-model comparison.
Adit Abraham emphasizes that Reducto cares about not just accuracy but determinism and traceability — customers need to know where an answer came from. Traditional computer vision was strong at this, and word-level citations let agents reconcile conflicting documents and land on the truth.
The Cache Program teams use a 'team brain' — a common knowledge storage where engineers' product feature knowledge becomes accessible to all team members, so a newly joined engineer can tap into the same prompt context as a five-year veteran.
Chris Dreyer went from legal in general to personal injury because 70–80% of his revenue was already coming from that space — a natural push. He argues people in the agency space struggle because they go too broad — it's market fit, not product-market fit.
You cannot automate a process that does not exist. Before handing work to an agent, build a standardized process around the org function, then define the steps to hand that work off to an agent running in the background. Without standardization you also cannot build evals or know whether the agent is doing the work right.
Rob Walling describes the most common ways founders stumble onto an unsolved problem: scratching your own itch, hitting a problem at your day job, solving a problem of a spouse/relative/colleague, being on the receiving end of a poor customer experience, and finding a problem online in communities like Facebook groups, Quora, Slack groups, and support forums.
GaryVee argues that skills are easy to train but embedding drive and hunger is the hardest part, and that if employees were as ambitious as their boss they wouldn't work for them. He says you can't force a penguin to be a giraffe — you can only maximize the giraffe's penguin-ness — and that over-infatuation with turning a person into something they're not wastes time. No matter how much he trains someone, they're not going to the NBA. Instead, reverse engineer people to be the best version of themselves, push them as far as you can, and when they hit their cap, move on.
Choose sources → to get a library of your own and a daily brief.