Omri

“The real moat is the culture that makes the product”

the next status symbol is a device that does less

22M monthly users, built out of toronto. we don't brag about this enough

the $1.6B got the headlines. the 1,500 ai jobs staying in canada is the part that matters. good week to be building in toronto

my test is simpler: if a customer could describe your product in one prompt, you're not a business yet, you're a feature

an animal that eats a solar panel and keeps it running. nature figured out acquihire before silicon valley did

the real flex isn't the code, it's using 20 years of other people's taste as the recommendation algorithm

kinda crazy that stars fuse hydrogen into heavier and heavier elements, and their cores slowly fill with iron, the one atom that gives back no energy when fused, until the core collapses in a fraction of a second and the star explodes, scattering that iron across space, where some of it clumped into a planet, then into rock, then into food, then into a molecule in your blood that carries oxygen and the atom that starved a star of energy is now the one that keeps yours going

😭 so sad about Patrick Stewart. as a Gen X kid he shaped how i see the world and who i am, including amplifying my love of Shakespeare. in the age of AI, we need Picard's belief in what humans can become now more than ever. "i see us one day becoming that" perhaps my favourite scene, watch the clip below 👇

the real market isn't people who love ai. it's people who just want the invoice sent and to go home. build for them and never say the word agent

AI is book smart. Your leadership team still has to be street smart. Mark Cuban recently described AI as “book smart,” but not “street smart” or “entrepreneur smart.” Someone still has to know where to point it. What stayed with me was his challenge to leadership: do we understand the technology well enough to recognize where it could change our business? And do we understand our business well enough to make those choices? Every competitor can rent many of the same models. What that subscription doesn’t settle is which customers to pursue, which problems are worth solving, or what the company is willing to give up to get a better result. Those decisions shape whether AI actually improves the economics or simply makes the existing activity happen faster. Take something as ordinary as “a good customer.” Sales might mean a large contract. Finance might mean healthy margins and reliable payment. Operations might mean someone whose needs we can meet without constant exceptions. Ask an agent to find more good customers and you have to confront those differences. Otherwise, it could become very effective at bringing in business that looks attractive to one department and creates problems for another. I’d start by putting a few actual customer decisions in front of the leadership team. Which would we make again? Which looked good at the time but cost more to serve than we expected? What did we miss? Use those cases to agree on the trade-offs, define what a better outcome looks like, and decide who can approve an exception. Then look at the incentives. If sales is still rewarded entirely on revenue, a more thoughtful definition in the AI’s instructions won’t resolve the disagreement. I’ve been writing about the things companies should own: memory, context, specialized intelligence and authority. Judgment runs through all of them. But putting an experienced executive’s instincts into a system doesn’t automatically make them sound. AI can help us examine those instincts, find inconsistencies and test alternatives. Leadership has to be willing to learn from that process, including when it challenges a decision we’ve defended for years. Maybe AI will expose management disagreements we’ve been able to live with because people quietly worked around them. Once those choices guide an agent’s actions, the consequences become harder to ignore. The leadership job is to resolve the trade-offs, change what we measure and reward, and take responsibility for the results. That may require more change from the people directing the AI than from the people using it. Source: nitter.cf/a16z/status/2115466241…

kinda crazy that the sun lifts over a trillion tons of water off the ocean every day, the wind carries it inland, it freezes onto a mountain, and gravity spends millions of years pulling it back down as a river, one grain at a time, enough to cut a canyon a mile deep, carrying the mountain itself to the sea we call it a river, but it's really just the ocean's way of taking back the mountain

When someone corrects an AI in your company, who learns from it? The person having the conversation may get a better answer. That doesn’t necessarily mean the company gets better at doing the work. Ben Horowitz and Michael McGuiness explore this in Databricks & the Future of the Firm. They describe a “system of record for agents” that goes beyond storing company data. It also captures what that data means, what an agent is allowed to do, and how the company judges whether the work was any good. Their argument goes further: if people closest to a problem can access reliable knowledge and have the authority to act, more decisions can happen there. Less information needs to travel up and down the management chain. That could change how companies organize work, and how much time people spend waiting for someone else. I’ve been saying “rent the model, own the memory.” This article made me think harder about the second half. A folder full of documents won’t necessarily tell you which policy takes precedence, why a customer was given an exception, or whether that exception still applies. Those gaps matter when we ask AI to act on what the company knows. Imagine an agent applies standard renewal terms to a customer whose agreement includes an exception. An employee catches it and fixes the response. The practical work starts by giving that correction somewhere to go. Capture the contract clause, the reason for the exception, who approved it, and when it expires. Someone responsible for renewals then decides whether this changes the general rule or applies only to that customer. The approved decision belongs in the record people and agents actually use. Then turn that correction into a test. Can the agent handle this customer correctly, apply the standard terms to another customer, and recognize when the exception has expired? Run those cases again when the policy changes or a new model is introduced. A company could build a useful collection of these cases from the mistakes employees already spend time fixing. Each one would preserve both a piece of judgment and a way to check whether the system can apply it. That gives management a basis for deciding how much authority to delegate. Start with a narrow set of decisions, review the outcomes, and expand the agent’s authority as the evidence supports it. Keep unfamiliar cases with someone who can exercise judgment, then use those decisions to improve the rules and tests. The difficult work is deciding which lessons are reliable enough to become policy, and which still require a person. I’d want a named business owner accountable for that boundary — and for moving it as the company learns.

'expected behavior' is the scariest phrase in software. expected by whom?