You cannot excel simultaneously at running, squatting, gymnastics/parkour, sprinting, and being both jacked and lean — there are limits to the cake. The more you diversify your training goals, the worse you will be at each one individually. When choosing hybrid training goals, pick only two from a list such as: pull-up rep out, 100 cal test, back squat, 5k, vert, bench, deadlift, 60 minute run.
Guest's endorsed position
The guest argues that getting kids to front squat, back squat, clean, snatch, deadlift, walking lunge, and run sprints is the core of good coaching. A young coach who spends five years on trendy brain-linking warm-ups instead will realize they would have been a far better coach without them. He suggests girls softball teams push their front squat toward 1.5x bodyweight, clean more, and increase power output for better bat exit velocity rather than buying into trendy warm-ups.
Guest's endorsed position; the 1.5x bodyweight figure is flagged as a possible numeric error
Find an approach to continuously mine for signals where AI agents are going wrong, so failures don't happen again. Manual log review is not feasible at scale with increasing automation and longer traces, so scalable approaches like LLM-as-a-judge are needed, with occasional manual inspection of eval outputs.
Don't fully automate an agent with no human oversight and ask it to 'fix everything' — that produces slop, low quality, and unsatisfied customers. But also don't go the full manual way: don't instrument everything yourself or try to catch all failure modes and review all data yourself, since that can be automated. Find the ideal balance: bring your taste and domain experts in to build datasets and curate the problem and direction, work with your coding agent to automate instrumenting, failure-mode finding, and evals (e.g., 20 experiments instead of one), and gradually automate the eval loop over time. Avoid ending up in the bottom-left corner (full manual); be the tasteful AI engineer.
Use experiments to test what happens when you change the model, system prompt, or agent tools/data, and prototype changes before promoting to production. Pushing changes because 'it looks kind of good' — running on vibes — is the common but wrong approach, and merging into main or prod without passing QA is not a best practice.
Lincoln Murr recommends that websites gate their content behind very small microtransactions charged specifically to agents — e.g. one ten-thousandth of a cent per read — and points to AWS and Cloudflare (which he says control roughly 40% of internet traffic combined) as evidence that microtransaction gating via X402 is already being enabled.
traffic share figure flagged as possible numeric error; publisher-reported
Websites can gate their content behind a very small microtransaction specifically for agents. As an example, a Substack publisher could charge an agent one ten-thousandth of a cent to read their content, which is trivial to the user but a meaningful revenue driver. Agent microtransaction payments should ultimately be fully abstracted away from the user, in the same way AI subscriptions are handled.
A method for understanding agent behavior: manually review production traces, filter the sample to specific cases (e.g., disputes, credit card topics), and accept that reviewing thousands of traces is infeasible so the process must be simplified.
A method for tuning LLM-as-a-judge: judge a representative sample yourself, run the LLM judge, and optimize for coherence with human raters; start loose with a few too many false positives on production data and tighten over time; check the judge periodically to keep it representative.
It would be an extremely positive development if the labs tried their best models on basic cryptographic assumptions and the models failed to break them, and the labs disclosed that failure. Labs could spend $20-50 million running their most cyber-capable unrestricted models to try to break cryptographic systems. Another thing the labs could try is using these models to give much stronger evidence that cryptographic problems are hard. These are pure math questions that frontier labs could be turning their models on.
The $20-50 million figure is a hypothetical estimate.
Armstrong recommends giving an AI agent its own financial account by pasting a single prompt command into the agent; Coinbase provides tools for this with no KYC process.
Everyday users should do nothing about AI cryptography risk right now — don't panic, continue on as normal. Keep using current cryptographic systems unchanged; nothing has changed until new information about standard crypto emerges. The worst thing you can do right now is take action.
They initially oversized the decoder, which allowed it to play the game by itself even with zeroed-out spikes if it had enough biases. The fix was to massively undersize the decoder, run ablations to verify the spikes carried game information, and set the linear decoder's bias to zero, which stopped overfitting and enabled progress.
anecdotal account from the guest
In a low-stakes environment building an internal chat agent, you can start with tracing and some monitors and go to production directly without heavy offline evals. In low-stakes environments, build out test cases over time and add them as you progress. If you are building a customer-facing case, especially in regulated environments, you probably don't want to risk the agent doing something it's not supposed to do. In regulated industries, the foundation is always tracing: know what steps the agent is taking, what inputs and outputs it sees, and what tools are being called.
endorsed
Doneyli recommends using experiments to prototype changes — a new model, a system prompt change, or new tools/data access — before promoting to production. He discourages 'just running on vibes': pushing changes to prod because 'it looks kind of good' is what everyone has been doing, and it's not systematic.
A method for AI-assisted agent development: work with coding agents to automate instrumenting, finding failure modes, running evals, and generating many experiment candidates (e.g., 20 instead of one), but with structured evals rather than open-ended 'fix everything' prompts — and install the Langfuse skill (markdown agent skill files) as the starting point.
Annabell recommends that for some things you should get your taste in — get your domain experts in the room to build out the datasets, and to curate what actually is the problem that your application should be solving and which direction you want it to go.
Doneyli recommends keeping domain experts in the loop working with development teams to define datasets and quality criteria, because human judgment remains essential to agent evaluation.
You cannot control what your users input into the system, so you must catch PII (like credit card numbers in clear text) during instrumentation and eval building. Enable PII identification so any data deemed PII is automatically masked before it gets persisted into the observability system, and use a masking function that automatically masks any defined PII so it is not persisted in plain text. In the demo, a user inputting a credit card number in clear text shows the card redacted in the trace before persistence.
Teams shipping agents must continuously mine for signals where things are going wrong and make sure they don't happen again. Monitor aggregate metrics on what is going on and detect signals where things are going wrong. Monitoring should include finding the interesting traces you want to dig deeper into, not just aggregate metrics. Track quality signals such as compliance guardrails, truthfulness, and how many users are frustrated (e.g., writing all-caps messages).
endorsed
No published protocol matches this selection.
The creators behind most of what cleared.
323 cleared insights match
179 cleared insights match
154 cleared insights match
Choose sources → to get a library of your own and a daily brief.
