For CEOs struggling with where to start with AI adoption, I have good news, bad news, and a suggestion. First, the good news: Everyone is feeling extreme FOMO compared to everyone else who seem to be having fabulous success with AI. Just a reminder, social feeds are not reality. I know exactly 2 business people building deep, challenging software and they are the kind of people you'd expect: always with a new hobby, a new obsession, surprisingly deep in the oddest areas. If that's not you, that's ok, but don't expect software engineering to suddenly be the moment you find your inner tinkerer. Second, the bad news: When most CEOs talk about AI, you could swap the word "AI" with the word "magic" and not lose any information in the sentence. Of course it'd be great to have this magic thing achieve a great business outcome, somehow, at a low cost. But that doesn't exist. So let's reset. In an enterprise setting, most AI hype is around the english-to-code-and-back translation capabilities. You can take ideas articulated in English and get code out and vice versa. So why isn't that magic unto itself? Simple: translation doesn't improve the inputs and with an LLM can make things worse. What this reveals is that software engineering is no longer the bottleneck. The actual bottleneck now is software architects and it's more acute than ever given the newly infinite supply of software engineers. The teams with great architects are the ones running away with the AI windfall. To convince yourself that architecture-as-bottleneck is true, imagine building a house with infinite high-skilled laborers but no architect. Day 1 would be really easy and you'd feel like you made a ton of progress. But after a few days, you'd run into all kinds of issues, interdependencies, rework, and things would grind to a halt. You'd never build a house without a plan and you can't build software that way either. Third, a suggestion: My view is that being a good architect is a cognitive style more than a skill set. You likely already have a person inside of your org who thinks this way: radical systems thinker, loves process, always generalizes, reduces things to their core essence. Obsesses about beauty, simplicity, elegance. They could be anywhere in your org, but ops is a good place to look. One easy way to find this person: ask everyone what they are building with coding agents. See who surprises you. If nothing obvious surfaces, ask people about their hobbies. Look for a self-taught tinkerer. This person WILL need coaching, but they'll also know how to use AI to level themselves up. I'd find this person, free up a bunch of their time, and see what happens. It's worth a shot and something you can do today. You might just conjure up some AI magic of your own.
Confirmed, claude code ultracode will crush tokens.
I hit a breaking point. 5+ terminals open with the same title. scrolling up to remember which was which. Getting lost in gits, commits, stashes, stomps, pushes, pulls, deploys. So I created a message bus where agents can talk to each other and leave notes about what they are working on and how to deconflict. And a process for them to manage all code changes and merges behind the scenes after I approve the feature build itself. This is a pattern. Outgrow my current stack little by little, hit a wall, then productize the thing i'm struggling with. Super fun. And also really strange why something for a non-engineer prosumer doesn't already exist.
Real estate and mortgage behave more like defense than finance. The incentives all point the same way - higher prices: - Real estate and mortgage are 'ad valorem' industries - 'Ad valorem' has the same incentives as 'cost plus': to inflate prices - Real estate is + 6% - Mortgage is + 3% (but actually +9% fully-loaded) - Real estate and mortgage are highly regulated which limits innovation - Only the buyer has the incentive to want lower prices - So, federal debt subsidizes mortgage credit, supporting buyers and further pushing up prices Like work inflating to fill available time, house prices will always inflate to soak up any excess income/savings consumers have. Net net: AI may yield margin enhancement for real estate and mortgage companies and their tech providers plus some cool new user experiences, but I don't see how the consumer will see any affordability benefit.
