Michael hochstat

Knowledge is free knowledge.dotadda.io/session…

The next Salesforce competitor might not look anything like Salesforce. David Sacks just explained something important about the future of enterprise software. Salesforce is allowing Claude to become the front end for its CRM. Think about that. Instead of logging into Salesforce, navigating dashboards, opening records and building workflows, users can interact with Claude. Claude accesses Salesforce data, executes workflows and takes actions through the existing system. Salesforce still provides the infrastructure, data, permissions and business logic. But the customer interaction is moving somewhere else. This is the bigger SaaS question: What happens when the application is no longer where the customer works? The architecture could look something like this: Data → Systems of Record → APIs/MCP → AI Agents → User And this isn’t theoretical. Salesforce and Anthropic have already launched their partnership, with 37 prebuilt sales skills. For anyone building software, this changes the competitive landscape. The value may increasingly move from the application interface to: Proprietary data. Business logic. Context. Permissions and governance. Reliable execution. The application doesn’t necessarily disappear. Its role changes. And here’s the part I find most interesting: If Claude becomes the interface for Salesforce, what stops it from becoming the interface for every other SaaS product? Who owns the customer relationship then? Who owns the workflow? And who captures the economics? The next generation of enterprise software may not be another application. It may be an agent that operates across all of them.

That’s actually an important distinction. You’re not anti-AI. You own an AI company and use AI to produce your own content. Your argument is much more interesting: I’m not criticizing AI. I’m criticizing the way companies measure its success. I use AI. I build with AI. I own an AI company. But I also know the difference between generating something and producing something valuable. If AI writes my LinkedIn post, that’s productivity. If it invents a statistic, that’s a liability. If it produces a research report, that’s output. If the report gets the investment decision wrong, that’s a problem. And if I have to spend more time correcting AI than it saved me, that’s not productivity. The point isn’t to slow AI down. It’s to stop confusing AI adoption with AI results. I believe in AI enough to demand that it actually work.

First it was AI confusing cats and dogs. Then it started hallucinating facts. Now it’s writing code faster than humans can fix it. What’s next? AI will probably automate the mistake, write a report explaining why the mistake was actually a success, and generate a PowerPoint to convince management. Meanwhile, someone will be laid off for asking questions. The technology keeps getting smarter. Apparently, the definition of productivity keeps getting easier.

I checked the Reuters investigation and the hiring-manager survey. Your central argument is supported, but a few details need tightening. The strongest part is the Meta data: AI-generated code increased dramatically, but so did technical and security incidents and the time employees spent fixing them. There are also two important distinctions: The 60% figure was a restructuring scenario, not an actual 60% workforce reduction. The 32% figure refers to organizations restoring positions previously eliminated through AI or automation, not necessarily every company rehiring the exact same employees. 1. Number-by-number fact check Claim Verdict What the sources say Teams faced cuts of up to 60% Reuters confirmed this was considered in internal scenario planning. Zuckerberg canceled the next layoffs He canceled the planned November restructuring wave on May 19, hours before the May 20 cuts. Engineers produced 220% more code Code changes to internal software platforms and infrastructure rose 220% year over year. Major technical/security incidents rose 40% Reported in Meta internal posts reviewed by Reuters. Firefighting time rose 70% Employee time spent resolving those problems increased by as much as 70%. 32% of hiring managers reported jobs restored Robert Half survey of 2,000 US hiring managers; 32% of those who eliminated positions later reinstated the same or similar roles. Sources: Reuters investigation, August 26, 2026; Robert Half 2026 Midyear Labor Market Outlook. Reuters +2 2. Two additional numbers you should include These make the Meta story even more interesting. 220% versus 36%: Internal code changes increased 220%, but changes resulting in new or upgraded features reaching Meta users increased only 36%. Employee morale: Meta’s internal employee sentiment score dropped from 74% favorable to 55% favorable. That’s a much better illustration of the productivity problem than code output alone. Reuters 3. Corrected LinkedIn post The AI layoff boom may be running into a rather expensive problem: the humans still have to clean up after the AI. I just read the Reuters investigation into Meta’s AI workforce transformation. The numbers are remarkable. Meta explored cutting some teams by as much as 60%, replacing portions of their work with AI agents. The restructuring was planned in two waves. Then Zuckerberg pulled the plug on the second wave on May 19, just hours before the first layoffs began. Here’s what Reuters found in Meta’s internal data: Code changes to internal software platforms and infrastructure increased 220% year over year. But new or upgraded features reaching Meta users increased just 36%. Major technical and security incidents increased 40%. Employee time spent firefighting those problems increased by as much as 70%. Employee sentiment fell from 74% favorable to 55%. Read those numbers again. AI was generating substantially more code. But the increase in actual product improvements was much smaller, while technical problems and cleanup work increased. Then there’s another data point. A Robert Half survey of 2,000 US hiring managers found that 32% of those who eliminated positions because of AI or automation later reinstated the same or similar roles. Among employers reporting these reversals: 40% cited institutional knowledge AI couldn’t replace. 38% needed more human oversight and quality control. 35% said productivity improvements fell short of expectations. We’re seeing a pattern. Companies are measuring how much work AI can generate, but not always measuring the cost of checking, correcting and managing that work. Reducing headcount is easy to put in a spreadsheet. Replacing experience, judgment and institutional knowledge is considerably harder. Maybe the AI revolution isn’t eliminating as many jobs as executives hoped. Maybe some companies are simply discovering that productivity isn’t measured by how much work a machine produces. It’s measured by how much useful work actually gets done. Sources: Re

