Creator perspective · MoneyEarly language models on benchmarks like GSM8K had to output the answer directly with no chain of thought; allowing a model to think before answering improved math performance. In RT2-style VLAs, the model likewise has to output an action directly, with no ability to allocate more compute for a more complex task.
Y Combinator · Robot-Use Agents: Why General-Purpose Models May Win in Robotics
Report · MoneyRobinhood Ventures Fund One is a publicly traded venture capital fund: Robinhood raised capital via its IPO retail channel from individual customers, invested it in startups, and the fund trades on the NYSE like a stock — holdings include OpenAI, fintechs, and AI companies like Eleven Labs and Cursor. Robinhood Ventures Fund 2 partners with Y Combinator to give retail investors exposure to seed and Series A startups — the earliest stage where risk is highest but opportunity is greatest — and is run by Rich Aberman. The vision: it should be an anomaly for a startup's seed and Series A cap table to lack individual retail investors as a big constituent.
No Priors: Artificial Intelligence | Technology | Startups · Tokenization, Agentic Trading, Harmonic: Robinhood's Mission to Expand Ownership with CEO Vlad Tenev
Report · MoneyVlad Tenev reports that Robinhood launched an agentic trading product earlier this year built on a Robinhood MCP, and also has an agentic credit card. Over 150,000 customers have opened agentic trading accounts in the few months since launch, and tool calls have been following a smooth exponential, with the last announced number around 30 million tool calls per day.
Publisher-reported metrics from the CEO
No Priors: Artificial Intelligence | Technology | Startups · Tokenization, Agentic Trading, Harmonic: Robinhood's Mission to Expand Ownership with CEO Vlad Tenev
Report · MoneyRobinhood is tokenizing private companies (currently outside the U.S.) with the aim of making them real-time tradable 24/7 in fractional quantities, like a stock or cryptocurrency — e.g., trading Anthropic 24/7 in low dollar amounts, with ownership quality at least as good as an LP interest in an SPV.
No Priors: Artificial Intelligence | Technology | Startups · Tokenization, Agentic Trading, Harmonic: Robinhood's Mission to Expand Ownership with CEO Vlad Tenev
Report · BodyIn Dr. Søberg's study, winter swimmers did their cold and sauna protocol two to three times per week. Each visit consisted of three rounds: three cold dips and two sauna sessions, starting in the cold and ending in the cold water.
Andrew Huberman · Essentials: Cold & Heat Exposure for Metabolic Health | Dr. Susanna Søberg
Principle · SocialThe creator describes walking toward the seating area first and sitting on the couch so that, by default, she has to sit next to you; he argues this seating and positioning creates a strong vibe.
Coach Kyle · Watch Me Create Attraction on a Live Date (uncensored infield)
Principle · SocialThe creator advises getting closer to your date physically over the course of the interaction, noting he was at times literally one foot away from her face, and that you should stay physically close from the start of the date.
Coach Kyle · Watch Me Create Attraction on a Live Date (uncensored infield)
Creator perspective · MoneyAn ontology layer adds a level of accuracy to agent answers that cannot be achieved with simple keyword search alone; in the demo, the agent without an ontology answered a counterparty question but the answer improved and became correct with the ontology layer, and a naive agent without a semantic layer got 'what does concentration mean' wrong while the intelligent agent got it right.
AI Engineer · Everyone Has the Same AI Models. Here's What Sets Agents Apart - Richmond Alake, Oracle
Report · MoneyThe team uses majority voting on held-out evaluations to help mitigate variance from individual LLM judgments.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Report · MoneyAfter clarifying the evaluation rubric, human alignment of the LLM judge increased from 57% to 64% while mean error decreased from 0.96 to 0.74.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Report · MoneyThe team updated rubrics so harmless formatting differences are not semantic failures, introduced title normalization guidance, clarified handling of concurrent employment, seniority inference, and geographic references, and calibrated rules separating low scores from partially correct outputs.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Creator perspective · MoneyAnalysis of judge-human disagreements revealed specific failure modes: the judge sometimes evaluated plausibility rather than completeness (accepting a plausible job title while missing omitted signals); its lowest scores were often overly harsh because it was better at recognizing strong outputs than distinguishing weak ones; it penalized harmless formatting differences like capitalization of 'full-time'; and it made inconsistent assumptions about what could be inferred from a member's existing employment information.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Principle · MoneyThe team describes a ladder of optimization targets with increasing complexity: start with prompt updates because it is the easiest and most straightforward way to keep iterating the AI system; then optimize the structural context of the AI system (L1 ladder); at the L2 ladder, expand scope to automatically design the entire workflow, which is a kind of meta information design; after workflow design, optimize the harness code itself; and at the end of the ladder, go further to optimize the optimizer code itself.
AI Engineer · An Agent That Fixes Its Own Mistakes - Saloni Gandhi & Ziwei Guo, LinkedIn
Creator perspective · MoneyDaksh Gupta catalogs the failure modes of a naive review agent: it will miss a lot of bugs because bugs are interactively large (fractally infinite); it will miss context and misunderstand how parts of your codebase work, maybe not understanding how one of your libraries works; it will be very expensive, with people typically spending between $3 and $25 per pull request using coding agents; and it will produce nitpicky comments — technically correct but overly pedantic feedback.
hedged
AI Engineer · How to Build an AI Code Review Agent - Daksh Gupta, Greptile
Report · MoneySupabase is announcing a final change for declarative schemas covering all of Postgres — a challenging engineering effort. With it, you can have a full file representation of your Postgres database shape declaratively in your folder.
Latent Space · Supabase’s Plan for a Billion AI-Built Apps
Report · MoneySupabase's P1 strategy is to partner with as many frontier AI platforms, labs, and models as possible, which drives attach rates up. US AI labs currently deliver the most traffic, which is why Supabase primarily partners with them, though the speaker notes that could change.
Partnership focus could change as traffic patterns shift
Latent Space · Supabase’s Plan for a Billion AI-Built Apps
Creator perspective · MoneyPosition your product as an alternative to an established incumbent (e.g., 'open source Firebase alternative') instead of a technical description (e.g., 'real-time Postgres') to get traction — the technical tagline got zero attention.
Latent Space · Supabase’s Plan for a Billion AI-Built Apps
Creator perspective · MoneyTereza explains that automatic model routing decides first which model is most optimal and can switch to a different model if it fails the task or runs into trouble. She argues routing helps with reliability and speed, not just saving money: open source models are often faster, and if one LLM provider fails you can switch to another automatically.
AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory
Creator perspective · MoneyTereza describes routing as working in four parts: assign the task, classification, threshold, and choosing the cheapest model above the threshold. In classification, she says to look at the structure of the prompt, the codebase, how difficult the task is, and what tools are being used. After setting a threshold of what is enough to accomplish the task, you choose the cheapest model above that threshold.
AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory
Creator perspective · MoneyRouting works in three parts: assign the task (organizations can give different permissions and different default models to different people — like marketing, sales, or engineers — which is very useful), classification (look at the structure of the prompt, the codebase, how difficult the task is, and what tools are being used), and a threshold (choose the cheapest model predicted to accomplish your task).
AI Engineer · Rise of the Software Factory — Tereza Tížková, Factory