Report · MoneyGEPA is a reflective prompt optimization technique for agents that uses an evolutionary loop along with a novel Pareto-based candidate selection. It is akin to doing reinforcement learning in text space: instead of just receiving a reward score, one obtains the score along with textual feedback that can be very domain-specific.
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
PrincipleElie Bakouch endorses that for a real benchmark, you want to run multiple seeds and put all models in the same condition rather than relying on a single unstructured run.
AI Engineer · We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect
PrincipleElie Bakouch endorses giving agents a scratch pad file to write into as active memory so they can persist notes and decisions across context, and giving agents the ability to fetch current human records at any time so they can improve on the latest state of the art.
AI Engineer · We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect
ReportThe proposed system is a multi-agent system where a bunch of generators interact together; it can use both a closed model and open source models, which are super effective for the cost. In the discovery loop, the generator models suggest ideas, then you run the speedrun to get the reward. A judge in the loop gives quality feedback on the generated ideas, and the judge can have taste about whether a method is good or not, operating outside the loop. After judging, you can decide which method you want to scale to a larger number of parameters and number of tokens. The approach is inspired by Alpha Evolve by Google and a bunch of papers released since then.
AI Engineer · We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect
PrincipleElie Bakouch endorses requiring that a result pass a statistical threshold to validate a speedrun record, so it is not just seed optimization or luck.
AI Engineer · We Let Claude Code and Codex Race Human Researchers — Elie Bakouch, Prime Intellect
ReportGit revert does not reverse all changes a coding agent makes, because agents also modify runtime state (tool registries, middleware, configurations) and trigger external effects such as deployments or database writes.
AI Engineer · Your Coding Agent Can't Undo Its Own Mistakes - Tanmay Sah & Dolly Sah
Creator perspective · MoneyVinoo Ganesh describes how the combination of deterministic storage and pointer-only reads allows verifying that retrieved information actually came from the source it claims to come from, and that this verification mechanism becomes fairly globally applicable.
AI Engineer · AI Can't Make Up a Number It Can't Look Up - Vinoo Ganesh, Kepler
PrincipleThe narrator distinguishes complex (a property of the domain) from complicated (a design failure), arguing that complicated is what happens when complexity is poorly managed — e.g., interfaces designed as an afterthought or machine logic imposed on users without understanding their mental model.
NNgroup · Complex vs. Complicated
Creator perspective · MoneyJesse Hall explains that turn detectors relying only on the transcript see a pause and assume the caller is done, interrupting even mid-sentence — for example responding after '$500' when the caller was about to say '$55.23'.
AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit
Creator perspective · MoneyJesse Hall describes that when the gap between the caller's last syllable and the agent's first audible sentence exceeds about 1.5 seconds, the conversation starts to feel off; the right question is how human-like an agent can fit within the latency window.
AI Engineer · How to Build Voice Agents That Don't Fall Apart When You Interrupt - Jesse Hall, LiveKit
Principle · MoneyRob Snyder recommends that before building, founders ask themselves honestly whether they would buy the product right now — and if not, don't build it. This is a self-test that applies the pull concept to the founder's own situation before investing in construction.
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Creator perspective · MoneyRob Snyder says the pull framework can help set the bounds for pricing: ask what the value of getting the project done is, or what the costs of the alternatives are. Ask what the prospect currently spends with their existing solution (e.g., market research consultants) and consider charging the same price — e.g., $50,000 a year, the same as their current consultants — to give an intuitive frame.
guest's recommendation with an illustrative dollar figure
Lenny's Podcast · How to find product-market fit (but, actually) | Rob Snyder (Author, The Power of Pull)
Principle · MoneyTo prevent data leakage and ensure document-level security, every hit is re-authorized in Postgres, checking citation, value, and document across every computation.
AI Engineer · Do Not Let Models Invent Numbers - Vinoo Ganesh, Kepler
Principle · MoneyThe invariant that makes numbers citable, deterministic, and reproducible is displaying both the calculation and the source. Every number and fact in the platform is clickable to its source; clicking reveals the source document (10K, 10Q, 8K, investor presentation) and the actual calculations behind the scenes.
AI Engineer · Do Not Let Models Invent Numbers - Vinoo Ganesh, Kepler
Creator perspectiveThe task models how work is actually assigned in the profession: a PM at an event-driven M&A fund (holding period 3-12 months) tells the analyst that company A will be bought by company B and asks whether to go long, short, or pass — requesting an investment memo detailing the evidence and reasoning. Nobody gives an analyst a simple isolated task.
AI Engineer · Why Frontier Models Fail as Buy-Side Analysts - Janie Gu, Google DeepMind & Suyeol Yun, LinqAlpha
Principle · MoneyMany scraping, automation, and workflow jobs kill the browser and Chromium instantly as soon as the task is done. Jay discourages this: instead, leave Chromium running and close the windows out. This is presented as one of the key changes Browserbase made to their verified browser.
AI Engineer · How to Build Financial Agents That Work Across the Web - Jay Sahnan & Arik Bird, Browserbase
Principle · MoneyWith Playwright, as soon as a website changes — the button moves, the text box changes from first name to just name — your script breaks when looking for that XPath or trying to click on an element or selector. Jay notes this is fine for one-off tasks run locally, but scaling a company like Ramp or Bill requires reliability every single time.
AI Engineer · How to Build Financial Agents That Work Across the Web - Jay Sahnan & Arik Bird, Browserbase
Report · MoneyDan Fu and others constantly try to optimize four key inference KPIs: time to first token, time per output token (decode step), throughput per GPU hour, and goodput at some service level.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Creator perspective · MoneyIndividual operations like layer norms, RMS norms, and softmaxes have very low arithmetic intensity. Writing a standalone kernel for such an operation yields very low flop utilization. This is historically why training only reaches about 50% end-to-end flop utilization and decode about 5%.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club
Creator perspective · MoneyWhen evaluating quantization quality, standard benchmarks like GSM8K, MMLU, and Complexity are used, though some are saturated (e.g., MMLU at 90+ percent). They look at KL divergence of individual layers to measure how much activations changed during quantization. When any model failure is reported, it's added to the test case suite, and the end-to-end agentic pipeline is evaluated with hundreds of test cases built from every reported model failure. Agentic evaluation loops like the Together compiler project — where the model optimizes its own kernels — are also used to detect quantization degradation.
Y Combinator · Not All Open Model Tokens Are the Same | YC Paper Club