VMC: Aggregate Is Scarce Where It Matters The world isn’t running out of rocks. It’s running out of permitted quarries in the right places. Crushed stone, sand and gravel are geologically abundant. But having high-quality reserves near growing cities and construction markets is a different story. Aggregate is heavy and relatively inexpensive per ton. Transportation costs matter. Approximately 80% of Vulcan Materials’ shipments move directly by truck, giving nearby quarries a significant cost advantage. And building a competing quarry isn’t easy. You need the right geology, land, zoning, environmental permits and community approval. Vulcan’s competitive position: 16.6 billion tons of proven and probable aggregate reserves. 425 active aggregate facilities. Exposure to some of America’s fastest-growing metropolitan markets. The pricing power is showing up in the numbers. Q3 2026: Aggregate shipments declined 3%. Freight-adjusted selling prices increased 3.8%. Volumes fell. Prices still increased. The investment thesis: Vulcan’s moat isn’t that we’re running out of rocks. It’s that you cannot easily build another permitted quarry near Atlanta, Dallas, Nashville or the next major growth market. In this business, location, permits and transportation economics matter more than how much stone exists underground. That’s a real competitive advantage. Sources: Vulcan Materials 2025 Form 10-K and Q3 2026 Form 10-Q.

Reed Hastings thinks AI will teach every child in the next 2–4 years. Are we ready to let our kids be the experiment? The Netflix co-founder, former math teacher and Anthropic board member has a big vision for education. His argument is simple. He started as a math tutor and eventually became a teacher. One-on-one tutoring worked. But put 30 or 40 children in a classroom and individual attention disappears. A private tutor costs roughly $100 an hour. AI could theoretically give every child that same personalized attention. Sounds great. Then Kara Swisher challenged him on the evidence. Her point: research suggests AI in classrooms can actually make learning worse. Hastings didn’t disagree. He admitted AI tutors aren’t very good right now. But he believes they’ll improve dramatically over the next few years. Here’s what the research actually shows: A 2025 PNAS study found students using unrestricted ChatGPT performed worse when they had to solve problems without AI. A Harvard randomized trial found students using a carefully designed AI tutor learned more physics in less time. A two-year Tennessee study of Khan Academy’s Khanmigo found modest improvements in math achievement, but gains were similar to traditional Khan Academy practice. Other research found that simply giving elementary students access to AI doesn’t guarantee they’ll use it or learn more. There is a pattern here. AI can help students get answers. Carefully designed AI can even improve learning. But getting an answer and understanding why it’s correct are two very different things. And a teacher does considerably more than deliver information. Teachers develop critical thinking, recognize confusion, build social skills and understand when a child needs help. Hastings may be right about the technology. But we’re still trying to establish whether it produces better long-term educational outcomes. There’s also an uncomfortable question. Hastings sits on the board of Anthropic, the company behind Claude. That doesn’t invalidate his argument, but it makes independent evidence even more important. We keep hearing that AI will transform everything. Healthcare. Finance. Education. But technology isn’t the same as an outcome. If an AI tutor is doing all the thinking, what exactly is the child learning? And perhaps the most important question for parents: Are we making education fairer, or are we turning an entire generation of children into the test group? #AI #Education #ArtificialIntelligence #FutureOfLearning

