34,068 results · Most relevant
When auditing an AI-generated financial model, check that every assumption is backed up with a real number from a citation, not just a citation without a number. AI can fail to identify the correct number even when the citation is real, and can draw assumption boundaries (e.g., 5-30%) that exclude the actual disclosed number (45%).
Extra agents pay off the most in planning on ProgramBench, where discovery is the bottleneck. Allocating more agents to planning achieves almost the same performance as the heavy-prompt (PIV) setup at half the cost. Each planning dollar buys almost 1.6x more reward than any other stage (implementation or verification). Planning delegation yields 2.4x more reward per call than delegating throughout implementation.
Replacing Sol workers with cheap Luna workers while keeping Sol as lead achieves almost the same performance at 60% less cost. A strong lead (Sol) with cheap workers (Luna) achieves the same performance as Sol plus Sol but at much lower cost. A 16-planner team (Sol lead, Luna planners) is 38% cheaper than an all-Sol planning team and 70% cheaper than workers at every stage (PIV).
Jesse Hall advises keeping the gap between the caller's last syllable and the agent's first audible sentence roughly inside 1.5 seconds so the call still holds, and notes that getting down to 500-600 milliseconds between turns is when the conversation really starts to feel human-like.
Prospects without pull can be enthusiastic — 'that's the future of our industry' — but aren't saying 'I need this right now'; leaned-out enthusiasm like 'let me think about it' or 'send more info' is not pull, and neither is 'we could use this in Q4.'
Rob Snyder recommends using the pull framework to stake out a claim — state which specific person is stuck — and then go test whether they try to buy before a product exists. If prospects try to buy, you might have something; then see if it repeats. If a prospect doesn't try to buy after the Hail Mary question, treat it as a learning signal that you might not have pull: 'cool, we've learned something.' He notes the Hail Mary question has led startups to completely pivot and take off.
Rob Snyder advises selling to the specific individual at a company who has a project on their to-do list they can't do themselves — 'You're not selling to SpaceX. You're selling to the head of legal at SpaceX who has a project on her to-do list.' This operationalizes the person-level view of pull into a targeting rule for sales.
The framework lets you stake out a testable claim — 'I believe this specific person is stuck' — before any product exists, then verify by seeing if they try to buy.
The thing that really matters is that the potential customer tries to buy — 'okay, cool, how do we get started?' — even before a product exists.
Every stakeholder should evaluate the agent — it is not just one role's job. Get everybody in the room who has a stake in the agent and have them define what success looks like according to their outcome. Stakeholders to include: product, engineering, support, experts, and security.
Dr. Scott Forbes states creatine supplementation improves cognitive performance after sleep deprivation, and even a single dose of creatine helped preserve some cognitive function when people were sleep deprived. In rugby players who got less than three hours of sleep, passing accuracy declined — but creatine maintained passing accuracy despite poor sleep. He recommends considering creatine supplementation when you have a bad night's sleep before an important performance (e.g., athlete before competition, surgeon after a night operation), and taking creatine when traveling across time zones to offset time-delay effects on the brain. Cognitive tests used in creatine research include the MoCA test, mini mental state examination, reaction-time computerized tests for processing speed, and
Consequential creator advice requires independent verification before Today can promote it.
Hari Kumar identifies one of the biggest problems in hedge funds: an analyst works on an idea for a couple of days, it loses PM interest or something else takes precedence, and it doesn't come back for months — so the analyst misses a key price window. He recommends building agents that track all the notes an analyst has taken, even from a single day, against the market, and flash the idea as interesting again to the analyst, PM, and trader when it becomes interesting.
The hosts state that communicating a product is on-device is the most important trust message for local AI products. A founder who applied to YC with an open-source text message analyzer that read iMessages entirely locally used a hero saying 'open source, completely on device' and communicated that the internet doesn't touch the data — which they learned is what users really cared about. An 'on-device LLM' label is also a trust signal displayed on AI sites, though the host questions whether the wording is necessary.
Vague, generalized copy (e.g., 'institutional intelligence and decision infrastructure') leaves visitors unsure whether they need the product. Copy should use specific language about who should use the product and why, and give specific examples of what the product does and the problems it solves. Copy that lists claims without explaining what the product actually does fails the visitor — even after reading 'years of history, alive from day one, nothing installed, no interviews,' the reviewer still didn't know what the product does.
Isabelle Williams argues that agents should operate using an agent token with scoped access issued from your application's backend, instead of taking the user's token or an API key, so agent actions can be audited, revoked, and permission-limited. Handing an agent the user's own token breaks auditing and attribution, prevents independent revocation of the agent without impacting the user's session and access, and prevents limiting the scopes of what the agent can do.
Commercial context: The creator describes this as their own product.
Giving an agent a user's API key instead of a scoped agent token means you cannot audit, revoke, or limit the agent's permissions.
Gary Vaynerchuk advises that every episode of a live show should be post-produced into as much derivative content as possible, because failing to do so wastes content that was already created.
To test whether a model can predict a company action, run the search at a specific historical point in time (e.g., one day, seven days, and one month before an acquisition announcement) rather than live, because a live search will find the announcement and trivially predict the outcome with 100% accuracy, making the test meaningless.
To adjust how you set up a model or agent, backtest whether it could have found a known discrepancy in a generalizable way, for example by creating a set of 50 sample sets and checking performance against them, using version-controlled search to see what information existed at a prior point in time.
Rob Walling advises that if you're bootstrapping, you shouldn't compete head-on when replicating an existing idea. Instead, bring at least one differentiator: a unique focus or position (best X for Y while others chase the broad market), a unique traffic source such as ranking high in Google, or a lower price.
Use saved training trajectories and a controlled intervention to explain a failure, separating equilibrium theory from observed empirical stability.
Demonstrate all 3 repaired bugs and defend why a passed gradient check is necessary but not sufficient for generalization.
Use the tractable fixture and posterior predictive results to separate approximation, sampling and model errors; justify one revised modeling assumption.
Choose sources → to get a library of your own and a daily brief.
Map each implemented component to the original paper and explain 2 reduced-model failure modes without claiming full-scale reproduction.
Use calibration and shift measurements to justify a monitoring or abstention threshold, and give an explicit confounding counterexample to causal feature-attribution claims.
Defend a causal claim or refusal using the DAG, numerical estimates and a counterexample showing why predictive model fit cannot establish identification.