I made a big change to my workflow recently - added the concept of a "workbench". A workbench is a deep dive into a topic, company, person, market, etc. where I want to do deep agentic research, aggregate data, synthesize, build reports, and create recurring processes. I was finding that I was going deep into different areas using the tools I've built, but then ending up with a bunch of small folders or repos splintered off from the main product. So now my stack looks like: Raw data artifacts, synthesized into A database, which enables A workbench of agent-led analysis, distilled into Context documents about topics, people, companies, etc., which power A chat product that knows everything I know, and never needs reminding Each workbench maintains its own state and mini kanban of ideas, next steps, open questions so when I ask the agent to resume, it picks up exactly where I left off (even on new threads or with different providers/models). I interact with the workbenches using Codex (much better UI and GPT 5.5 has been much stronger for me recently that Opus 4.7). I've also built a componentized app layer that I can plug into any of these workbenches to visualize or interact with the data itself. So what has emerged is a set of personal apps that all sit on my personal, local shared data and infrastructure. Basically G-Suite, but my own, private, and 100% future-proof because I can change model providers anytime, or go local / open source when it gets good enough, all with a one-minute code change. Doing all this building has made me think that personal AI will be all local data - you are the only person you will trust with this level of data across your life and it prevents provider/model lock-in. Local data with no lock-in, though, is something foundation models or hyperscalers won't build and to me explains why we're left with fragmented product experiences from big tech that feel surprisingly bad.
I took things to their logical conclusion (late) last night ... made an Adam agent that acts as the reviewer at each stage-gate in the build process. Got tired of saying the same things over and over ... and finally realized the obvious thing to do. Trained the agent on all my historical coding agent chats (which I export as logs, not default behavior, so a good tip). Here we go :)
Couple things I've done recently which have been helpful: 1/ I created a sitemap of the product I'm building. Coding agents start each session by reading the sitemap, then do the work, then update the sitemap at the end. 2/ I don't do ANY work now in a coding agent that isn't a story or a defect. This creates structure, memory, traceability. Ad hoc work seems crazy now in retrospect. 3/ Left myself a 1 hour voice note about product roadmap, dumped the transcript into claude code, which turned it into an optimized sequence of stories. 4/ Maybe the most powerful, I added a set of recusrive learning stages where the agents mine the build history for learnings, reflect, and update themselves. 5/ Which surfaced that the agents were cheating - a lot - so I went back to the Sitemap and tagged certain parts of the product as 'signal extraction' = use code and 'intelligence' = use reasoning for force the LLMs to think in the places where I wanted them to actually think. This has helped a lot.
I've had a few friends reach out to me to talk shop about AI and how to transform their orgs. Is AI coach a thing? Regardless, if you're thinking about how to kick-start AI use within your org, feel free to DM me. Happy to chat.
Couple helpful notes I've found for building with agents: 1/ I created a team with 5 personas: orchestrator, codebase expert, researcher, builder, qa agent. Each is assigned a notable public figure as a mentor to model their thinking after (eg researcher = Richard Feynman). 2/ I kick off the build process with a clear brief. The better my thinking the better the outcome. Slop in slop out. 3/ I created a build process with stages: research, scope, plan, build, qa, retro. Each stage has exit criteria that agent must prove is met and I must sign off on before moving to the next stage. 4/ The retro iteratively improves items 1 and 2 and 3 5/ Everything is tracked in a hashed log (not context file that gets overwritten) to create cumulative learnings. Still working on releasing an open sourced version. Working on productizing everything e2e so it's dead simple. Having a ton of fun doing it!
We’re not hiring traditional software engineers anymore. We’re hiring Architectural Operators — people who design systems that build software using AI agents, rather than writing code themselves. We’re looking for engineers who think in systems, move extremely fast with AI, and prioritize architecture and reproducibility over writing more code. Full details + challenge inside ↓
I've been having a blast with Claude Code. I'm building a few things right now, but here's one that's probably relevant to most people — I call it Dojo. A personal AI operating system. Six apps sharing one database. Every piece of data I generate flows into it: emails, texts, meeting transcripts, calendar events, documents. All indexed, vectorized, and retrievable in context. The core is a unified chat interface. Every major LLM sits behind one UI, and you can toggle between models mid-thread without losing history. Built on the enterprise API layer, so I own the data with full retention and no training on my inputs. Every conversation compounds into a knowledge base that grows with each use and never forgets. The LLM already knows about people, companies, and history from the graph, but if I want to load a specific call transcript, email thread, or wiki page into the conversation, I just @ it in. Chats connect directly to tasks and wiki pages, so every unit of work lives together in one system instead of scattered across a dozen providers. People and company graph scored by real interaction strength across every channel: email, calendar, calls, texts, etc. Feed it a name and it pulls the full relationship: how we met, what we've discussed, every touchpoint across years of communication. Each person and company gets an auto-generated wiki, updating in real-time as new data flows in. Now that most work calls are recorded, the amount of structured knowledge sitting in transcripts alone is massive and this approach pulls them out of their silos. Runs on a single Mac Mini. No cloud infrastructure. No SaaS fees beyond the API calls (very low after Claude Code $200/month). Next step is local models on dedicated hardware — all data on-device, open-source models, zero external dependency. That's where the world is heading. I'll be open-sourcing Dojo and a few other things over the coming weeks. Drop a comment if you want to get notified. FWIW, my social media usage is down 95%. If I have 10 minutes, I'm adding features...
