35,595 results · Most relevant
The first move for AI search visibility is creating high-quality transactional content on your website: directory-style roundups and versus articles. This is where to start before pursuing offsite tactics.
Ross Hudgens describes how percent visibility in GEO is measured: use prompt trackers with proxy search volume for prompts like 'what is the best X software for your industry', and track that over time. He recommends creating categories around each of your product use cases if you can, and tracking visibility over time.
The creator recommends running many agentic coding threads in parallel using coding harnesses from the labs, and argues the limiting factor is the human's brain — how much context you can keep in mind to keep all threads going. Agents also let humans work and think at a higher level, exercising more executive function.
Thor Schaeff recommends trying the Gemini Live model through Google AI Studio as the easiest way to get started: select the Gemini 3.1 Flash Live model, pick a voice such as Puck, enable grounding with Google Search for real-time information, and optionally try the Live Translate model; ai.dev/live is an alternative URL.
Fuad Ali recommends using harness-as-a-judge for access checks by connecting it to external systems so it can call out and verify whether the user was allowed to do what the trace shows.
Forrester argues you probably need both input guards and output guards from a security standpoint — together they catch 80 or 90% of what you need. Challenge five was an output guard, a general prompt-based guard: you can do it and it's worth doing, but you will pay the cost of LLM processing. Challenge six was an input guard, which can be done upstream — probably cheaper because it doesn't get processed by the LLM. Upstream filtering helps keep performance going in AI workflows, as long as you don't muddy the waters.
The 80-90% coverage figure is the speaker's rough estimate.
The presenters advise that if a tool like the search help library is not actually helping, remove it from the application, and to use the Langfuse skill with repository context to fix issues in the code — not only in the evaluator setup — so the application itself gets back to working. After finding issues in traces, fix them to improve the application over time and address the issues actually coming up.
Wolfram Ravenwolf recommends running five evaluation runs per benchmark with WolfBench (more if budget allows) and publishing all evaluation results so others can inspect them.
Treat the WolfBench leaderboard as more than a chart: load a snapshot, add your own runs, filter metrics and models, reorder or remove bars, and export. The interactive page lets you build custom reports, copy charts elsewhere, use the screenshot symbol to share results (e.g., on X), and share the URL so others see exactly the view on your screen.
Let AI agents automate your full workflow from beginning to end: observe issues, pull them down, create datasets, build fixes, and test them. The agent can pull down traces, review evaluations and annotations, do its own analysis, and come back with fix suggestions autonomously, including future suggestions like asking more questions or adding guardrails.
Duggal recommends setting up evals — LLM-as-judge, code-based evals, or some other eval setup — as the way to move beyond shipping on vibes.
Through the repo you can set up your own cluster and configure it in any way you want; the infrastructure is not strict and can be self-hosted, and with your own GPUs it is totally your own.
The belief that you don't need to hire juniors anymore and can just hire a senior person with AI has been debunked as a huge misconception; Atlassian changed its hiring pipeline to hire a lot more on the junior end and more on the senior end and less in the middle, and it has worked wonders.
Users type requests in chat that the product cannot do yet; those prompts reveal which capabilities to build as dedicated UI features and which experiences belong natively in chat versus in the app, accelerating product discovery for existing SaaS apps.
Two easy ways to integrate the app server: the recently released Python SDK, which handles lifecycle management, eases things like sign-in with ChatGPT, and gives a familiar interface; or just asking Codex to implement the protocol, largely pointing it at the documentation (sometimes nudging helps). You can also combine it with plugins like Build macOS apps to quickly spin something up.
Felipe Blanes recommends that as soon as you start getting customer signals, you make sure they feed your evals, and that your evals stay up to date and reflect exactly what your customer is trying to do right now.
Viren Baraiya says that for production deployments and Kubernetes cluster management, he wants a very well-defined workflow in terms of the sequence of steps executed when restarting a cluster. As part of this, the harness can send a Slack message to a human asking for approval before restarting a production cluster.
Do not open a store approach by commenting on her body (e.g., 'I saw your ass' through the window). Instead, open with a direct compliment like 'Hey, excuse me. I just wanted to say I actually thought you looked really cute.'
Kyle advises the audience to approach women they could go on a date with — including women who look busy, settings with many people around, and situations where they fear rejection — rather than making excuses.
Maintain a rolling frontier of bets that are deeply aligned and ready to trigger a ramp at any given moment — structurally carving out a three to six month advantage means winning all of those deployments, a perennial wedge that lets you drive in.
Choose sources → to get a library of your own and a daily brief.