Rosa Del Mar

Daily Brief

Issue 62 2026-03-03

New Low-Cost Model Option And Unit Economics

  • Google released Gemini 3.1 Flash-Lite as an update to its inexpensive Flash-Lite model family.
  • Gemini 3.1 Flash-Lite supports four different thinking levels.
  • Gemini 3.1 Flash-Lite pricing is stated as $0.25 per million input tokens and $1.5 per million output tokens.

Low-Cost Model Release And Unit Economics

  • Google released Gemini 3.1 Flash-Lite as an update to its inexpensive Flash-Lite model family.
  • Gemini 3.1 Flash-Lite supports four different thinking levels.
  • Gemini 3.1 Flash-Lite pricing is $0.25 per million input tokens and $1.5 per million output tokens.

Low-Cost Model Refresh And Explicit Token Pricing

  • Google released Gemini 3.1 Flash-Lite as an update to its inexpensive Flash-Lite model family.
  • Gemini 3.1 Flash-Lite supports four different thinking levels.
  • Gemini 3.1 Flash-Lite pricing is stated as $0.25 per million input tokens and $1.5 per million output tokens.

Maintainability And Capability Atrophy Risks From Ai Coding

  • Organizations that rely on AI to do everything risk eroding internal engineering competence over time.
  • ULMFiT uses a three-stage pipeline: pretrain on a general corpus, fine-tune on task-specific text, then train a downstream classifier.
  • Notebooks can be made Git-friendly using a notebook-aware merge driver that provides cell-level diffs and merge conflicts while keeping notebooks openable.

Understanding Maintainability And Behavioral Risk From Ai Coding

  • Jeremy Howard argues AI-based coding can create an illusion of control while producing code that maintainers do not understand.
  • Jeremy Howard warns that aggressive AI coding adoption can erode internal engineering competence and increase future maintainability risk.
  • Jeremy Howard describes ULMFiT as a three-stage process: pretraining on a general corpus, fine-tuning on task-specific text, and then training a downstream classifier.

Valuation Anchoring: Cash Flows, Buybacks, And Forward Vs Trailing Metrics

  • Token buybacks are described as primarily a mechanism to return capital to holders, and critics are said to lack a counterfactual for why buybacks are inefficient.
  • The primary sustainable crypto revenue pools are described as trading, DeFi borrow/lend, derivatives, and asset issuance that leads to trading, rather than payments transfers or Web3 social microtransactions.
  • For top-10 tokens, meaningful outperformance requires substantial new external inflows, while sub-$5B tokens can reprice significantly with marginal in-crypto flows.

Value Accrual Stack Apps Vs Layer1 And Cash Flow Mechanics

  • Token buybacks primarily function as returning capital to token holders, and critics lack a clear counterfactual for claiming buybacks are inefficient.
  • If agents remove consumer-commerce friction, they will prefer the most efficient (near-zero fee) payment rail such as stablecoin transfers, and this does not imply the underlying layer-1 token is a good investment.
  • An AI-driven 'rapid economic blow up' scenario is directionally plausible but likely overstated in timing and magnitude because market signals imply a slower path due to capex and scaling constraints.

Go To Market And Sales Leadership Hiring As High Error Surface

  • Brian Halligan reports that HubSpot used a two-pass role-play interview, and that improvement on the second pass was a strong hire signal.
  • Ben Horowitz asserts that the best tech companies tend to have founder-CEOs who are blunt and ask aggressive questions so truth surfaces and bad news travels quickly.
  • Ben Horowitz asserts that founder mode can be a necessary correction after over-deference to senior hires creates fiefdoms, but that it should not be interpreted as advice to avoid hiring experienced executives.

Ai Operationalization With Explicit Productivity And Quality Metrics

  • Robinhood is prioritizing AI in customer service and software engineering by deploying AI agents for support and tracking deflection rate, AI-generated commit share, and overall engineering velocity.
  • Tenev says traditional 40-act fund structures can be used in the U.S. as an alternative rail to provide retail exposure to alternatives and private assets.
  • Tenev says making Robinhood a customer’s primary financial institution depends on broad product coverage, especially being the place where a paycheck can land.

