Use web search as the primary tool for grounding AI agents
Richard Socher recommends using web search as the number one most used tool for grounding language models, agents, and chatbots in real-world knowledge.
Contested by Anthropic on AI Engineer
46,002 results · Most relevant
For understanding everyday Spanish through listening and conversation.
Richard Socher recommends using web search as the number one most used tool for grounding language models, agents, and chatbots in real-world knowledge.
Contested by Anthropic on AI Engineer
Richard Socher recommends using red teaming or rainbow teaming, where one AI tries to hack another AI in an open-ended fashion, to inoculate against attacks.
Richard Socher recommends phrasing every NLP problem as a prompt with text context, question/task description, and output to unify tasks into one neural network.
Richard Socher recommends testing a model's extrapolation ability by training it only on data before a certain date and then evaluating its performance on later events.
Richard Socher recommends using arXiv to publish papers without gatekeeping, then relying on community discussion (Twitter/X, word of mouth) as a filter instead of expert peer review. He argues that if you tell 10 friends about a significant breakthrough, someone will talk about it again.
Use multiple data sources (e.g., Crust Data, Clay, ZoomInfo) to build a dynamic data system instead of treating CRM as static. Don't treat CRM data as static; treat it as dynamic and continuously update it with multiple data sources.
There are stakeholders in your deals that you do not know exist (e.g., VP of Finance viewing your champion's recap) who can kill the deal at the finish line. Multi-thread to unknown stakeholders surfaced by AI before they can kill the deal.
Rob Walling advises that for direct snail mail outreach to work, you need a high enough annual contract value to justify the cost and significant cash up front to fund it, as you will send many letters or gift boxes that don't work as you optimize your process.
Rob Walling recommends mitigating platform risk by having a backup foundational model provider to switch between if the primary API goes down. He advises assessing the likelihood of AI regulation and API downtime as part of risk tolerance analysis. He notes that AI credits are currently subsidized (roughly 20% of true cost), and there is a risk of price increases if AI funding crashes, though costs may also decrease over time. The risk of AI being regulated out of use in B2B SaaS is considered low.
medium
Gonen Stein and Ofir Ehrlich recommend that data-rich companies sell their data to AI labs, and that AI product builders buy enterprise data upfront to train models before customer interaction. They cite Google's purchase of Spirit Airlines data as an example of this growing trend.
medium
Hodak recommends that BCI communication devices require explicit volitional intent from the user to avoid false activations, such as accidentally summoning an Uber during a meeting.
Hodak recommends prioritizing the retina as the neural interface target when the patient has an intact optic nerve, citing scientific and technical reasons.
Consequential creator advice requires independent verification before Today can promote it.
Jeff Cavaliere recommends fully locking out the elbows on triceps exercises to recruit the medial head for wider-looking arms. He advises against cutting short the final inch of the triceps lockout, as it is the hardest but most important part of the exercise. For the triceps pushdown, he recommends placing the bar further out in the fingers and letting the wrist bend back casually to take the forearms out of the movement and achieve full lockout. In week two, the audience should add 5 pounds and use rest-pause to get all 12 reps. In week three, they should add 10 pounds from week one and use rest-pause to get all 12 reps.
Jeff Cavaliere identifies several common mistakes: swinging the dumbbell way out in front of the body during a curl, which invites the shoulder to do some of the work, reducing biceps tension; having hands facing forward during bench dips, which creates internal rotation at the shoulder; and cutting short the final inch of a triceps lockout, which limits medial head recruitment.
Jeff Cavaliere describes the rest-pause method: go for as many reps as possible, rest 10-15 seconds, then continue until all reps are completed.
The host advises letting your full reaction through when you hear a funny joke or a surprising question. This will make the other person feel heard and, more importantly, prove to yourself that your unfiltered reactions are good enough, which is the essence of secure confidence.
The host advises naming your struggle first to give others permission to share. This deepens connections and makes you the conversational leader. By talking about the hard stuff without flinching, you show that you don't need to be perfect to feel good about yourself.
The host recommends training your 'silliness muscle' by telling your own embarrassing stories—ones where you didn't look perfect but can now laugh at. He warns against overusing this type of humor, as it can become another form of identity protection.
Jeff Cavaliere recommends the dead arm hang from a pull-up bar for spinal decompression, low back pain relief, shoulder flexibility, and posture improvement. He advises performing a scap pull every 5 seconds to engage the lats and retract the shoulder blades. If 5-second intervals are too difficult, start with 3-second intervals. Continue for as long as possible.
Jeff Cavaliere recommends the creeping lunge to develop glute medius strength, which controls the pelvis and prevents low back strain. The exercise involves stepping back with one leg from a squat position and returning slowly, alternating legs. Beginners can put their knee down.
Choose sources → to get a library of your own and a daily brief.