You’re referring to a significant reality check for the pharmaceutical industry, and there’s a bigger AI story here. On November 7, 2026, two major heart-disease drug trials delivered disappointing results at the American Heart Association meeting. The failures are particularly interesting because they expose the gap between finding something that looks scientifically promising and actually improving patient outcomes. What failed? 1. Novartis / Ionis — Pelacarsen Designed to lower Lp(a), a genetically linked risk factor for heart disease. Lowered Lp(a) levels by approximately 78%. Tested in more than 8,300 patients. Failed to reduce heart attacks, strokes, cardiovascular deaths or emergency artery procedures. 2. Novo Nordisk — Ziltivekimab Designed to target IL-6, a protein involved in inflammation. Successfully reduced inflammation. Failed to demonstrate a reduction in major cardiovascular events. Novo subsequently discontinued two related late-stage trials. Sources: Reuters +1 Here’s where the AI hype comes in STAT published an article today by Matthew Herper titled Why the latest heart drug failures should make us all more skeptical of AI hype. His argument is worth paying attention to. The pharmaceutical industry has repeatedly embraced technologies and scientific breakthroughs that promised to transform drug discovery. But identifying a biological target doesn’t mean you’ve found a successful treatment. STAT And this is exactly the distinction I think gets lost in the AI conversation. AI can identify patterns. AI can process enormous amounts of research. AI can identify potential drug targets. AI can accelerate research and development. But none of those things proves a drug works. A successful prediction is not a successful outcome. One important qualification: these two failed drugs weren’t established as AI-developed drugs. The failures are a warning against technological hype more broadly, not proof that AI drug discovery itself has failed. The LinkedIn angle The AI hype machine just ran into an old-fashioned problem: reality. Two major heart-disease drug trials have delivered disappointing results. Novartis and Ionis developed a drug that reduced a genetically linked risk factor by 78%. More than 8,300 patients participated. The result? No meaningful reduction in heart attacks, strokes or cardiovascular deaths. Novo Nordisk also tested an anti-inflammatory drug that successfully reduced inflammation. It didn’t translate into fewer major cardiovascular events either. Think about that. The science looked promising. The biological targets were reached. The measurements moved in the right direction. But the patients didn’t get the expected benefit. This is a lesson for the entire AI industry. Finding patterns isn’t the same as finding answers. Producing a prediction isn’t the same as proving it. And generating research isn’t the same as delivering an outcome. Wall Street has the same problem. AI can produce a 50-page investment report in minutes. That doesn’t mean the investment thesis is correct. We are becoming incredibly good at producing information. We still have to prove that the information actually works. Technology can accelerate discovery. It cannot replace evidence. This is a particularly good example of the distinction you’ve been making with DoTadda: data, analysis and evidence are three different things. And ultimately, outcomes are what matter.

Yes—but the precise claim is that Singapore materially reduces Apple’s tax rate, not that Apple pays no tax there. Apple’s fiscal 2026 10-K reports: •A $5.492 billion reduction in its calculated tax provision from the “Singapore tax rate difference, including incentives.” This lowered its effective tax rate by 3.4 percentage points. Apple 2026 10-K, Note 8 •Apple’s overall effective tax rate was 17.0%, versus the 21% U.S. federal statutory rate. •Apple nevertheless paid $3.253 billion in cash income taxes to Singapore during fiscal 2026. That is tax paid—not the amount saved. Apple 2026 10-K, income taxes by jurisdiction •Ireland also reduced the tax calculation by $3.464 billion, while minimum top-up taxes and other items partly offset these benefits. Verdict A fair formulation is: Apple lowers its global tax expense partly through Singapore’s lower tax rate and tax incentives. However, the $5.492 billion combines the ordinary difference in tax rates with incentives, so Apple does not disclose how much came from negotiated incentives alone. Singapore is also only one component of Apple’s broader international tax structure. Sources: • Apple 2025 Form 10-K — Segment Information and Geographic Data • Apple 2026 Form 10-K — Uncertain Tax Positions

