Eli Portnoy

Key lessons I learned founding 3 startups. 1 Optimize for size of pie, rather than slice 2 Join a CEO group 3 Success hinges on the quality of your team (so this needs to be #1 focus) 4 When things break, look inward first 5 Waiting to do hard thing will only make them worse

There is an AI trap that people are falling for over and over again. They take Claude or ChatGPT, connect it to a bunch of tools like Notion and HubSpot, and assume that the AI now knows everything that lives in there. People are going to get burned. And I don't think most of them understand why. So let's break it down. Three things about LLMs you need to internalize: - They have context windows. There is a max amount of information they can hold in their head at any given point. And the more you cram in, the worse the answers get. The NoLiMa benchmark found that at 32k tokens, 11 out of 12 tested models dropped below 50% of their short-context performance. - They want to please. If you give an LLM three restaurant options with only prices listed and ask "which will be the tastiest?", it will confidently pick one. There is some signal in price. But it is not the totality of signal, and you will end up with a bad decision. A study published in Nature Digital Medicine found that LLMs complied with illogical requests prioritize helpfulness over accuracy, even when they demonstrably knew the premise was false. - They are not great at planning. If you ask it to do something complex, it will take shortcuts unless you force it to map out a strategy and spell out exactly what tools to use. An ICLR 2025 paper called "LLMs Can Plan Only If We Tell Them" found that models struggle to even recognize whether a goal has been achieved. Without explicit structure, they wing it. Now let's apply this to HubSpot. Anyone who has worked with a CRM knows how messy they are. Duplicate entries. Missing notes. So when you ask Claude to go into HubSpot and figure out which accounts are at risk, what actually happens under the hood? It grabs a bunch of stuff. Some helpful, some not. Its context window fills up, so it doesn't grab everything. It likely grabs the easiest info to access, probably user-inputted notes if there are any. It doesn't plan a strategy. It just pulls what it can and gives you an inferred, bad answer. Which is why we are solving this problem at BackEngine - We connect directly to the source. Video recordings, emails, tickets, Slack messages, and CRM. - We organize all of it. Tagged by speaker, topic, customer, type. Every interaction broken into moments. - We dedupe it. And there is a lot of deduping to do. - We apply permissioning so it's clear what data is in scope and out of scope for every user. - We built dedicated agents with guidance on how to solve different types of questions, with clear and precise plans. - We give those agents access to highly organized data and very clear instructions for how to get what they need and only what they need. If you're plugging an LLM into your tools and expecting magic, slow down. Connecting an LLM to your tools is not the same as giving it structured, clean data. The gap between those two things is where bad decisions live.

Shalhevet's win over top-seed Gahr is only the 2nd time Firehawks boys have won 2 gms in state regionals - they reached regional final in Covid-condensed 2021 tourney Now they get a rematch with team that knocked them out of sectionals 2 wks ago, Colony @jhoopsamerica

I have a friend who owns a company that produces very high quality sports equipment. Super smart, very successful. We talk often about business and I appreciate how different our worlds are and the different perspectives we each bring. A few years back, his industry was going through an existential moment. Is there a place for quality in a world of cheap knockoffs on Amazon? Everything seemed to be moving there. But he is smart, and agile, and he faced the challenges head one. And while his world has changed, he has certainly figured it out and is thriving. I bring this up because when he would tell me about these challenges, I would empathize, and I wished he wasn't going through them, but I also felt like they were his problem and not mine. I am in software. We are not getting disrupted. We do not have existential questions to answer. So yesterday when I met up with him, it kind of hit me. What we are going through now in the age of LLMs is similar to what he was going through. There is a big, scary, threat that we need to sort through. And its not clear exactly how this plays out and what it means for us. But it also gave me hope. We all deal with existential questions. And just like he figured it out. Just like he evolved. I think we will too. As long as we face the challenges head on, stay agile, and evolve.