Sales Org Design And Vp Sales Hiring Signals

  • Brian Halligan states that HubSpot used a two-pass role-play interview (sell, receive feedback, sell again) and that improvement on the second pass was a strong hiring signal.
  • Ben Horowitz asserts that great founder-CEOs share independent, original thinking and an ability to attract followers "out of curiosity," which enables higher talent density via hiring.
  • Ben Horowitz asserts that top tech companies tend to have founder-CEOs who are blunt and ask aggressive questions so that truth surfaces and bad news travels quickly.

Expansion Toward Primary Financial Institution: Banking, Card Flywheels, And Mortgage Distribution Without Underwriting

  • Vlad Tenev says Robinhood began rolling out a product called Robinhood Banking a couple of weeks before the interview.
  • Vlad Tenev says Robinhood is prioritizing AI in customer service and software engineering and is tracking deflection rate, AI-generated commit share, and overall engineering velocity to validate net output gains.
  • Vlad Tenev says traditional 40-act fund structures can be used in the U.S. as an alternative rail to provide retail exposure to alternatives and private assets.

Limits And Drivers: Contested Sizing, Governance Constraints, And Non-Financial Risk

  • In early 2021–2022, Andrew Beatty and co-founders were seriously concerned about potential retaliation for disrupting large-scale money-moving operations tied to cartels.
  • BeatDap runs roughly 700 continuously updated models to detect music streaming fraud.
  • Early Facebook growth could be manipulated using likejacking by hiding a Like/Follow control (e.g., in a pixel) so users unknowingly like a page while clicking elsewhere.

Streaming Anti-Fraud: Telemetry-Based Detection, Model Operations, And Enforcement Levers

  • BeatDap is described as running roughly 700 continuously updated models to detect music streaming fraud.
  • Early Facebook growth could be manipulated via likejacking by hiding a Like/Follow control in a pixel so users unknowingly like a page while clicking elsewhere.
  • Music labels asked Andrew Beatty's team to use blockchain to create real-time receipts for song plays because streaming services typically provided only aggregated CSV play counts without usage-level proof.

Expert Belief Update / Legitimacy Signal

  • Donald Knuth expressed joy that his conjecture has a nice solution and that the solution illustrates progress in automated reasoning.
  • Donald Knuth stated that Claude Opus 4.6 solved an open problem that he had worked on for several weeks, and he stated the model was released three weeks earlier.
  • Donald Knuth characterized the event as evidence of a dramatic advance in automatic deduction and creative problem solving.

Legitimation And Belief Update Signals From A Prominent Researcher

  • Donald Knuth expressed joy that his conjecture has a nice solution and that the solution illustrates progress in automated reasoning.
  • Donald Knuth stated that Claude Opus 4.6 solved an open problem that he had been working on for several weeks, and that the model had been released about three weeks earlier.
  • Donald Knuth stated that he expects to revise his opinions about generative AI in light of this result.

Frontier Model Reported To Solve Open Research Problem

  • Donald Knuth stated that Claude Opus 4.6 solved an open problem that he had been working on for several weeks.
  • Donald Knuth stated that Claude Opus 4.6 is a hybrid reasoning model that was released three weeks before the reported solution.
  • Donald Knuth expressed that he felt joy that his conjecture had a nice solution and that the solution illustrated progress in automated reasoning.

Syld-Recent-Underperformance-Flows-And-Valuation-Conditions

  • The episode asserts that Cambria Shareholder Yield ETF (SYLD) ranks top decile versus its Morningstar category over both 10-year and since-inception periods, but ranked in the bottom 11% in 2025.
  • The episode claims that $10,000 invested in Berkshire Hathaway in 1965 would have grown to over $600 million by year-end 2025.
  • A Vanguard paper is cited as finding that among 2,085 active funds, 552 (26%) survived and outperformed from 2000 to 2014.

Case-Study-Syld-Recent-Drawdown-Flows-And-Valuation-Contrast

  • Cambria Shareholder Yield ETF (SYLD) ranked top decile versus its Morningstar category over both the 10-year period and since inception, but ranked in the bottom 11% in 2025.
  • Warren Buffett retired from Berkshire Hathaway in 2025.
  • Shorter holding periods increase the influence of randomness and luck on observed returns, making false conclusions about skill more likely.