Developers should use probes to detect if models are trying to hack something
Developers should use probes to detect whether a model is trying to hack something — a practical monitoring technique for AI safety.
endorsed
34,950 results · Most relevant
Developers should use probes to detect whether a model is trying to hack something — a practical monitoring technique for AI safety.
endorsed
Anish Acharya argues price is a measure of product-market fit and recommends a product exercise: imagine your product costs $1,000 or $10,000 a month — a software Birkin bag — and figure out what it would have to do to justify that price.
Mention that you were considering getting a coffee or tea soon without explicitly asking her to join, to implicitly invite her and let you ask her out without ever technically getting rejected.
A financial analyst agent should produce actionable recommendations, not just summarize data. No built-in template checks for actionability, so you have to build a custom eval that checks for it.
Dr Ghayas Issa advises maintaining a high index of suspicion for differentiation syndrome in any patient on a menin inhibitor when they start treatment. For revumenib, it occurs roughly within one to two weeks; for ziftomenib it can occur up to a month or two months after starting, related to the drug's longer half-life.
Consequential creator advice requires independent verification before Today can promote it.
The onsite is three back-to-back interviews. The mock cold email can be done live on a laptop or drafted beforehand and brought to the meeting; the interviewer asks why certain decisions were made and may have the candidate rewrite a sentence on a whiteboard. The test checks whether they applied the prep and used their brain, not whether they wrote the best email or copy-pasted from ChatGPT. The rep filtering test uses the first framework taught: from 40,000 venture-backed companies at Carta down to 25, with filters, not stopping until they hit 25; candidates typically get stuck around three filters. The rep wildcard tests hustle tactics that get you from 80% to 100%, like how to find accounts and who jives with AEs, and also tests cultural fit, seeing whether the candidate gets excited or
For most middle market or commercial sales motions, the host says SDRs should target 25 accounts worked and 200 cold calls, with 25 accounts by 75 contacts (three contacts per account). For enterprise accounts, closer to five contacts per account is recommended. Cold call set rate for top reps should be 20%, and sequence reply rates should be in the double-digit range.
Armand lays out four parts of a good cold email: what did I notice about you, the problem you solve, the solution, and a very simple interest-based or offer-based CTA. He adds that a basic four-sentence email got their sequences to double-digit reply rates.
The minimum viable hiring process has four steps total: a recruiter screen, a hiring manager interview (the first real interview, sussing out whether the candidate is smart and cares), a mock cold call run twice, and an onsite consisting of three interviews.
The mock cold call is run twice. The first call is expected to suck; what matters is whether the candidate listened to the podcast, did research on Carter, and tried to put it together. The second call tests whether they apply feedback; if they change nothing, that is a massive red flag. The test checks whether they can be coached and connect dots, since the number one most underestimated thing in SDR is the ability to think, and most people can crank the wheel but very few can actually think.
Average PR size increased from ~44 to ~72 lines. AI models tend to output 'mediocre' code, and long build times incentivize bundling AI outputs into larger PRs. This increases potential bugs, vulnerabilities, and review burden, reducing portability.
moderate
Gary Vaynerchuk advises posting more content on more platforms, saying that is what matters for everyone right now. He warns creative people that demonizing quantity is the great vulnerability of their career, and urges them not to demonize it.
Gary advises creators not to position themselves against AI. His view: 'AI is going to win' — refusing to use it is not a viable strategy. He frames the current landscape as offering two options: use AI to scale, or deliberately choose not to use it and potentially use that choice as a way of standing out. The host adds that AI is currently best at creating more content, whereas interest media rewards creating the best content.
Jeff recommends performing the dumbbell overhead extension leaning back against a bench with the dumbbell overhead, rather than standing, to get better gains and avoid shoulder problems.
The guest advises going organic first to understand what's getting views before post-producing for lower-funnel performance, because organic-first creative outperforms the best performance creative that the team guessed on, and paid media can disguise garbage creative — organic views reveal true creative quality.
The guest describes his process of taking organic content that performed well and adding performance elements to make it performance-oriented natively, using the grandma/granddaughter/lollipop example as the kind of organic content that can be adapted.
Hormozi's first move for a cash-constrained, near-bankruptcy business is to generate cash immediately, starting with an aggressive season pass offer sold to the existing email list, and reinvesting that cash into paid ads to feed the funnel.
The speakers recommend running verification tests over and over in a loop — infinite tests unlock the whole team to move faster because the agent can run the checks itself.
Everyday users should do nothing about AI cryptography risk right now — don't panic, continue on as normal. Keep using current cryptographic systems unchanged; nothing has changed until new information about standard crypto emerges. The worst thing you can do right now is take action.
View the maximum amount of sunlight you can reasonably get in the early part of the day — within the first hour of waking, ideally, but certainly within the first three hours. Take sunglasses off and go outside once the sun is out (blinking is fine), but never look at the sun or any light so bright it's painful. Sunlight exposure to the eyes early in the day increases dopamine receptors, allowing circulating dopamine to have a greater effect on motivation, mood, and pursuit.
Choose sources → to get a library of your own and a daily brief.