Explore

Work with coding agents to improve applications, but don't fully automate or go fully manual

Annabell recommends being in favor of working with your coding agents to continuously improve your applications — but warns against two extremes. Don't just ask your coding agent to 'fix my agent, make no mistakes', which produces slop rather than anything helpful; if you fully automate agent improvement and just let the agent go for it, you will end up with a low-quality system and unsatisfied customers. But also don't go the full manual way anymore — don't instrument yourself, catch all failure modes yourself, or review all the data yourself; these can be automated.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Use shadow runs and human-in-the-loop before automating

Run the process manually in parallel with the agent (shadow runs) to compare how the agent handled it, measure accuracy, and build ground truth. Start with a fully human-in-the-loop system, then over time slice the data into parts you're confident to automate versus things where you always want a human in the loop.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

For most athletes, prioritize acceleration over max speed

Attack acceleration first when working on sprint technique, because in most sports the max speed phase won't express much difference. Unless you're a winger getting big open breaks covering 40-50 meters, max speed matters less. Getting good at acceleration is the low-hanging fruit and big bang for your buck for most athletes.

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Define hard limits for regulated spaces

In regulated spaces, define hard limits — things that definitely shouldn't happen (e.g., a financial services agent should never give investment advice) — and ingrain them in the evaluation set. Even if a violation is never fully prevented, you will at least know about it when it happens.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Proposed alternative: passive optical device for diffusion steps

The proposed alternative is an application-specific passive device where the physics of light propagation performs a step of the diffusion procedure; by iterating on the optical system architecture, one can start from a random distribution of pixels and, just by letting the physics do its work, create new samples from the desired dataset.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Two-phase athlete assessment: strength first, then technique

Assess foundational strength first as a simple ratio to bodyweight, taking into account the athlete's current level and training status, and decide whether strength needs to be a priority. If strength numbers are already adequate for the goal, move to the second phase: the technical aspects of the movement or sprint — altering patterns, bringing in accessories or derivatives, and changing how the athlete runs or moves.

Method, not a verified protocol.

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Air bike conditioning protocols: intervals, steady state, and testing

For improving conditioning on the air bike: do 10 sets of 10 calories or one set of 100 calories once or twice a week, intensely; the same workout week to week with genuine effort produces good progress without elaborate programming. For longer work, do 60-90 minutes of long slow distance keeping heart rate at 70%, or interval work of six by six minutes at 85-90% heart rate without going above that ceiling, and occasionally push past threshold into the zone where you feel like you are about to die. You can get an excellent conditioning session in under 10 minutes.

Endorsed protocols from the guest

Method, not a verified protocol.

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Commingling memory and compute: the most promising near-term direction

The most promising near-term direction in neuromorphic computing is the commingling of memory and compute, done purely digitally, where a bunch of startups are working. One approach is combining SRAM and logic. D-Matrix is a startup that has done good work here, including a strong Hot Chips presentation; its innovation is essentially co-packaging: bringing SRAM very close to compute in the memory hierarchy on the decode side to address the decode ceiling. IBM has also done work in this area. D-Matrix feels like a makeable-now digital memory-compute company: it works, it can be manufactured, and the market can take it up. However, co-packaging SRAM close to compute increases the surface area of the packaging, which means fewer servers fit in the same amount of data center space, and the do[

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Monitor security-risk flags such as potential investment advice

Doneyli recommends monitoring for security risk flags where the agent might have given investment advice to customers, and filtering traces by categorical scores (e.g., security risk = investment advice) to find specific examples where the agent was flagged. In his example, 89 cases over the last seven days involved customers asking for investment advice.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Build a stakeholder dashboard of business-benefit and failure-mode metrics

Doneyli recommends building a dashboard of metrics to report to business stakeholders based on data collected from the agent. Track the number of agent interactions that didn't need to go to a human agent as a business-benefit metric, and calculate the cost of agent-handled interactions versus human agent cost to show directional business benefit. Also pinpoint the areas where the agent is tripping up by tracking failure modes in order of occurrence, and track thumbs up / thumbs down customer feedback displayed in an easy-to-consume dashboard.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Test guardrails by impersonating client personas

Impersonate different client types (e.g., everyday customers and ultra high net worth individuals) to test whether the agent respects guardrails for those personas.

endorsed

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Curate specialized datasets: red-teaming for prompt injection and use-case-specific sets

Curate a red-teaming dataset that looks for prompt injection attacks and any attack that would allow someone to circumvent guardrails and get information they shouldn't. Also curate datasets specific to a use case — e.g., knowing which cases have been disputed and whether the dispute has not been solved.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Schedule 2-3 weekly gym sessions for strength alongside running

When combining strength training with running, program two to two-and-a-half to three gym sessions per week for deadlifting or general strength work.

