35,595 results · Most relevant
Chris Dixon's framework holds that the best ideas look like bad ideas because the entrepreneur has a secret that makes them actually good; obvious good ideas get their profits competed away, so a good-idea-that-looks-bad beats a good-idea-that-looks-good. When a founder describes such a secret, probe for it — it is what differentiates the opportunity from a genuinely bad idea.
Part of the playbook is figuring out how to get the best people to work for you and keep them. Poach great people constantly — without someone great like a top CFO, the business will crumble. When your business grows fast, everyone starts DMing your employees on LinkedIn offering more money and equity to poach them. The guest's countermeasures: give equity to people that deserve it, vest equity for long-tenured employees who deserve it, give an equity package to amazing new hires, and offer crazy bonuses between 50 to 100% if employees do really well and the company hits targets. Bonuses get people sticking until the end of the year, then you have to set a new benchmark for the next year.
The guest says what's more important than getting more creators is getting your current creators to do more — e.g., convincing a creator doing 30 creatives a month (one a day) to do two a day. His top creator made almost $70k in commissions last month, will probably do over $100k this month, and makes 600 ads a month — all he does is toss content and lives and breathes tank tops. By contrast, most brands' creators treat the work like a side hustle, making a video then a few more days later, which limits results.
You can determine where software sits on the disposable-durable scale using four factors: consequences of failure, cost of replacement, dependencies, and cost of maintenance. Err on the side of software being durable — assume all four factors are higher than you expect.
none stated; one atom flagged possible_numeric_error in provenance
Adit Abraham warns that benchmarks have become marketing — everything claims to be state of the art, making it hard to know what works in your context. He advises auditing how your golden dataset was created (human-made sets drop information; LLM-made ones inherit the cheaper model's mistakes, e.g., Sonnet instead of Opus), avoiding overly strict benchmarking rules that penalize faithful alternative representations, and stress-testing vendor benchmarks by adding small noise (tilt one degree, slight Gaussian blur) — a publisher whose results degrade far more than competitors under noise is suspect.
Adit Abraham explains that in multi-step document pipelines (splitting, extracting, editing), each step needs high accuracy because mistakes compound — you can't have 90% accuracy on individual pages when processing a dozen files. Traditional OCR accuracy sits in the 70-80% range, while modern approaches reach 97-99%.
Adit Abraham argues that documents are ultimately just images — a 9x11 grid where information can be constructed any which way, not just characters on a page. Reading a document properly means understanding structure: with two columns, knowing to read the first column then the second; with a table, preserving rows, columns, and merged cells.
The creator sees movie critics as the opposite of what they need — snobs, opinionated, whose job is to criticize art rather than make it — but wanted to test whether movie connoisseurs might hold a key to better recommendations. The test method: find a group of well-respected movie critics and seven very polarizing films that some critics loved and others hated; rate the polarizing films based on your own personal tastes and find which critics better align with your taste, which could let you find stuff worth watching. The creator's theory is that the best recommendation isn't a crazy complicated algorithm at all — maybe it's just humans finding the critics whose taste is closest to yours. They already follow entire YouTube channels built around people telling them what movies are cool to.
The vendor side of the stack included the harness, the model, FleetKit, and a code modernization plugin — any of which could be replaced with another vendor, model, or framework. The non-replaceable 'Geico' side included context and knowledge base (software standards, past decision archaeology), a list of tools the agent could access plus instructions for using them, methods (skills and quirky domain-specific optimized ways of developing software), and judgment (who can decide what and at what gate, from both SME and leadership points of view).
In finance you can't just hand the decision-making loop to a team of agents because the other side of the trade is a client — if it's a commodities trade, it's a physical cargo; if it's a company, it's a team of people. When something goes wrong you need a human to stand up and say 'I made the decision for this reason.' If an agent leaks your API key, all it says is 'oops, I'm sorry, I shouldn't have done that, please rotate it' — unacceptable in finance, because an agent is just self-correcting and one small error exponentially gets bigger.
endorsed as the speaker's argument
Every investor is looking for the same thing: edge. In practice edge consists of three things — knowing what's going on before anyone else, knowing your blind spots (the second and third order derivatives of what's happening), and having conviction so you can size up on a trade you believe in. If everyone building in the investment space has access to the same models and systems, the question becomes what actually constitutes edge — and the answer is what you can't download, what sits in an investor's head.
endorsed framing by the speaker
When an investor asks for your opinion, they don't just want to know what you think — they want to know how you came to that reasoning. The approach described: measure hallucination, measure every claim, check how it's traceable, and give the user the assumptions used and how they led to the decision. Moshi leaves the decision-making with the human because that is what humans do best — there's no machine more powerful than the brain.
endorsed; the measurement claims are the builder's own description
GaryVee describes how his company operated: he was thrilled to keep everyone on the payroll forever for vibes, people only got fired when everyone else thought they sucked — and nobody told the person, only him and his brother the co-founder. He frames this as an anti-pattern: failing to deliver candor to close employees for years means problems fester invisibly until they surface as a shock.
Software that is authorized to actually do the thing (like move money), not just recommend it, is a durable moat. Stripe Connect and Reemby are given as examples.
The learning must not be easily stealable, exportable, learnable, or simulatable; it should come from observing the real world on a real-world clock rather than from synthetic data that can be generated quickly.
A buyer will not replace a known HIPAA-compliant system with a vibe-coded replacement because the compliance risk is too high. Similarly, a buyer would not accept a vibe-coded e-signature replacement for a million-dollar contract, and a hypothetical buyer (Tom) would not take the compliance risk of a vibe-coded EHR.
The guests dig into the practical challenges of proving gains are real: variance, reward hacking (including reward hacking occurring two levels down in an agent chain, where the agent that built the agent that built the agent hacked something), and the enormous difficulty of distinguishing genuine improvement from noise. LLMs are fundamentally super stochastic; specific agents on specific benchmarks have levels of noise that are sometimes unpredictable and extremely weird; even two B200 GPUs can have slightly different behaviors on the physics, some of which replicates at the high level. When they try to replicate other RSI announcements, they realize the results are noise or cherry-picking.
based on the guests' internal experience
Owen offers a definitional test for recursive self-improvement: ask what the optimizer is and what the optimizee is. If the improver is still a human using AI tools, it is not RSI — it is autocatalysis. The distinguishing feature of RSI is that the model/agent itself becomes the thing being optimized, enabling reinvestment of cognitive improvement into itself.
Owen reports that many published papers claiming recursive self-improvement fail when tested at scale on real benchmarks and economic problems, despite sounding compelling on paper. He advises attempting replication before believing RSI or AI breakthrough claims — most public announcements turn out to be noise or cherry-picking. His company assigns a dedicated engineer, the 'Viceroy of variance,' whose explicit job is to attack and destroy the team's results and find what's wrong with them.
Owen proposes evaluating RSI claims from a control-theory perspective: an autocatalytic system has no closed loop — there is a gap, and humans must pour in resources to make it work. A true closed loop can produce constant or growing marginal gains over time, which is the fundamental difference.
Choose sources → to get a library of your own and a daily brief.