WSJ BREAKING: Breakfast is now just “Dinner That Starts Earlier” The Wall Street Journal has officially declared that the first meal of the day is no longer recognizable, mostly because people are out here treating a one-pound ribeye like it’s a bowl of oatmeal. Previously vegan concert pianists are now documenting their “high-fat carnivore mornings” for 647,000 followers: pound of ribeye in butter, then six eggs, then presumably a standing ovation from their own arteries. New breakfast rules according to the WSJ: •Cereal is for people who still believe in the concept of “later.” •Your coffee should now be paired with a 2019 Cabernet and a side of existential dread about what you’re eating at 7:12 a.m. •If your plate doesn’t look like it could star in a steakhouse commercial, you’re doing it wrong. •The food pyramid has been quietly replaced by a food rectangle. The rectangle is a ribeye. Somewhere, a Kellogg’s executive just stared out a window and whispered, “We never stood a chance.” Meanwhile, dinner has been demoted to “the meal you eat after you’ve already had dinner.”

AI JUST STOLE $129 MILLION. AND IT DIDN’T NEED TO HACK A COMPUTER. The Department of Justice has announced charges against three foreign nationals accused of using what prosecutors call “super intelligence” to defraud companies around the world of approximately $129 million. Read that number again. Here’s how the alleged operation worked. Criminals identified employees at targeted companies, often people working in accounting or finance. They used AI to replicate the voices and images of corporate CEOs. The fake CEO contacted an employee, claiming there was a confidential financial transaction that needed immediate attention. The employee was directed to contact attorneys at major international law firms. Those attorneys were also impersonated using AI. The fake lawyers then instructed employees to wire millions of dollars into accounts controlled by the alleged conspirators. The money was transferred through bank accounts across multiple countries and converted into cryptocurrency to conceal its origins. The alleged victims included: A Missouri hospital system: $4.8 million. A Rhode Island bank: $2.7 million. A California hospital system: $3 million. A UK financial technology company: $11.6 million. A Massachusetts mutual fund company: $4.6 million. A Florida pet food company: $5.2 million. And those are just some of the identified victims. The alleged operation ran from 2024 through February 2026 and targeted companies around the world. Three defendants have been arrested in Spain and the United Kingdom. They face federal charges including wire fraud, aggravated identity theft and money laundering. All are presumed innocent unless proven guilty. Now comes the part businesses should be paying attention to. We have spent years investing in cybersecurity, authentication systems, compliance departments and financial controls. But what happens when the voice on the phone sounds exactly like your CEO? When the attorney sounds real? When the transaction appears legitimate? When the people involved are all supposedly working on the same confidential deal? This is no longer just a cybersecurity problem. It’s an evidence and verification problem. AI can manufacture convincing information. It cannot establish that the person on the other end of the phone is who they claim to be. The next generation of corporate fraud prevention won’t just be about detecting fake images or voices. It will require independent verification, transaction controls and evidence that exists outside the AI-generated conversation. The technology is getting smarter. Are our controls keeping up? Source: U.S. Department of Justice, Southern District of New York, November 5, 2026. My take: There’s a bigger business lesson here: AI doesn’t have to break into your systems if it can convince your employees to authorize the transaction themselves. That’s a much harder problem to solve with traditional cybersecurity alone. Source: Original DOJ indictment announcement⁠. See also AP News +1.