State-of-the-art frontier AI models are <<$10 per hour if you use them heavily non-stop for the full hour ... That's insane for having a genius side-kick. You aren't using them enough!
I've been having a blast with Claude Code. I'm building a few things right now, but here's one that's probably relevant to most people — I call it Dojo. A personal AI operating system. Six apps sharing one database. Every piece of data I generate flows into it: emails, texts, meeting transcripts, calendar events, documents. All indexed, vectorized, and retrievable in context. The core is a unified chat interface. Every major LLM sits behind one UI, and you can toggle between models mid-thread without losing history. Built on the enterprise API layer, so I own the data with full retention and no training on my inputs. Every conversation compounds into a knowledge base that grows with each use and never forgets. The LLM already knows about people, companies, and history from the graph, but if I want to load a specific call transcript, email thread, or wiki page into the conversation, I just @ it in. Chats connect directly to tasks and wiki pages, so every unit of work lives together in one system instead of scattered across a dozen providers. People and company graph scored by real interaction strength across every channel: email, calendar, calls, texts, etc. Feed it a name and it pulls the full relationship: how we met, what we've discussed, every touchpoint across years of communication. Each person and company gets an auto-generated wiki, updating in real-time as new data flows in. Now that most work calls are recorded, the amount of structured knowledge sitting in transcripts alone is massive and this approach pulls them out of their silos. Runs on a single Mac Mini. No cloud infrastructure. No SaaS fees beyond the API calls (very low after Claude Code $200/month). Next step is local models on dedicated hardware — all data on-device, open-source models, zero external dependency. That's where the world is heading. I'll be open-sourcing Dojo and a few other things over the coming weeks. Drop a comment if you want to get notified. FWIW, my social media usage is down 95%. If I have 10 minutes, I'm adding features...
State-of-the-art frontier AI models are <<$10 per hour if you use them heavily non-stop for the full hour ... That's insane for having a genius side-kick. You aren't using them enough!
Using Claude code feels like being a kid again - magic. I say that without hyperbole. If you haven't tried it, you really need to. For a non-technical person, it'll be the easiest way to get a gut-level feel for the transformation that AI is going to bring. Use Claude in your browser to explain to you how to set up Claude code.
