37,539 results · Most relevant
AI safety discussions should be grounded in concrete, identified risks (e.g., novel cybersecurity threats) rather than speculative super-intelligence scenarios. Security claims should be validated using physics-based and systems-based constraints (e.g., entropy, bit rates) rather than abstract 'super-intelligence' capabilities.
The center of AI innovation has shifted from model creation to the integration of models into traditional software systems. This involves using language models to return probabilities for conditional logic (e.g., 'if X percent') rather than binary true/false checks, reviving 1960s-70s probabilistic programming research.
medium
Commercial context: The creator presents this as their own offering and directs viewers to get it.
AI agents should be treated as untrusted entities with oracle-like capabilities in threat models. Unlike human employees, they can execute thousands of security probes rapidly without fatigue, requiring granular access controls and monitoring of internal APIs for denial-of-service style attacks.
Karri Saarinen describes a feature roast practice: attendees critique the whole feature with raw, non-personal feedback ('you can be as nitpicky or confused as you want'), framed as 'This is how I see it' rather than a personal attack. The building team can interact with attendees and ask questions when they don't understand the feedback. Afterwards, the lead synthesizes the feedback, groups it, and turns it into actionable issues to fix.
A multi-step workflow: kick off a project with a Slack command, create a Hilltop PRD, have an agent break it into units of work, review and approve via Linear tickets, and the agent starts work automatically. Humans stay in the loop, and the agent reevaluates the project to keep the plan fresh.
A step-by-step method for making existing components agent-ready: audit and refine current components, start with meaningful layer names, check how components resize, and set up component properties for each component.
A weekly cold-and-heat protocol: 2-3 visits per week, each with three cold dips and two sauna sessions, starting and ending in cold water, with 1-2 minute dips at ~15°C or colder, 10-15 minute sauna sessions at ~80°C, totaling 11 minutes of cold and 57 minutes of sauna per week.
Bakken describes how he identified his golden zone: he ran a session, then another a couple of days later, measuring lactate and controlling it with heart rate across consecutive sessions; he also measured competitors and Kenyan runners to narrow down his own training zone. He recommends runners test the golden zone in their own training, ideally taking a lab step test first to establish control values, then doing a field test outdoors about a week later to compare treadmill heart rate, outdoor heart rate, and perceived effort, cross-checking lab results against field conditions close in time.
hedged
The best approach to establishing training zones is to take a lab step test first for control and guidance. About a week after the lab test, do a field test outdoors to check whether treadmill heart rate is comparable to outdoor heart rate and perceived effort. Thereafter, use lactate testing to cross-check intensity once in a while, not on every session, since testing is expensive.
hedged
In a normal winter training week for elite athletes, Bakken prescribes double threshold — two threshold sessions in one day — on Tuesdays and Thursdays, with the harder X session on Saturday where you really push. The two weekly threshold sessions are split into intervals: one short interval session (45 seconds to one minute) and one longer session (six-minute intervals), keeping total running time in threshold sessions to about 25 to 35 minutes. The X session is a hill workout in winter and more flat work in summer when preparing for track workouts.
When starting to optimize an AI system with RSI, begin with prompt updates because they are the easiest and most straightforward way to keep iterating. After prompt optimization, the next step (L1 ladder) is to optimize the structural context of the AI system. For the L2 ladder, expand scope to automatically design the entire workflow, a kind of meta information design. After workflow design, the next step is to optimize the harness code itself.
The RSI process is broken into seven steps. (1) Run the agent and generate samples: given a query, generate output and log the processes in traces (e.g., LangSmith traces) — these trajectories are what you optimize. (2) Run the traces through the evaluation engine (e.g., the LLM judge) to get judge output showing where the system is good and where it is not. (3) Agent reflection / root-causing: understand the root cause of bad output — agent, model, prompt, tool, or other issue; the root cause can even be a judge issue or rubric issue. (4) Based on the root cause, generate an edits proposal — currently just prompts, but in future a new workflow design, harness code, or even optimizer code. (5) Validate the edits, since some are good and some bad, using the same evaluation engine (samejudge
Run the agent and generate samples logged in traces; run the traces through the evaluation engine (LLM judge) to see where the system is good and not; perform agent reflection (root causing) on judge output; generate an edits proposal based on the root cause; validate the edits using the same evaluation engine (same judge, same rubrics); after the prompt is updated, run evaluation again on a holdout set to confirm the new system performs better; then make the updates the new baseline and continue iterating at a good cadence.
Cluster the failing rows from production traces by failing dimension using pure clustering without any LLM calls; within each cluster run an LLM to diagnose each row and assign a root cause plus a tag; merge near-duplicate issues because LLM-generated root causes are free tags with slightly different wording; then group and rank canonical issues by frequency to identify major bottlenecks.
Collect a bunch of production traces from your agent, get a human to annotate about 50 of those trajectories with very detailed feedback (e.g. this is a long response, this is a short response, this is a good response, this uses this terminology), use GEPA to optimize an LLM-as-a-judge prompt, then use that judge prompt to go back and optimize your agent and deploy it, forming a data flywheel for continuous improvement.
For subjective tasks that are hard to evaluate, GEPA can learn evals for the task from production traces. Collect a bunch of production traces from your agent and have a human annotate about 50 of those trajectories with very detailed feedback — specifying things like response length preferences, quality judgments, and terminology usage. Once you have human annotations, use GEPA to optimize an LLM-as-a-judge prompt. Use the optimized judge prompt to go back and optimize your agent, then deploy it, creating a data flywheel of continual improvement. This eval-learning flywheel is a successful paradigm that some leading teams in production are already using.
Erina Karati lays out an ordered recipe for long-horizon agents: (1) freeze the harness, (2) define scenarios, (3) log traces, (4) score behavior, and (5) expose only a small policy surface. The steps are presented as an ordered method for making agent behavior measurable and safely improvable.
Jesse Hall lays out a benchmarking method: take a complete working agent with a real backend, put simulated callers against it (callers that pursue goals, change their minds, and introduce background noise), define each scenario as what the caller wants, what actions the agent should take, and what outcomes should be, then score those outcomes; swap the model stack while keeping the same agent and scenarios so every model sees identical input, and run the loop again.
To pick models, test the conversation: take a complete working agent with a real backend and put simulated callers against it. Simulated callers can pursue goals, change their mind, and have background noise added to make the scenario real-world. Run text-based simulations to test business logic, and audio-based simulations to test speech recognition, turn-taking, and the entire voice pipeline. Define each scenario as: what the caller wants, what actions the agent should take, and what the outcomes should be. Score agents on outcomes — in a hotel benchmark, did the correct reservation hit the database; in a financial benchmark, did the correct authorized amount get charged. Don't score on completion alone, because agents can game completions — verify they took the right steps: check that a
When the happy path doesn't happen and a prospect gives a lukewarm feel, Rob Snyder recommends throwing the 'Hail Mary': asking 'Is there anything on your to-do list that you would pay me to do right now — that you would pay somebody to take off your plate?' He reports that quite a few startups have totally pivoted and taken off from this question.
The pivot outcomes are described anecdotally.
Choose sources → to get a library of your own and a daily brief.