Snowflake earnings: what matters most The key question is whether product-revenue acceleration is durable, not simply whether Snowflake beats the reported-quarter target. 1. Product revenue and next-quarter guidance — most important Snowflake guided Q3 FY2027 product revenue to $1.588–$1.593 billion, representing 37%–38% growth. FY2027 guidance was raised to $6.07 billion, or 36% growth. Q2 FY2027 earnings release Watch for: •Product revenue above the guided range. •More importantly, whether the next-quarter and full-year outlook rises. •Management commentary that consumption remained strong after quarter-end. Because Snowflake recognizes revenue as customers consume computing resources, forward consumption commentary often matters more than bookings. 2. Core consumption versus AI-driven growth Management said AI contributed roughly half of the recent acceleration, while core workloads also strengthened. Snowflake reported: •More than 9,100 Cortex Code accounts •5,800 Cortex AI/CoWork accounts •Stronger platform consumption among AI adopters—but management did not quantify the uplift. Q2 FY2027 earnings call The best signal would be Snowflake finally providing: •AI-related revenue or consumption •Consumption uplift among AI customers •Evidence that pilots are becoming production workloads Customer counts alone are less valuable if monetization remains unclear. 3. Net revenue retention Net revenue retention was 126% last quarter. Q2 FY2027 Form 10-Q—MD&A Interpretation: •Stable or higher: Existing customers are expanding consumption, supporting durable growth. •Continued decline: New customer additions and AI products may not yet be offsetting slower expansion in established workloads. 4. Margins: is AI growth economically attractive? The FY2027 non-GAAP product gross-margin target was 74.0%, while Q3 non-GAAP operating-margin guidance was 15.5%. Management said newer AI workloads currently have lower contribution margins. Q2 FY2027 earnings call Watch whether: •Product gross margin stabilizes. •AI inference and model costs continue pressuring margins. •Operating-margin guidance rises despite AI investment. •Snowflake can generate leverage without slowing product development. Rapid AI growth accompanied by falling gross margins would make the quality of the acceleration less compelling. 5. RPO and large customers Last quarter: •Remaining performance obligations, or RPO: $9.00 billion, up 30% •Approximately 54% expected to convert within 12 months •Customers spending over $1 million annually: 828 •Customers spending over $10 million: 65 Q2 FY2027 earnings call RPO growth should ideally remain near product-revenue growth. However, RPO is not a perfect leading indicator because Snowflake’s customers can consume faster or slower than their contract schedules. 6. Cash flow, stock compensation and dilution Q2 stock-based compensation was $423.6 million, approximately 27% of revenue. Unrecognized compensation expense was about $3.8 billion. Q2 FY2027 Form 10-Q—financial statements Check: •Stock-based compensation as a percentage of revenue. •Adjusted free-cash-flow margin versus the 23% FY2027 target. •Diluted share-count guidance. •Whether share repurchases meaningfully offset employee-equity issuance. Simple scorecard Strong report: Product-revenue beat, forward guidance raised, NRR stabilizes, AI monetization becomes more measurable, and margins hold. Mixed report: Good reported revenue but cautious guidance, with growth driven primarily by unquantified AI adoption. Weak report: Slower consumption, declining NRR, limited AI monetization evidence, and simultaneous gross-margin pressure. Bottom line: Focus first on the forward product-revenue guide, second on core versus AI consumption, and third on gross-margin durability. An earnings beat without stronger guidance probably will not be enough to make the report genuinely exciting. Sources: • Snowflake Q2 FY2

Yes, it works, and plenty of people run it. The post oversells it a bit. What works •Claude Code can read and write plain markdown files, so a folder of linked notes suits it. •Obsidian is just a viewer for that folder, so nothing needs integrating. •Having Claude read each source once and file the key points into wiki pages makes later questions faster than re-reading everything. Where it falls short •"Smarter" is overstated. It doesn't learn. It keeps better notes, and quality depends on how good your CLAUDE.md rules are. •Drift and errors: Claude can summarize wrongly or merge pages badly, and the mistakes then sit in your wiki looking authoritative. Keep raw/ untouched so you can check. •Scale: at hundreds of pages, Claude needs an index page or good search to find the right ones. Without that, answers get spottier. •Cost: every ingest uses Claude Code usage, so big PDFs add up. •Setup: it won't work well from "paste the idea and build it." You'll tweak the rules a few times. Quick test: make a vault, drop in 3 articles you know well, have Claude ingest them, then ask questions you know the answers to. If it's accurate, keep feeding it. Want me to write that starter CLAUDE.md?

Kind of, but it's less about ABC having the Finals and more that NBC owns what happens next. Brunson is the reigning Finals MVP of a Knicks team that won its first championship in 53 years, and NBC is in its first season with NBA rights. They've leaned into that: •SNL: NBC's own show had him host its season premiere, and NBC Sports' release says he became only the fifth NBA player to host. •Opening night: NBC gave the Knicks' banner-raising and the 76ers game a prime slot in its tip-off tripleheader, with promos built around the Knicks. •Regular schedule: NBC had already put Brunson and the Knicks in marquee slots last season, including Sunday Night Basketball and the MLK Day quadruple header. The Finals being on ABC doesn't give NBC a reason to hype him, and it may even work the other way, since the network that didn't air the title run wants to claim the champs going forward. Mostly it's good business. New York is the biggest market, Brunson is the face of the champs, and NBC gets cross-promotion between SNL and the NBA. Sources: •Bleacher Report: Brunson's SNL promo •NBC Sports: tip-off tripleheader •NBCUniversal: Knicks-76ers promo

