Flag interruptions, transcription drift, accuracy, and task success as audio-specific eval metrics
The source recommends flagging interruptions, accuracy, transcription drift, and task success as audio-specific eval metrics.
36,915 results · Most relevant
The source recommends flagging interruptions, accuracy, transcription drift, and task success as audio-specific eval metrics.
The source recommends creating audio evals via natural language prompting with Claude, specifying sentiment analysis with GPT audio, and catching frustrated user tone on audio spans as an eval target.
The source recommends viewing the entire trace tree for everything that happened in a session rather than disparate events from distant systems, including session lifecycle, audio input, transcript, conversation items, output, response lifecycle, and token counts.
The source recommends viewing tool call execution in line with the rest of the audio to debug tool failures, and that when a voice agent calls a tool, view the tool call in line with the audio.
The source recommends replaying the whole conversation using the session ID to stitch every turn into a timeline, and to see who spoke when, when the agent paused, and where the user got cut off.
Voss recommends taking a problematic trace and adding it to a dataset so you can run evals on it, and creating an evaluator based on the problem that Signal has found.
Jesse recommends finding a place in your daily diet for a high-quality protein shake as one of the easiest, most digestible, most convenient ways to meet daily protein goals. The first thing he does when he gets home is make a protein shake. Where you reach for your protein is up to you, and you should choose between vegan and regular whey formulas to find the one that's right for you. He adds that the underlying advice stands on its own merit despite the shameless plug behind him.
Rather than fighting and escalating, respond non-defensively. Assume positive intent once, then respond non-defensively; if someone is deliberately rude they won't stop after one comment, and continuing to pretend you don't see repeated disrespect gets you walked over.
Think of menin inhibitor therapies for KMT2A-rearranged leukemias or NPM1-mutant leukemias. Dr Issa hopes these therapies will be broadly available soon for all patients who could benefit.
Consequential creator advice requires independent verification before Today can promote it.
To verify token savings, open the data pack and inspect stored conversations with embedded token-use traces down to the individual tool call, proving whether the agent burned fewer tokens.
endorsed
When a champion says a stakeholder will only join later, propose meeting them immediately to avoid last-minute blockers. If the champion refuses to introduce you to a key stakeholder, challenge them by pointing out the risk of that stakeholder blocking the deal later. Frame why access is in their best interest; if they still refuse, fall on your sword ('maybe we did something wrong' / 'this is going to cause risk in this project') and pipe-gen back into the org.
Ask to connect your CEO with the prospect's CEO rather than requesting direct access to a CXO, because executives are more likely to accept a meeting with a peer. Their CEO will not meet with an enterprise AE no matter how good you are. Send the ghost-written message via the executive assistant to set up the C-suite connection.
Ask your champion whether the executive sponsor knows about the evaluation, and whether a key influencer is actually informing the executive sponsor. If the executive doesn't know, that's a red flag the project is not a top-three priority. Laura's example: the champion revealed Scott has no idea the evaluation is happening because Pina doesn't tell him until a technology decision is made.
When you suspect a blocker, name their behavior to the champion and ask whether the person is doing their job or pushing for a competitor. Then ask the champion what they think you should do, because only insiders understand the politics. Also ask how the competitor is being positioned internally.
Prioritize AI agent features that perform invisible, proactive cost-saving workflows (e.g., HSA reimbursements, flight price drops) over general productivity efficiency tools.
high
AI labs and security teams often publish breach post-mortems that resemble selective memory or legal defense documents rather than technical root-cause analyses. They should implement structured vulnerability reporting protocols with clear obligations, including all raw data (e.g., Slack messages), rather than relying on incomplete reports.
Andrew Chen recommends that when there are two co-founders, they should split roles: one business co-founder focuses on interviewing customers and making deals, while the other technology co-founder focuses on building the product using AI coding tools. This splits outward and inward work effectively.
Build an AI-powered dashboard that pulls insights from all customer-facing interactions (Salesforce, Intercom, Momentum, Talkdesk) to identify the biggest opportunities for product improvement. This automates the process of gathering product signal from customer support, sales, and success teams, saving hours of manual work.
Create an internal AI agent (like 'Owen') that automates 90% of builder coordination work by listening to GitHub, Slack, Notion, Linear, and Google Meets. This frees top individual contributors from the burden of meetings, status updates, and alignment tasks, allowing them to focus on building.
Build an AI-powered pre-call research system that automatically researches every prospect lead the moment it is submitted. It pulls in competitive reports, the nearest successful customer for social proof, and estimated gross payments volume. This saves reps 20-30 minutes per demo and can increase call volume by over 90%.
Choose sources → to get a library of your own and a daily brief.