36,915 results · Most relevant
Gary Vaynerchuk observes that the content that gets views is often the extremities of random stuff — his blueberries-are-underrated green room video was one of his most successful ever — and that sometimes it's the human stuff, sometimes the emotional stuff (like telling a young woman 'you're somebody now'), and the third type is the tactical: something of value.
Alex Lieberman and CJ Hess are generally bearish on vertical-specific AI solutions in different functions or industries. Spending all day in the guts of enterprises reveals enormous nuance, inertia, systems to wrangle, and people to wrangle. Vertical AI products where an agent is the product doing work in a specific vertical will struggle going forward unless there is a strategic compliance or data moat, as the ability to spin up your own agent quickly compounds and models get better and tasks fall into general model capability. Vertical products that change the process and have an agent on top of an app and tool built around it can still be super helpful and successful.
the speakers' own interpretive stance
Of the roughly 100 enterprises Tenex talks to, 90% have rolled out Claude or Codex to employees, and training/enablement is at roughly 60%. Most people guess 70-80% of enterprises have rolled out coding agents to the engineering org, but the actual number is closer to 50% — every other enterprise either hasn't rolled out a coding agent or what they use is closer to not having one. The biggest diffusion gap is agents in production for non-engineering work, which drops to like 5%; most enterprises with more than 1,000 employees do not have an agent in production doing work that isn't building software.
sample is enterprises Tenex speaks with, not all enterprises
GaryVee argues that in the current landscape the merit of a single post matters more than what a person has done to get to that post, and that a person's third TikTok post — if it is the right piece of content — can fundamentally change the course of their career, while a very good fourth post can lead to 15 to 20 new clients and leads, an employee hire, or whatever they are pursuing.
Because core ExaSearch supports arbitrary semantic complexity, you can write complex predictive features as freely as you want (e.g., 'any adverse media against reputable VPs operating within VC funds who have been on the AI Engineer podcast and said something bad'), but then backtest them to see if they are actually performative rather than assuming they work.
For a product or agent whose job is to define such financial figures, version search lets you go back and see what information existed on the internet before the market figured it out. You can check whether an SEC filing would have revealed a discrepancy in an accounting footnote before the market knew about it, and it can reveal reporting from less reputable vendors that nonetheless found the information first, suggesting you should consider those sources differently. This allows you to potentially trace financial discrepancies before the rest of the market found them.
medium
For financial services customers, backtest new search features against existing processes to prove they are better rather than just believing search will be better; for example, ask what an adverse media scan would have said one month before a portfolio company entered the portfolio, and how that compares to today. Also determine whether a new tool like Exa will add benefit by trying it before committing, and use Exa's Snapshot to backtest financial, risk, or prediction models with web search features to verify they add value.
one atom flagged as possible numeric error (one month before)
Adding one or two agentic features to an existing well-performing model can improve its performance; an adverse media scan runs multiple searches across multiple trajectories in an agentic way, then condenses all results into a structured format.
GaryVee's condition: if you can ship your product (like cookies), live shopping works for local businesses. When you're local and can't ship, live shopping is harder because TikTok's audience is global — a mismatch between the platform's reach and a local service area.
The creator argues that the excuse of 'I don't know what to do' is dead — it died with the internet and is completely finished with AI bots. He says if you're unable to ask basic questions of an AI agent in Claude, Grok, or ChatGPT, then you're just not interested in doing the work; either you're mute and cannot speak or type, or you are not interested in doing the work.
GaryVee argues that live shopping is a major, under-discussed business opportunity and that the right move is to spend the next six months learning the live shopping ecosystem on TikTok and Whatnot, including how to buy at wholesale and sell.
Data ownership and neutrality are so important that large companies should not be beholden to any one lab, especially given labs are building vertical companies that compete with their customers (e.g., Rogo, Harvey).
speaker's strategic argument, endorsed
There aren't enough tinkerers in the world — or people with time to be tinkerers — to help their businesses, and that is why AI is not diffusing. Silicon Valley firms use AI effectively because everyone works next to engineers and tinkerers, while most of the world lacks that proximity. If you can make a non-tinkerer get the same value as a tinkerer in the world of AI, you could be a ginormous business today because AI can solve so many business problems.
speaker's theory and argument
If you can make a non-tinkerer get the same value as a tinkerer in the world of AI, you could be a ginormous business today — the intelligence is already there; the problem is deploying it effectively at scale.
GaryVee says he learned in one year about building interview intellectual property what would have taken 800 hours and took him eight, and tells the audience that once they build their own AI agents and harnesses to handle mundane day-to-day tasks, their capacity to accomplish things will expand dramatically.
When the parent agent must judge output quality and decide when the contract is fulfilled, use a stronger model rather than a weaker one — that's where you get a better result.
While the default is to let agents figure out the how, at certain points you want to weigh in. OpenProse has a feature/syntax that lets you pin the how for specific steps, balancing flexibility with control.
Dan argues that once you can name the thing that you want, any harness can self-shape, fan out, and run any sort of loop to get you there; the question to focus on is what the good best practices are for naming the thing that you want.
If a Prose program's dependencies (skills, CLIs) can't be found, don't run the program — and don't fall back to alternatives. This preserves the integrity of the declared contract.
The opportunity in the AI era is larger than ever because you can now access labor budgets, which are much larger than software budgets. The average business spends much more money on salaries than on software. AI can 10x businesses that prior to AI may have struggled to grow 2-3x in the same time frame.
Choose sources → to get a library of your own and a daily brief.