If you mean the November 6 news, your interpretation is directionally right but too broad. Microsoft reportedly wound down an Oracle cloud-capacity arrangement because the servers used older-generation Nvidia GPUs that no longer met Microsoft’s requirements. That suggests part of Oracle’s rented AI infrastructure fleet was technologically behind Microsoft’s needs—not that Oracle’s entire database and software stack is obsolete. Microsoft itself reportedly declined to comment. November 6 report Investment read-through for ORCL •Negative: Raises utilization and residual-value risk for older GPU capacity, especially if hyperscale customers increasingly demand the newest chips. •More important risk: Oracle is making enormous infrastructure commitments. Rapid GPU obsolescence could shorten asset lives and weaken returns on that capex. •But not proof of broad technological inferiority: Microsoft still supports Oracle AI Database@Azure and calls Oracle a critical system of record for enterprise data. Microsoft Oracle-on-Azure announcement •Separate structural issue: Microsoft has explicitly called on-premises Oracle databases expensive, hard to scale and increasingly inconsistent with modern enterprise agility—while promoting migration to PostgreSQL on Azure. Microsoft Azure modernization commentary Bottom line: The strongest bearish conclusion is not “Oracle’s whole stack is old.” It is: Oracle may be taking substantial balance-sheet risk on AI hardware that depreciates technologically faster than expected, while its legacy database installed base remains vulnerable to cloud-native alternatives. Sources: • From legacy to leadership: How PostgreSQL on Azure powers enterprise agility and innovation • Zero downtime migration from Oracle to Azure SQL • New Oracle Autonomous AI Database – Dedicated option on Oracle AI Database@Azure • Microsoft is winding down its deal to rent AI computing from Oracle

This is actually more interesting than a fund blowing up. Katamaran Capital says it is closing despite being profitable since inception, returning all investor capital, and having its short position in Life360 work exceptionally well. So why shut down? Because being a good investor and running a successful hedge fund are two different businesses. The statement tells us three things: They say the strategy was profitable. They say the closure was a business decision, not a performance decision. They returned capital in an orderly fashion rather than liquidating under pressure. But it leaves out the most important information: why the business was no longer worth operating. There is some useful background. Katamaran was founded by Priya Kodeeswaran, a former Millennium and Man Group portfolio manager. The fund was backed by Brummer & Partners. Its reported assets were approximately: $320 million in March 2025. FN London More than $475 million in a January 2026 interview. Hedge Fund Alpha That makes this particularly interesting. This wasn’t obviously a tiny fund that never attracted capital. There is also an important distinction: its publicly reported SEC 13F holdings are not its total AUM. They only cover reportable U.S. securities. Holdings Channel My interpretation: this is probably an economics and strategic decision, but we don’t have enough evidence to establish the actual reason. A few possibilities: Standalone economics: Running a hedge fund is expensive. Staff, technology, compliance, prime brokerage, data and administration eat into management-company profitability. Capital economics: $400–475 million can produce a respectable business, but it may not produce the economics the founder wants. Opportunity cost: A successful portfolio manager can potentially earn more operating inside a large platform than carrying the costs and responsibilities of an independent firm. Market structure: A low-net long/short strategy needs differentiated opportunities. Crowding, volatility and financing can make that increasingly difficult. I would not assume this was caused by the November selloff. The fund was already wound down in October. Here’s the bigger issue I see. We’re hearing about hedge funds cutting positions, reducing teams, returning capital and now an apparently profitable fund closing. Meanwhile, the industry keeps selling the idea that more data, more AI, more analysts and more sophisticated systems create an investment advantage. But there is a distinction between generating investment returns and building a sustainable investment business. And that distinction deserves considerably more attention. One question I would want answered: If a profitable fund with hundreds of millions in assets decides the business isn’t worth continuing, what does that say about the economics of running an independent hedge fund today? I wouldn’t call Katamaran a failure. But I would call it a warning worth investigating.

Go for it

Ai should work like the guy at the hotel who answered my question right. It’s my job When Maybe