The 50-year mortgage discussion misses the point entirely. It's money printing by another name. I recently walked a family member through mortgage optimization. Running just 4 basic scenarios across a single lender, I found alternatives that save 5-10% of the house price over 7 years, in cash! The actual solution isn't creating more artificially cheap credit. It's fixing the broken quote process. I've documented all the levers and the math. Use this as a roadmap for your next mortgage. Link below ↓
Compromise kills the protocol flywheel. Protocols are counterpartyless, code-native, eternal. VCs need counterparties, documents, and exits. A structural mismatch runs deep between how protocols operate and how capital flows. I propose Bitstrapping—a new way to raise capital, aligned and onchain. ↓ Link below
Protocols are going to eat startups from the bottom, exactly like startups did to corporations. First solve problems even startups can't economically justify, then move upmarket. Building Elephant Protocol showed me why. Decentralized protocols are at least 10x more efficient across five operational factors: talent acquisition, tool selection, iteration speed, cost structure, error correction. People often ask why we built Elephant Protocol as a protocol instead of a startup. That's why. Some problems only make sense at protocol economics. Wrote about how the efficiency compounds and what it means for venture-backed companies. Link below ↓
Prediction markets would predict fire risk better than insurance premiums. Participants would get granular—roof condition, lot slope, wind direction, distance for fire hydrant, etc. But insurance premiums do the opposite. Two houses on the same block pay the same price. One burns, one stands. The homeowner was deprived of valuable, possibly life-saving information. Elephant Protocol data enables insurance companies to act like prediction markets, which we all know is how they should operate. https://lnkd.in/ed_anxek
A major P&C insurer tested Elephant Protocol data on their hardest underwriting problem: roof age in disaster zones. One county in Florida revealed $67M in preventable Hurricane Ian losses. https://lnkd.in/e3ghsYTJ
Elephant Protocol just put Lee County's entire permit history on-chain. Here's why that matters for every insurance market in America.
We have an intuitive sense that using travel agents or insurance brokers is both slower and more expensive than DIY. Why go through all the tedious effort to share your information and the nuances of your preferences with someone else just to have them re-type into a computer and charge you for the service? LLMs + onchain data will (very soon) do the same for real estate agents and mortgage brokers. Why pay someone 6% if you can easily DIY using an LLM who has perfect knowledge, data/systems access, and never flakes on you? Tell me I'm wrong!
Soofi Safavi and I have outlined the 1st principles problems in real estate in mortgage on pages 3 - 6 of the Elephant Protocol white paper. If you understand these few issues deeply, you're ahead of industry insiders who spend 100% of energy keeping the Rube Goldberg "working. https://lnkd.in/eB_j-eqN
If you buy a house, get a mortgage, and sell the house 7 years later ... you will pay which of the following (as a % of your house price) in total transaction fees: A) 6% B) 9% C) 12% D) 17% The correct answer is D, 17%. These numbers 100% tie to the BLS and NAR data. The common refrain is that "housing is broken". In fact, the system is working exactly as designed since it was designed by intermediaries who make a living maximizing and concealing this offensive total fee amount extracted from consumers. Over the next 12 months, Elephant Protocol will put all 150m residential properties on chain and give consumers the ability to transact without intermediaries - and their fees.
The killer app here is a mortgage that fully circumvents the current GSE system -- a loan fully collateralized by staked BTC, without a cash downpayment. Staking can progressively be reduced as the BTC growth outpaces the price of the house. Easier access to credit without underwriting and consumer benefits from BTC appreciation >> house price appreciation. Assets in reserve are nowhere near the limiting factor in access to GSE or jumbo mortgage credit. Volatile assets anyway don't really make strong reserves after the actual origination of the mortgage.
If Compass really cared about the consumer, they would argue for an open listing service. Their lawsuit is just internal squabbling between members of a distribution cartel about how to divide up the spoils of overcharging.
Mortgage rates are rising ... did you know putting a cap on your rate costs about 100 basis points (bps) of the loan value as a pure fee? That's insanely high. Can anyone explain why it cost lenders 100 bps to hedge against rate movements? I would love to speak to someone in a secondary department if anyone has an easy intro.
Housing -- costs, quality, choice, liquidity -- is the single issue with the biggest potential to unlock economic value and life satisfaction for Americans. Using unit transactions as a gross proxy for market depth and health, we haven't seen such a poor market since 1995.
Really creative book by George Mason University Prof Bryan Caplan on why housing could easily be affordable if it weren't for so much regulation. There's absolutely nothing special about housing that prevents it from being in abundant supply and at low prices ... I think the form factor of the book also tells us something about how we should think about reaching and talking to consumers in an attention economy.
