35,800 results · Most relevant
The speaker advised involving product managers, QA, and support teams in writing evals because they know what failures look like, and letting any team member read traces or annotations and contribute to criteria — evals are not just an engineering surface.
The source recommends mapping WebSocket events to an easily understandable set of semantic conventions, using Open Inference semantic conventions which are open source and free, and using OpenTelemetry-compliant semantic conventions for GenAI tracing.
The source recommends classifying user emotion from actual audio rather than transcript for sentiment analysis, and classifying tonality and sentiment from audio because the transcript cannot capture it.
The source recommends moving to agent-as-judge for complex tasks, using RBAC checks to verify the agent accessed the system it was supposed to, and hydrating evaluation with data from accessed tools and third-party systems.
Listen to the audio with the trace open alongside it to understand failure cases — reading transcripts alone is not enough.
The source recommends using one auto-instrumenter that captures everything rather than bespoke parsing per vendor for voice agent tracing.
The source recommends running evals directly against the audio with evals that are based on audio models, rather than transcript-only evals, and running evals against the actual audio spans.
Voss deliberately asked the agent a question he knew it could not answer — 'Show me all the toys you've got under $6' — because the non-pre-run app lacks price filtering, which he knew in advance. Clicking the live button in the Arize UI then shows the failing query arriving in the traces, where you can see it fail.
Charles Frye recommends that at every moment some kernel should be running on the GPU, and CPU-side scheduling work should never block GPU progress. He notes the GPU can't send the response directly to the user and wouldn't be good at running an HTTP server, so CPU work is necessary — you just want to make sure it doesn't block. He warns it is very easy to accidentally write a host-device sync that blocks GPU progress, even though CPU-GPU communication is necessary.
The creator recommends leading with the so-what rather than the how: start with the insight that matters most and why it matters, since the audience doesn't need to understand methodology before understanding why they should care.
The creator recommends giving the audience one clear thing to walk away with — one headline, not 12 findings — so stakeholders can summarize the research to others.
The creator recommends connecting each finding to something the audience already cares about — a goal, a risk, or a decision on the table — and explicitly tying findings to decisions currently on the table, since research that maps to nothing actionable is easy to ignore.
Jesse says to include lean protein like chicken, fish, or beef in the non-sweetened portion of the meal plan for muscle gains — protein is an inescapable truth when building muscle is the goal. But don't eat boiled chicken with steamed broccoli; there are many ways to enhance protein flavor without a lifetime of bland food.
Evals are a way to answer questions about your AI system: which LLM is the best choice for your needs (especially with new models dropping all the time), how the system performs across diverse real-world scenarios (e.g., English vs Japanese, Python vs TypeScript), cost efficiency (high performance without excessive costs), brand consistency (does AI reliably reflect the company's voice and standards), feedback (are we learning from users and improving iteratively), and how to know when something breaks or gets worse.
Fire underperformers who resist coaching quickly because they will scorch earth and block progress — don't wait nine months for them to leave. Trying to save everyone, including people who don't want to be saved, will cause the entire sales organization to fail. Avoid spending three to six months warming the team up; make bigger, harder decisions faster to avoid a dead sales floor.
endorsed
Get reps focused on the right accounts and personas before demanding more calls, emails, or better copy. The highest activity rep is almost never the highest performing rep; a prospecting performance issue is usually a targeting issue, not a copy issue. Skills can raise a rep by a letter grade, but the wrong cold calls on the wrong account are F cold calls no matter what.
endorsed
The core technique: map the informal, political 'shadow' org chart in your prospect's company to understand who actually has influence over key executives, not just who holds titles. Laura credits this technique with helping her reach 305% of quota and finish as the number one rep out of almost a hundred sellers.
Rob Walling recommends that aspiring founders without industry experience get a day job to gain exposure to problems. He says you will inevitably see problems at the day job and see how the company spends money — including on software you think is bad. He advises being deliberate about learning from the job: where they spend money, where the gaps are, and what problems could be solved with software. He shares his own path of learning to code nights and weekends from the library, buying books, and eventually getting a job coding.
For ADK agents, Warrick describes applying the Temporal model class to wrap the model, wrapping tools with the Temporal activity tool, setting up a worker and passing in the Google ADK plugin, and then running the agent as with any standard ADK setup.
For the LangGraph integration, Warrick describes passing Temporal into each node of the graph, setting up what the metadata looks like, using LangGraph's interrupt to call out from nodes, using the wait condition in a loop to track the interrupt signal and pause the workflow, passing the LangGraph plugin to the worker, and building out graphs for the multi-agent setup. She notes LangGraph is more than a framework but is being used as one here.
Choose sources → to get a library of your own and a daily brief.