Here is my early hypothesis on how I think the "SaaSapocalypse" plays out. Please tell me where you think I am wrong! 👇 Phase 1: The Wild West (Next 0 - 6 Months) 🤠 We are going to see lots of experimentation and lots of failed projects. Non-technical operators are trying to build full-stack apps and quickly realizing that while AI can write code, it can't architect robust data structures or enterprise permissioning. They are going to learn what it means to support software. To evolve software as the business changes. To support user requests. And to manage permissions, security, and data. We will learn the hard way what we should and should not try to build ourselves. Phase 2: Pro-Serv 2.0 & Opportunity Cost (6 - 12 Months) ⏱️ I think in about 6 months we are going to figure out that there is a very clear divide. Vibe coding UI, agents, and workflows that match what we want perfectly is a great idea, but vibe coding the foundational backend elements isn't worth the risk or the hassle. In other words, I don't want to vibe code my CRM, the data flow between all my systems, and all the infrastructure it supports, but it would be amazing if I could quickly spin up a fully built-from-the ground dashboard on top of it and a bunch of agents that it ran. It would show me EXACTLY what I want to see and do what I want to do, without all the crazy, awful bloat and the insanely bad UI. So we start building a bunch of UI. 1 of 1 UIs and agents just for our workflow. But then we hit a second issue, opportunity cost. A Marketing Director could vibe code their own campaign dashboard, but they really shouldn't. They have actual marketing to do. But vibe coding is now fast enough that you can easily hire someone to do it for less than it costs to pay the SaaS licensing fees. Just like you hire a consultant or an agency to build you an ad campaign (even though you could do it yourself), you'll hire professionals for this. Companies will buy rock-solid headless backends, and pay micro-agencies a fee to build a 1-of-1, perfectly customized UI on top of it. The headless backend manages privacy, security, permissioning, and all of that fun stuff. And the professional services agency maintains the UI, adds features, and so on. Phase 3: The Internal "Vibe-Ops" Era (12+ Months) 🏢 Companies see the massive ROI of 1-of-1 software. It's amazing. It works really well, it does exactly what you need and not what you don't, and it connects all of your different headless systems in a really easy to use package. But a new issue emerges, Shadow IT. All of a sudden there are hundreds of different UIs, and agents, and workflows. And someone leaves, and no one really knows how to support it. We've created a disjointed mess of critical software that no one owns. So companies ban micro-agencies and instead we see a new dedicated role emerge: the Internal Vibe Coder team. This centralized team works internally to understand the business use cases, spins up UIs, workflows, and agents, and maintains and supports them. The corporate tech stack shifts entirely. Instead of paying premium subscriptions for monolithic SaaS, companies will just buy access to secure backend infrastructure and MCP servers. Need a new tool? The internal Vibe Coder sits with your team, maps your exact workflow, and spins up a bespoke UI by Friday. The Takeaway 🎯 ❌ The losers in the SaaSapocalypse will be the traditional, rigid SaaS platforms that force users to adapt to their workflows. ✅ The winners will be API-first backends, MCP server providers, and the people who specialize in translating business problems into bespoke, disposable UI, agents, and the like. Are you seeing teams try to vibe code their own tools yet? Where are they hitting a wall?

Something strange is happening. I am hearing more companies say they want to try and build their own jerry rigged solutions to core infrastructure instead of buying software. I get it. I vibe code all day long. I am building all sorts of tools that make me go faster. But they break. And when they do, I just build them again. They are tools for 1 person. And they are simple. But its different to try to replace core infrastructure that multiple people are using. It might seem fun. It might seem efficient. But your saving a few thousand dollars to build something that will take a massive amount of time to build, maintain, and evolve. And when it breaks, it will literally bring down the company. I just think we're getting a bit too excited too quickly. And we're going to create a lot of needles messes.

Deni Avdija says he’s fed up with the hate he receives for being from Israel, calls out the lack of education on the topic, and demands respect, saying those outside the Middle East shouldn’t speak on it “You don’t have to love what I stand for or how I look, but if I’m a good player, give props. All this hate … for no reason. Like, I’m deciding things in the world. I’ll be honest: What do people expect me to do? This is my country, where I was born, where I grew up. I love my country; there are a lot of great things about my country. But obviously, not everyone is educated and knows what is going on, and that’s what pisses me off. Because if you are educated and know what is going on, it’s fine to say what you think and say who you think is right or wrong. But if you are not educated and you are not part of the Middle East, and you don’t understand how long this goes back and understand the consequences and everything … just don’t say anything.” (Via nytimes.com/athletic/69752…)

Now that Israel's attacks on Gaza have ended but large scale violence is continuing-- perpetrated by Hamas--I will be watching to see how much concern is expressed by pro-Palestine groups. If there is none, it will confirm my suspicions about underlying antisemitic motivations.

This is going to age very well.

@JamCristopher @FlyByKnite @LawMurrayTheNU

8/8 Sometimes the boring explanation is the right one: Not every business decision is a conspiracy. Sometimes it's just VERY bad business 🤷‍♂️ /end 🧵

1/8 The @PabloTorre take on Kawhi/Ballmer/Aspiration is getting wild traction but it's completely backwards and stupid 🤔 Let me break down why this "salary cap circumvention" theory makes zero sense...

🚨 Breaking — Senior Israeli analyst @bardugojacob: Israel 🇮🇱 may sue New York Times for $10 BILLION for the blood libel it spread, including personal lawsuits against the journalists who wrote those articles. _

You are truly one of the biggest demagogues of our time. Truly. I explained how your commentary has been harmful to Iranians and because you are chronically incapable of self-reflection, you respond with an emotionally manipulative deflection about “the children of Gaza,” something I’ve repeatedly expressed sympathy for—only disagreeing about the cause—because every single word that comes out of your mouth is a logical fallacy to deflect from ever having to respond to anything that comes at you. I’m not going to engage you further Dave, because you have proved here and for the past 1.5 years that you are a ruthless manipulator and a dishonest propagandist with a very shallow level of understanding and a lot of sophistry to conceal the fact that you really aren’t capable of engaging complex ideas. That’s why you constantly turn to manipulation and deflection to maintain the curtain that conceals nothing behind it. That’s why you get so fired up and defensive when people poke holes at the curtain, terrified that challenging your ideas will reveal that there’s no foundation of knowledge or machinery of intellectual depth behind it. So you have nowhere to go but emotional manipulation every. Single. Time. Trying to engage you, crack open your narrow mind, show an iota of sympathy for Middle Eastern people outside the ones you can exploit in bad faith, or cause you to reflect on your harmful rhetoric has proven completely pointless. That obstinance proves the opposite of intelligence. I’ve held back so many times from saying what I really want to say and what so many others are thinking. On behalf of everyone who has been hurt by you: You are a massive cunt. With love.

Does this sound familiar to anyone?