Following the VP Debate commentary on housing ... to make housing more affordable, the simple answer is to build more housing. More housing is strictly beneficial to house prices at all levels. Reducing transaction costs (ie red tape) to build will benefit Americans at all income levels and at all house price levels. Subsidizing downpayments simply papers over the problem and will be inflationary with no benefit to anyone as prices will adjust up accordingly. Great podcast here on the topic with supporting research that steps through popular counterarguments to building housing as the key solution to creating affordable housing. Michael Munger https://lnkd.in/g9KXhWq8
A Clear Example of AI's Disruption to SaaS and Workflow Software At Staircase, we built a fully AI-powered CRM in just 1 engineer-week and now $0 ongoing maintenance costs. First, we integrated all customer interactions—emails, calls, SMS, website visits, mailers, account data—into a unified data structure linked to each customer ID. Next, we built a custom GPT using OpenAI, designed to act like an expert salesperson. Now, anyone on our team can ask the AI: - What key customer activities happened yesterday? - What follow-ups are needed today? - Which customers are satisfied or dissatisfied? - Provide a full history for John Doe. The GPT then crafts tailored responses for each customer, considering their specific history, tone, and language. Our team can iterate on these drafts with the AI and then send the final message via email, SMS, or social media DM from within the GPT itself simply by saying "send as email" or "send as text". This ensures we send the right message at the right time through the right channel, all within a single, easy interface. Way, way more business value than a tradition SaaS CRM.
Ever wonder what's really in your mortgage rate? It's more than just a risk-based rate. Fees and commissions are layered in, adding an extra 20% to your total cost. For a $1M loan, that’s $283K more in interest you could avoid by paying the fees and commissions directly! Time for more transparency?
Yesterday, the Fed cut rates by 50 bps, but mortgage rates remained flat to slightly higher. Why? Markets had already priced in the cut. However, strong economic data (housing starts, jobless claims) pushed the 10-year Treasury up by ~5 bps. Given that mortgages are typically held 7-10 years, the 10-year T-bill serves as the benchmark rate for mortgage pricing.
Did you know that for every $1 your lender makes you pay $13 in extra interest that could be easily avoided which adds up to 23% of your home price? Here's the breakdown: There are three mechanisms lenders can use to make money on a mortgage: A) Charge you a transparent fee B) Roll fees into the loan amount C) The interest rate spread For each $1 the lender makes, here's how much hidden cost the mechanism itself drives to the borrower: A) $0.00 extra cost B) $1.20 extra cost C) $13.10 extra cost This is holding lender profits constant, so this waste is purely a choice by lenders. Yet, lenders always choose the spread because it's hidden from view and they all understand the game theory of acting together. As a consumer, think about this: your lender makes a $20,000 commission on your $1m loan and your lifetime interest goes up by $260,000 simply because they don't want you to see their fee. That's 23% of your home price of pure waste. Note: These numbers are for a full 30 year hold. The extra cost is half as much for a 10 year hold.
Mortgage payments have skyrocketed to 2.2x their 2020 levels, but incomes have only grown 1.2x. Right now, tight supply is keeping prices high. If inventory grows as interest rates decline, prices should fall and perhaps very quickly. Eventually, life needs will force people to move—even at a lifestyle markdown. https://lnkd.in/eK8RhHbg Rukmini Callimachi #realestate #housingmarket #mortgages #interestrates
Excellent book on the future of modeling and prediction, particularly relevant for housing prices. Chapters 4-6 convincingly make the case that housing prices are inherently and primarily cyclical due to real estate agent pricing strategies, independent of external factors like interest rates or lending policies. In terms of game theory, the "price high, then cut" approach has several flaws: - Creates a deflationary effect, making it optimal to wait - Sets off a feedback loop, reducing perceived value - Leads to a thin market by targeting top bids I've always been skeptical of the Zestimate which use price to predict price. Farmer’s ideas on agent-based models offer more robust and intuitive approach. Doyne Farmer
Simplest way to lower housing costs? Eliminate the 'interest rate spread' that lenders tack on to loans. By forcing lenders to charge fees transparently instead of hiding them in the interest rate, monthly payments could drop 10%-15% overnight. Lenders would have to operate more efficiently — and that’s a win for consumers. Check out more details: https://lnkd.in/eR4XDfkF
