Approach AI adoption from impact, not AI-for-AI's-sake
Start from the outcome you want and how you'll measure success, rather than adopting AI because it can do something. The specific tool matters less than the result.
Explore
34,921 results · Most relevant
Worth exploring
WHY · 34197 cleared results from 497 sources on record
Measure token spend first — it's the highest-ROI move
If an organization isn't measuring token spend, start with measurement before anything else, because the biggest ROI comes simply from measuring. Once spend is visible, you can right-size which teams and workloads spend on what, and the data informs workload decisions.
AI incidents are infrastructure failures, not model failures
The controls that could have stopped all three incidents discussed are probably already inside the organization, just waiting to be configured. Many controls are already owned today but haven't been set up for the AI world. The audience is urged to take protective steps right now, as soon as this week, rather than keeping an eye out down the line.
Boundaries and love can coexist
The host urges listeners not to let their 'no' and their 'I love you' stand in opposition; both are necessary for a healthy relationship. He recommends practicing the phrases 'I love you, no' or 'No, and I love you' out loud to internalize this, and shares that not wanting to cut off love in order to draw a boundary kept him in relationships that didn't feel good for an incredibly long time. He also recounts firing someone after many violations and how starting the call with 'I love you. This is so hard.' made it flow much easier.
Make any conversation topic about the two of you together
The creator's highest-rated technique: whatever topic comes up, find a playful way to make it about you and the woman together, because men typically fail by never doing this.
Stand out in a world that wants to blend in: be really fucking good and be fully you
The way to stand out in a world that wants to blend in is to be really fucking good and be fully you. The meta of building a personal brand is just being you — wear what you want to wear and change it when you want to change it. To have X factor, you have to be able to put other things in the bucket — add something spice and everything nice beyond your stock traits.
Define and codify code quality standards before building agentic workflows
Nnenna Ndukwe recommends defining what code quality means to you and your team before building agentic workflows, then codifying those standards so AI agents can extract and adhere to them, and making them distributable across CLI, IDE, and CI workflows. The same standards should be checked against across planning, code generation, and code review phases.
Strongly endorsed by the host; no stated uncertainty.
Design with threats in mind from the start, including when prompting AI tools
Security measures should be put in from the beginning of app creation, and designing with threats in mind during the prompt phase makes it easier to build an app that is harder to hack. When creating something with AI, also instruct the tool to design a testing process and a security review, not just the app.
Reframing agent ROI: what did the agent take off your plate?
When engaging business stakeholders like a CFO, the very first question they ask is ROI. A common fear when deploying agents is that the only measurable thing about an agent is whose job it takes away from, and that fear survives because we mostly haven't measured the alternatives. The reframe: ask what did the agent take off your plate, and what did that freed time allow your business to do? A good intern lightens your workload: routine work moves off your desk and you get sharper at the work that's left. Apply the same test to an agent. When an agent handles routine tasks (like balance checks at 7 and 9), the banker spends the next morning with the customer who's rethinking their mortgage. The agent made the banker better at banking — that is the column that belongs on your scorecard. So
If you're not collecting tracing, nothing else works
Fuad Ali emphasizes that if you're not collecting tracing, nothing else works — getting tracing started is step one.
Tool adoption is the gate on every capability from better tools
Measure your adoption of the tools, because getting adoption numbers up is what gets you the last part of the way; every capability you can get from better tools is gated by the tool adoption number. An elegant tool that gets reached for a third of the time is just a third of a tool.
endorsed
Run evals multiple times — determinism matters more than per-shot accuracy
It's quite important to run evals multiple times because it's not about per-shot accuracy, it's about your determinism.
endorsed
Read your traces; the scoreboard can bury bigger problems
Go into your traces (e.g., LogPire or wherever you send your traces) and actually read them. If you just look at your scoreboard, you might get answers that look directionally correct but are burying underlying issues that are much larger problems inside your architecture.
endorsed
Output guardrails cost money per request; stop bad prompts at the input instead
Output guardrails cost money per request, even when the output is redacted, so consider stopping bad prompts at the input instead. Accepting a prompt, processing it, and catching bad output on the back end costs money you probably didn't have to spend. If a hacker discovers you only sanitize output on the back end, they will likely DDoS you to run up your costs. The speaker states that, legally, he would never actually DDoS anyone — the DDoS scenario is hypothetical. With the input guardrail on, the jailbreak prompt was blocked before reaching the LLM, and it was fast.
Don't rely on the model itself to not do the thing you don't want it to do
Don't rely on the model itself to not do the thing you don't want it to do. Today's LLMs are far more probabilistic — that's their whole job — which is why you can't rely on model behavior alone.
Teams not in high-risk industries can start with tracing and monitoring and iterate fast; offline work comes with maturity or high risk
Lotte Verheyden notes that teams not in high-risk industries can start with just tracing and monitoring, make simple fixes based on what they find, deploy to production, and iterate fast; the offline part (data sets, experiments, evaluation) makes sense once the product or setup matures, or in a high-risk environment where any mistake has bad consequences.
Benchmarking methodology: uniform conditions, multiple runs, and change-as-new-benchmark
Use uniform conditions across every evaluation run so results are comparable. Changing any single variable in an evaluation setup makes it a new benchmark. Always do multiple runs when benchmarking — a single run is benchmark lottery: a random result and just a ballpark number.
Variance between best and worst runs signals how much to trust a model
If the best and worst run scores are very close, there is not high variance and you can expect the same score. If they are far apart (e.g., 79% to 60%), there is a lot of randomness involved and you cannot really trust the model to execute the test even with a high score. A very high consistency score means the model is very reliable because it can be expected to solve the task all the time.
A cheaper, faster model can end up slower and more expensive due to token usage
It can happen that a cheaper and faster model takes longer and costs more than a bigger model because it just uses so many tokens. Before choosing a fast and cheap model, compare whether it is really faster and cheaper, because all those tokens also take time. Include cost as a metric when evaluating models.
Humans need a compressed review surface
When agents can generate thousands of lines of implementation detail, humans need a compressed review surface. A good diagram lets us see the shape of the system in a way code often obscures. Use diagrams not as decorative documentation after the fact, but as a way of making systems legible.
Protocols you can run
WHY · 59 of 59 published protocols
Adversarial Generators and Failure Analysis
Use saved training trajectories and a controlled intervention to explain a failure, separating equilibrium theory from observed empirical stability.
Backpropagation Without Autograd
Demonstrate all 3 repaired bugs and defend why a passed gradient check is necessary but not sufficient for generalization.
Bayesian Inference and Graphical Models
Use the tractable fixture and posterior predictive results to separate approximation, sampling and model errors; justify one revised modeling assumption.
Source analyses
WHY · 665 of 665 current source analyses
Choose sources → to get a library of your own and a daily brief.