Arca's approach: agents on the whole long-horizon outcome end-to-end
Arca put agents on the whole long-horizon outcome end-to-end, rather than on isolated tasks.
35,800 results · Most relevant
Arca put agents on the whole long-horizon outcome end-to-end, rather than on isolated tasks.
Arca's overall design gives agents the primitives for driving long-horizon work: staying aware (intense signal system), knowing what to do next (operational context), doing the work (harness), long persistence (memory systems), and bringing the right people at the right time.
To Farza, 23 views counts as a launch: 'We did something new at the company. We talk about it to the world.' If there's a chance you get more than one view, you should talk about what you did. Making small everyday launch videos was normal in his company because new classes were dropping almost every day, and he hoped one might go viral — which never happened, but he made the videos anyway.
Launching in 2026 means showing your work — that's what launching is. Launching on TechCrunch or the New York Times is showing your work, but in 2026 you can show your work every day and actually build an audience by doing so.
Farza Majeed argues that launching should build a channel you can keep repeating and find users from, not a one-time event. This is his interpretation of what launching is for, grounded in his own experience of turning launches into an ongoing distribution channel.
Farza describes demo-driven development: work backwards from a demo you want at the end, using it as proof the thing is ready to go, then publish the demo.
GaryVee argues that social media is the greatest business opportunity for individuals, and that starting a business with no money and no employees is possible by simply posting about what you already do.
Andy Jassy argues AI will get everybody to the advanced starting spot much more quickly by eliminating manual research and note-taking, but the actual work of putting the dots together — building a strategy, plan, new product, or customer experience — is not something AI will do for most people most of the time; building amazing experiences still requires human judgment, ingenuity, invention, and experimentation, so people should use AI tools as huge aids while critical reasoning and thinking remain very important.
Jassy argues every customer experience is going to be reinvented with inference and AI, and leaders should evaluate how AI transforms each product from the ground up. If you run a successful product at scale and treat AI as merely incrementally improving the experience, you will very likely not have the winning experience three to five years from now. To think more than incrementally, get teams to reimagine applications from scratch — as if building them today with all the new technology available.
endorsed prediction/advice
A model has crossed the usefulness threshold when the time you spend reviewing its work becomes much smaller than the time it saves you, leading to regular use and eventual delegation. Below that threshold, if checking and fixing its work still takes as long as doing it yourself, the model is not yet practically useful. The threshold depends on the work being done and how much checking or human review it requires.
Carmen Li argues that for capital markets and resource allocation the key variable is efficiency. Buying stocks via NYSE or NASDAQ doesn't guarantee the best price, but it is more efficient and liquidity is assured. She applies the same logic to GPU procurement: when you need compute within 48 hours on reserve contracts, going to an exchange with verified compute resources is the easiest way, analogous to going to NASDAQ to buy stocks. Calling 300 salespeople is impractical when you need compute in 48 hours, though talking to cloud salespeople regularly is still valuable because they know a lot of things.
Consequential creator advice requires independent verification before Today can promote it.
The key is consistency: making sure you have facts, not opinions, across your website — opinions are more for humans, facts are more for agents. Agents don't care about images or design; they really only care about facts. This is a very different way of thinking about marketing — marketing in facts, not opinions — it's hard and weird, but you start to get used to how the agents think.
None stated.
The speaker frames the core principle of AEO as consistency: making sure you have facts, not opinions, across your website — opinions are more for humans, facts are more for agents. You have to think a little bit outside the box about what you can say that is truly a fact that the agents will see. AEO is a fundamentally different way of thinking about marketing: it is marketing in facts, not opinions — hard and weird, but something you get used to as you learn how agents think.
Having one centralized surface, place, or queue to monitor everything is beneficial in an agent-led growth motion; a centralized prospector lets an agent filter noise in inbounds/outbounds and ensure responses to current customers, renewals, and support tickets.
Jarmak warns that relying on public benchmark results without testing in your own environment leads to models that may not do what you expect in your actual environment.
Identify what level of risk you are willing to accept before granting agents autonomy. Come up with boundaries for when it is safe to autonomously deploy versus stop and escalate to a human; granting agents autonomy without establishing boundaries for when to stop and escalate to a human is unsafe.
Do not over-index on public benchmark results because they may not do what you expect in your actual environment due to the spiky nature of model training. Public benchmarks are primarily helpful for gaining a broader understanding of the landscape and producing a shortlist of options to explore within your own organization or tooling.
Production is truth — after deploying changes, you need measurements and systems in place to get relevant data in the actual deployment environment. Deploying changes without measurement systems in place in the actual deployment environment means you cannot trust your evaluation values.
The talk's central thesis is to approach AI agents with a research scientist mindset of relentlessly questioning everything rather than trusting them, framed as 'lies, damn lies, and evals.'
AI apps and LLM products need lots of data and context to be valuable, which raises the trust/privacy bar for products running locally with access to the user's computer. Products that run locally and access user data must proactively address what data is accessed, how it's used, retention policies, and security. Products that go above and beyond to make users comfortable about data access, usage, retention, and security will have more success getting downloads and word of mouth.
Choose sources → to get a library of your own and a daily brief.