The two core remedies in the Sitzer-Burnett v. National Association of REALTORS® class action settlement are not going to meaningfully reduce fees for buyers and sellers (h/t to the NAR lawyers). The settlement's two core remedies are hugely ineffective: First, you can't defeat a pricing cartel by simply asking for lower prices; they have market power. The core remedy of the settlement is to require borrowers be reminded of their already existing legal right to negotiate. Consumer negotiations will be ineffective not least because the game theory for agents to keep fees high is so well established. My personal view is that the core function of the NAR as well as the aggregation of real estate agents into national brokerages is to facilitate and enforce this pricing coordination. Second, the settlement required removal of the commission amounts from the MLS itself. All this will do is create inefficiency as agents are expected to *call* each other for manual commission discovery. This just creates a further drag on productivity with no benefit. Here's a rational way to think about agent commissions: Seller Fees: The truth is that most homes would sell without a real estate agent when priced at 80% - 90% of the appraised value. This is why banks cap lending at this amount because they have high confidence a house priced at this level would sell very quickly. It follows then that commissions on this 'baseline' amount of 80% - 90% of the appraised value should be 0%. With current pricing, the 6% agent sell side fee captures a huge portion of any gains above the 'baseline' price for the majority of houses that do sell near appraised value (ie at 100%). Buyer Fees: The correct class action settlement would simply have been to ban the practice of sellers paying for the buying agent. Each agent should be paid directly and transparently by their own customer. In this world, buying fees would immediately drop since buyers would suddenly be faced with having to pay 3% to agents for listings the buyer themselves found on Zillow. When the 3% buyer's fee is made that stark, no one would pay it. Many agents will say that the seller paying the buyer fees helps homebuyers since it allows buyers to use all of their available cash solely on the downpayment, mortgage, and closing costs. The simple legislative fix would be to allow buyers to roll agent commissions into the loan amount. This fixes the problem directly and disarms the NAR backwards rationalizing of their anti-competitive commission structure. Fee Mechanisms The buyer/seller should have to pay agents directly and not via net mechanisms at the closing table. Sending a $50,000 wire to an agent (equivalent to a @Tesla Model Y!) feels very different than $50,000 coming off the top. Agents know this, of course ... Mary Ann Azevedo https://lnkd.in/e8YWJbz7
Yes, this is progress, but this misses the larger point that the entire housing system is designed to extract maximum commissions and fees from homeowners. It's a feature, not a bug. The homeowner pays, as % of their home value: 3.0 % buy side agent commission 2.0 % loan officer mortgage commission 0.5 % title insurance 3.0 % excess interest cost 3.0 % sell side agent commission Plus >$5,000 in ~20 different "junk" fees. In the deepest, most liquid real estate market in the world, it costs you 11.5% of your home value simply to transact in and out. The average home price is roughly $350k, so these transaction costs eat up more than a full year of after tax income. THAT is the big story. https://lnkd.in/ePiJrWng
We integrated #ga4 with #openai so that every aspect of #ga4 is now available via a prompt. Here are a few key use cases: Intuitive Queries and Custom Reports Simply ask your question in natural language: "What products are trending this Black Friday?" or "Show me the user demographics for Cyber Monday sales" and get instant, tailored responses. Make swift, informed decisions. Real-Time Data Analytics Our GPT provides real-time analytics whereas many dashboards within GA4 operate on a lag. Stay on top of the details. Predictive Analysis Forecast trends and customer behaviors, preparing you to meet demand, adjust strategies, and maximize sales opportunities. Be proactive vs reactive. Simplify Complex Data Turn complex data sets into clear, actionable insights. Understand your customers' journey, preferences, and behaviors in a straightforward, easily digestible format. See the matrix. You can see an example of our Google Analytics GPT in the video below. To spin up your own GA4 GPT, DM me. #blackfriday #cybermonday #ecommerce #ga4 #googleanalytics #googleanalytics4 #openai #llm #gpts