Creator perspective · MoneyLloyd lays out the components of an effective software factory: a set of automations, a way of providing context and skills, a way of bringing humans in at the correct time when things get stuck, and self-improvement capabilities (loops). A built or bought agent factory includes a set of ways of getting work into the factory at the top, a control plane for figuring out how work gets distributed across the factory floor, and cloud sandboxes where the work happens — including figuring out which agent to run, the harness and the model. He also recommends setting up a data plane below the factory that lets agents remember what they've done, learn, and improve over time.
AI Engineer · Software Engineering Is Becoming Factory Engineering — Zach Lloyd, Warp
Creator perspective · MoneyWhen asked to clean up a feature flag, experienced engineers think through a string of questions first: is the rollout frozen, what about the flag next to it, which team owns it — a flag is never just a flag. That bundle of questions is judgment, and it's the most valuable thing in the building — it currently surfaces once in a code review then is gone, and the next person relearns it the hard way.
AI Engineer · Scale the Judgment, Not the Model — Andrew Orobator, Reddit
Principle · MoneyNobody can hand-audit 300 skills for drift, so automated drift detection is necessary. Give the agent a bedtime: a scheduled pass where it reads its own skills, finds stale ones and contradictions, and opens drafts for review. The system sleeps, cleans house, and wakes up sharper — a self-maintenance cycle for agent skills.
AI Engineer · Scale the Judgment, Not the Model — Andrew Orobator, Reddit
Creator perspective · MoneyFrancois describes a worldview change on how to perform machine learning, influenced by conversations with Francois Chollet and his 2018/2020 'On the Measure of Intelligence' paper on transduction versus program induction. Transduction (learning a function theta that maps X's to Y's) is wrong because it is slow and information inefficient — it needs many pairs of X's and Y's, and with a small amount of data you need a lot of inductive bias. Instead, what you need to act is a generator that maps from X's to Y's: theta takes in few-shot pairs of X's and Y's and emits the function that maps from X to Y — and that is what code is. This is like a coding interview: given a coding problem and some examples, write the function — you are emitting an F that will map from X to Y.
hedged
Y Combinator · Robot-Use Agents: Why General-Purpose Models May Win in Robotics
Creator perspective · MoneyTenev explains tokenization by analogy to stablecoins: custodian real-world assets in a traditional institution, then mint and burn tokens against the reserve so the tokens can trade freely on-chain. He extends the same mint-and-burn wrapper pattern to any real-world asset — stocks, private company shares (via SPV), art, collectibles, and real estate.
No Priors: Artificial Intelligence | Technology | Startups · Tokenization, Agentic Trading, Harmonic: Robinhood's Mission to Expand Ownership with CEO Vlad Tenev
Principle · MoneyTenev states his belief that everyone should be able to invest in private companies, and that the vision behind Robinhood Ventures is for individual investors to benefit from private companies and be a huge shareholder.
No Priors: Artificial Intelligence | Technology | Startups · Tokenization, Agentic Trading, Harmonic: Robinhood's Mission to Expand Ownership with CEO Vlad Tenev
Creator perspective · MoneyVlad Tenev argues that private markets have made it really easy for high-growth successful companies to raise money at massive scale, which did not used to be the case. In the 80s and 90s, companies like Microsoft and Amazon went public at valuations in the hundreds of millions, so individual investors who got in during the public offering could make 1,000x. Now, if an AI lab goes public at a trillion or two trillion valuation, an individual investor at IPO would need to get in at quadrillions to make a 1,000x return, which he thinks is unlikely. Gains from late-IPO companies keep getting concentrated into a smaller and smaller group of venture capitalists or institutional investors.
Historical claims and a forward-looking judgment ('unlikely'); numeric figures flagged as possible_numeric_error in provenance
No Priors: Artificial Intelligence | Technology | Startups · Tokenization, Agentic Trading, Harmonic: Robinhood's Mission to Expand Ownership with CEO Vlad Tenev
Report · BodyBrown fat is a healthy kind of fat tissue that, when activated, increases metabolism and acts as the body's first responder to keep temperature up. Two tissues can increase thermogenesis: brown fat (always temperature-regulating) and muscles (which secondarily start to shiver). Cold is the most potent stressor activator of brown fat because brown fat is the body's temperature-regulating organ.
