35,595 results · Most relevant
In the eval results, vector search and agentic search achieved the same level of accuracy, but vector search cost four times more than agentic search. Scoring was binary: if the test passed the Microsoft TypeScript-Go test suite, it got 100%; if it failed, it got 0%. The team also ran multiple trials per task, a best practice for evals since LLMs are non-deterministic and it helps you trust the score you get back.
The first fix to prospecting was mapping out a 'golden path' by breaking organizations into archetypes: pre-enablement (0-250 employees), 250-500 (VP plus some managers), 500-1500 (second-line leadership and directors), and organizations with multiple sales orgs. The team highlighted in green the people to target at and above the line, with the sole enablement hire plus a VP of sales as the main target.
Laura walks through asking the champion about relationships (e.g., whether Pina and Scott get on, whether Pina knows about the project, whether she's telling Scott), revealing influence paths and that the economic buyer may be unaware of the evaluation.
Pienaar reports that Cleric added grounding into its agent in the form of a critic. With grounding, the agent's confidence-versus-accuracy line sits at 45 degrees, meaning confidence matches accuracy; without grounding, the agent becomes extremely overconfident. An agent using only vibes will always give a very confident answer, but with outcome signals, if the agent says it's 90% sure, you can trust that signal — and you can complete the loop by taking the signal from production back to your agent.
For each stated task, Composio returns not only the correct tools the agent needs but actually a plan of how to use them — for example, for Slack, you have to find the Slack channel ID before you can start querying for messages.
In the example diff, a parameter (called 'inertia check') was removed because a refactor to a dependency bump was needed. Many people using an LLM without web search will conclude the diff looks consistent and looks like a cleanup, which is not actually the case. A human reviewer would instead go to Stack Overflow, ask whether the random variable was removed, paste the error code if the code isn't compiling, get the desired response, and make a PR to the repo to resolve it.
The parameter name 'inertia check' was stated with low confidence (0.5).
Perceived delivery rate increased by ~4.5%, but the METR study found actual productivity dropped 19% while perception rose 20%, creating a 40% spread. This suggests developers feel faster but may not be generating more value.
high
Hodak describes the three elements of Science Corporation's pipeline: vision work, bio-hybrid neural interfaces, and a perfusion program (Vessel), which together form the minimum set to drive a significant revolution in medicine within 10-15 years.
Jeff Cavaliere states that Conor McGregor's knee injury was not due to lack of training or preparation, but likely due to over-enthusiasm and adrenaline from attempting a spinning roundhouse kick right at the start of the fight.
Consequential creator advice requires independent verification before Today can promote it.
The first thing to think about when building evals for high-stake applications is to define the most ambitious success. Setting this goal gives the scope of what can be designed further. Extend the what-must-happen questions into other WH questions and first principles, thinking about what life would be in AGI with a model that has no inabilities and performs like the human being emulated.
Roland Gavrilescu argues that the next stage is to productize always-on, long-running, horizon agent tasks in ways that can scale with your customers.
Aamir Shakir recommends fitting really good context so the LLM can focus on what is relevant, which means needing way less LLM cost.
When exploring traces, take inventory: what are users asking for, what did the agent do, which tools were used, what were the answers, and would you ship the answer. Think like a stakeholder (product person, engineer) and make mental notes of what you would measure before defining eval criteria.
Ramp makes it easy for companies to manage spend, but as it grew it had to interact with more hard-to-reach portals, creating the gap of how to fetch receipts efficiently at scale. Ramp initially solved this by putting together a team of humans who would log onto portals and fetch receipts manually. Jay recounts that Ramp now spins up a browser agent that goes to Delta, logs in, captures the receipt, and matches it to the transaction after a flight is booked on a Ramp card.
Sofia Jakovcevic tells a hypothetical scenario: a security engineer reviews a support agent's policy and requests a one-line confirmation-before-action addition. Asking a support agent to confirm before taking any action improved action safety by 17 points (measured one week after the change was submitted, per publisher-reported qualification). But an agent that confirms before every single action causes tasks to go uncompleted when users leave sessions mid-confirmation, so the session is never completed to the end.
Dane Knecht predicts that over the next 18 months the majority of engineering teams' time will be spent on security, citing a Cloudflare keynote stat that over 50% of engineering time will go to security. He advises planning budgets and staffing accordingly.
explicit prediction with a stated time window; keynote stat is secondhand
The speaker asked Muse to return the top three or four competitors for each customer the company had ever had historically, along with each competitor's domain name, then fed that list (company names and domains) to their outbound agents.
The Muse audit flagged things that 10K missed — noting 10K doesn't really have a browser and is stuck in Replit. Muse went through all three of the company's websites and pointed out where agents got confused on event years — some agents referenced past SaaStr Annual, some referenced the 2027 website — leading to confusion on attendance numbers and demographics. Muse recommended cleaning up the year confusion by standardizing everything to one set and republishing everything.
possible numeric error in source
Gary Vaynerchuk says the best way for a brand to start thinking of itself as a media property is simply by getting started and beginning to make content. For example, a flower shop owner should start making content literally — that is how you begin acting as a media company.
Jeff identifies the dumbbell overhead extension as one of the best tricep exercises you can do, dumbbells or not, and recommends it as the first of the two exercises.
Choose sources → to get a library of your own and a daily brief.