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Prioritize fundamental strength work over gimmicky brain-training warmups for kids

Get kids to front squat, back squat, clean, snatch, deadlift, walking lunge, and run sprints instead of spending time on gimmicky brain-training warmups; a coach who spends five years on such gimmicks instead of fundamentals ends up worse, not better.

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Air bike benchmarks: 50-cal and 100-cal tests

The 50-calorie and 100-calorie air bike tests are useful conditioning benchmarks for field sports. Around three minutes for 100 calories indicates good conditioning depending on the sport's demands; about a minute for 50 calories means you are a beast. Nobody reaches the three-minute 100-cal mark without significant conditioning time — it cannot be faked. The 50-cal times converge more across sports than 100-cal times, and benchmarks differ by sport (a front row rugby player will not match a jiu-jitsu or judo athlete). The test is low impact the day after, so you can still squat the next day.

Publisher-reported benchmarks; sport-dependent

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

A/B skips as a proxy for sprint quality

The host uses A-skips and B-skips as a proxy drill to assess and develop sprint coordination, bounciness, and body awareness. If A-skips become bouncier and springy and B-skips become more cyclical and smooth, sprint quality will very likely improve down the line. Conversely, a B-skip that looks like the hamstring is about to tear off signals poor sprint technique. He notes A/B skips benefit athletes for roughly their first six years of training, are a great tool for teaching middle school kids rhythm and timing, and are used by sprinters like Noah Lyles for warm-ups, while acknowledging ongoing debate about their value in track and field specifically.

Host's confident heuristic; he acknowledges debate around A/B skip drills in track and field

Garage Strength · How To Get Strong and Still Be Athletic w/ Sika Strength

Oversized decoders let silicon play the game while the cells never learn — an easy way to cheat

An oversized decoder can play the game by itself — if you zero out the spikes, the decoder can still play with enough biases. If the silicon decoder overfits, the cells never learn — they can keep their stochastic behavior while the decoder plays the game for them. It is very easy to cheat performance out of these systems: massively oversize your decoder, run backpropagation on it, and suddenly you have brain cells that play anything. Startups and companies get this wrong a lot of the time. In fly-brain connectome experiments, people use massively oversized decoders, so it is never the fly doing anything, just silicon. The fix was to massively undersize the decoder, run ablations to confirm the spikes carry information, and set the linear decoder's bias to zero; undersizing gave better

Presented as the speaker's strong claim based on their experience; the fly connectome example is attributed to 'people' generally.

Y Combinator · What If We Stopped Using GPUs? | YC Paper Club

Reduce unpredictability and reproduce outcomes for customer-facing agents in regulated industries

When building customer-facing agentic solutions in regulated industries, reduce unpredictability and reproduce outcomes as much as possible, then figure out failure modes and unwanted results so you can improve upon them. Banks face financial liability, legal liability, and a bunch of risk when putting out LLM-powered customer-facing solutions, so accuracy of information is critical for the institution. Don't put out inaccurate or hallucinated information in front of your customers.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

Human taste plus automation beats full automation

Fully automating agent improvement with no human time investment produces slop, low quality, and unsatisfied customers; asking a coding agent to 'fix my agent, make no mistakes' is an anti-pattern. Bringing human taste and domain expertise in while automating what can be automated lands at the ideal balance.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse

The AI engineering loop: trace, monitor, datasets, experiments, evals

The foundation of agent improvement is a loop: trace all agent steps into an inspectable log, monitor production traces to find what goes wrong and right, build datasets representing production behavior, run experiments by varying one element at a time, and continuously close the loop by feeding production learnings back into datasets.

AI Engineer · Evals for High-Stakes AI Agents — Doneyli De Jesus & Annabell Schäfer, Langfuse