Andrew Huberman · Essentials: Cold & Heat Exposure for Metabolic Health | Dr. Susanna Søberg
Principle · BodyYou don't have to start with a cold shower or a cold plunge if you're not ready for that yet; just exposing yourself to cold wind has also been shown to activate brown fat. If you don't want to be in an awake cold-exposure state, you can sleep in the cold and maybe won't notice it as much. Sleeping in a 19 degrees Celsius room activates and grows brown fat, as studies have shown.
The suggestion to sleep in the cold includes a hedge ('maybe won't notice it as much').
Andrew Huberman · Essentials: Cold & Heat Exposure for Metabolic Health | Dr. Susanna Søberg
Report · BodyDr. Marius Bakken describes a normal winter training week for elite athletes on the Norwegian method: double threshold — two threshold sessions in one day — on Tuesday and Thursday, the harder X session on Saturday (usually a hill workout in winter, more flat in summer for track preparation), a longer easy run on Sunday, and an easy day on Monday.
Garage Strength · The Genius Who Solved Endurance Training w/ Dr. Marius Bakken
Creator perspective · MoneyAI agent systems can be organized into three planes: data plane, control plane, and intelligence plane; the intelligence plane is described as the key thing missing in current AI agent architectures.
AI Engineer · Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle
Creator perspective · MoneyThe original reflection taxonomy did not include a way to explicitly say the LLM judge could have been wrong, so the reflection agent sometimes fell back to blaming the rubric even when the real problem was an incorrect LLM judgment.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Creator perspective · MoneyAt the foundation is the base agent (question answering, summarizing, chat, workflows); next is agent evaluation where gaps appear (bespoke workflows, one-off analysis, manual inspection, few standardized processes); above that is agentic reflection (systematic investigation and root cause diagnosis); at the top is the RSI loop connecting evaluation and reflection to optimization and validation.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Creator perspective · MoneyRoland Gavrilescu describes how Cursor and Cognition went from building the best product, to building the best evals for the product, to finally building the best models based on the previous two artifacts. He endorses this product → evals → models progression as the recipe for everything going forward, noting code was the first domain where it was successful.
AI Engineer · The Loop Is the Product — Roland Gavrilescu, Introspection
Creator perspective · MoneyEvaluating tool calls is an open problem: tools depend on the state of the world at the time they ran, which differs when you re-run evals (e.g., a Slack support bot evaluated with a different model).
AI Engineer · How a Trading Firm Ships New AI Models the Same Day - Noyan Tokgozoglu, Optiver
Report · MoneyLog everything — every tool call, input, output — via OTEL logging, and feed it into the evaluation system.
AI Engineer · How a Trading Firm Ships New AI Models the Same Day - Noyan Tokgozoglu, Optiver
Creator perspective · MoneyGEPA produces detailed problem specifications — how to make sense of input, purpose and context of the pipeline part, key observations and lessons from the data — unlike prior prompt optimizers that use model idiosyncrasies (e.g., 'my grandmother will be really angry if you don't generate a good prompt').
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
Principle · MoneyOptimizeAnything provides three optimization modes: single problem, multitask search across n related problems, and build-a-skill for generalization to new deployment problems. If you have a single problem, like one matrix multiplication kernel to optimize, use the single-problem mode. If you have n related problems with potential information transfer, like a matmul kernel plus a dot product kernel, use multitask search mode. If you optimize on a fixed set of problems but deployment will surface new problems, use the build-a-skill mode, because you care about generalization. In math prompt optimization, training examples differ from deployment queries, so generalization modes matter.
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
Creator perspective · MoneyGEPA's algorithm takes an AI pipeline written in any agentic framework or raw LLM calls, runs it on a few examples, and collects domain-specific feedback from whatever information the environment contains.
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA
Creator perspective · MoneyRL with verified rewards learns only an O(1) score from each rollout, discarding chains of thought, tool calls, environment responses, and error messages that contain diagnostic value.
AI Engineer · Beating RL With Reflection: GEPA and Optimize Anything — Lakshya A. Agrawal, GEPA