35,595 results · Most relevant
The talk's core principle: models should never write numbers, only cite them. Kepler's guarantee is that models cite numbers rather than write them, and the model is structurally incapable of producing output it cannot cite.
The agent submits an investment memo (long/short call) together with a reasoning graph in which every fact points to the exact tool call it came from. Former buy-side analysts write down the key reasoning steps that any good memo would include, and each step is a node in the graph. For the Sonoco task, agents must rebuild the company as it was in May 2025, determine what the new business really earns, net out what was sold, and rebuild the debt. For each reasoning step, the memo is attacked with a what-if test, e.g., if management promised cost savings from the EBSIS deal but the savings were zero, would the call change, and does the memo explain this with clear reasoning about the moving parts. Memos can reach different calls and both can score well as long as the work behind each call is
In the spirit of the bitter lesson, the best cost model is no cost model: run the programs on the real hardware you'll run on and see how long they take. Measuring on real hardware is a dumb but very effective solution: moves that programmers don't think are beneficial can turn out to be very beneficial when run on the GPU.
Dr. Scott Forbes states creatine is extremely safe as a dietary supplement, citing a review of more than 685 randomized controlled trials whose authors concluded there is no increase in frequency or prevalence of any side effect compared to placebo. The podcast host affirms creatine is safe based on the research.
Dr. Forbes cautions that the popular study claiming you can gain muscle mass on creatine without working out measured lean mass, not direct muscle mass; a DEXA study showed about a 3% increase in estimated lean mass just from manipulating muscle water and glycogen, so lean mass should be interpreted cautiously.
none stated
Coach Kyle argues that the presence of family, friends, or a crowd doesn't mean anything — a lot of people around, nobody cares — so you should approach women you see out in your day to day anyway, including when they're with others.
Data should not just be dumped on investors' desks; it must be aligned to an internal ontology with enrichments like countries, concepts, markets, and the time period discussed in the document. This alignment to the internal ontology and world model is what gets great leverage from the data.
Humans need precision and a high signal-to-noise ratio because they can only read so much — even a fast reader manages about 300 words a minute. Agents are far less sensitive to precision: if the data is there, they'll find it, and anything you hand to them will get attended to. Agents scale super quickly and consume about 100 times faster than humans — about 30,000 words per minute.
Ben Horowitz says that when founders fail at the CEO job, it almost always comes down to a crisis in confidence — they know they don't know what they're doing but still must make decisions. Hesitating makes the company nervous, slows everything down, and creates politics as people try to fill the leadership void; the job is to make high-quality decisions on time even when you don't know if the decision is right.
Ben Horowitz says when a founder knows a decision is necessary but fears the consequences, hesitating guarantees the feared outcome — founders think about all the bad things that will happen, hesitate, and get wrecked. The rule: if the only viable choice is bad, do it immediately without hesitation. When he went public, all the bad things he thought would happen did happen, and it was a bad time — but the alternative was bankruptcy after talking to literally every investor in the private markets.
Every entrepreneur has flaws and self-doubt — even Zuckerberg at 22. The question is not whether they lack weakness but whether they are world-class at something, the best in the world at a thing, with a secret nobody else has.
Hari Kumar explains that in fundamental investing you cannot show AI ROI immediately because ideas take three months to a year to manifest in stock price impact, so ROI must be measured in a process-based manner — asking process-by-process whether AI is actually changing how someone works or just adding thrash.
Make sure your products match your promises to create a virtuous cycle of branding. The virtuous branding cycle: what you say is the first level of branding, then people consume your product and it's what other people say, then a customer who heard what others said buys it, and finally it's about what they say. No matter how good the marketing was, when someone buys a product they will make a decision about it based on their experience.
Operating model option 1: build and operate the full stack — hardware, site; your team arranges power, permits, assembly, and management. Option 2: own servers in a co-location facility with more or less assembly and operation handled by a partner. Option 3: an off-take contract buys defined capacity where the provider owns and operates the asset — but may not give complete root access to bare metal, with access levels depending on tenant vs multi-tenant configurations, which also affects pricing.
Arca gave its agents signals from the external world (emails, calls, documents, meetings, custodian events, market movement), all operational context on what to do about each signal, a harness with tools, guardrails, and persistent memory over the long horizon, and the ability to collaborate with people.
The consistent advice since the team started 15 months ago: stair-step — look at two or three (maybe four) agent tools, run a small test, pick one and get it working even for a subset of use cases, push it as far as you can, then add another when you see gaps. Early gaps were outbound vs. inbound: they used Qualified (Salesforce spot) because it was optimized for qualifying inbound while other tools did outbound. In the early days they ran many different outbound agents because so many vendors wanted them to try their tools.
None stated.
Set up benchmarks within your own organization that are as reflective of actual workloads as possible. Identify a target workload when setting up evaluations of realistic work associated with your particular task loads, and evaluate realistic work rather than generic benchmarks.
Track whether your landing page converts as the real metric, rather than how it looks. Think about what you are ultimately trying to do with the landing page before designing it.
The reviewer infers that AI agents love to overdo it, and recommends evaluating whether a design trend is a 'vibe coding smell' before adopting it, as some patterns signal AI-generated work.
The Magic Patterns guest endorses the idea that shipping only the vibe-coded output without polishing the extra 20% results in a lower-quality product.
Choose sources → to get a library of your own and a daily brief.