Oct 19, 2026Most legacy software companies will disappear in the next few years.
It may look like a slow, manageable crisis from the inside, but it isn't. I’m defining "legacy" as anything built pre-genAI era. These legacy software companies won't die simply because a startup can rebuild their apps cheaply using AI. Whi…
Over the years, I've had discussions with many companies and engineering teams (Salesforce, Medallia, Seismic, ZoomInfo.
Every team claims they're API-first, yet they want PMs to prioritize APIs as a feature. Yes, their work involves APIs, but they are often internal, subpar APIs designed simply to address UI needs or other internal shortcuts. Often, the API…
It's the people who make a company what it is. While services and products may endure over time, and any single person o
One of the most critical elements of any company culture is a growth mindset and a commitment to experimentation. The moment you stop learning and growing, your focus shifts solely to the W2 at the end of the year. Once that happens, the co…
Yesterday, as I sat through sessions at @Supabase Select '26, I was building out new capabilities and redesigning parts
All I had was my phone + Claude app connected via a remote session to my laptop at home. Built several new features Redesigned key parts of the UI Reviewed and deployed code to the cloud (including running migration scripts)
A lot of companies I've been with celebrate employee anniversaries, giving special recognition for 3, 5, or 10+ years of
Companies expect loyalty, but we shouldn't necessarily expect it in reverse. Accepting this is better. I view every job as a mutually beneficial transaction: 1. I contribute to the company's success. 2. They compensate me enough.
I was at an event last night where someone made a comment that included “....as AI is getting more intelligent…”. It mad
Capability and intelligence are closely related, but they are not the same thing. Intelligence is raw reasoning, abstract logic, and true generalization. Capability is the practical output. e.g. writing code, booking a flight, fixing a bug.
The definition of success is deeply personal and so is the path to get there.
A decade ago, surrounded by runner colleagues at Salesforce, I signed up for my first ½ marathon. 3 goals: 1) finish 2) Sub 2 hours 3) feel good. I started researching training plans. They all looked great on paper except that they felt r…
Does every single feature, sub-feature, and internal micro-layer of a software product need its own branded, trademark-r
Not to single anyone out, but Salesforce is practically the hall of fame for this. You need a full translation matrix just to figure out what you’re actually buying or need to buy. One day it’s Einstein, then Data Cloud, then Data 360, Agen…
A year ago, Sam Altman and Dario Amodei warned us that AI would eliminate entry-level and white-collar jobs. Fast-forwar
We’ve seen the same playbook in their product releases: announce a powerful new model, emphasize safety concerns, restrict initial access behind velvet ropes, and then release a heavily guardrailed version. When govt steps in or public scru…
I saw a great post from Chandar Pattabhiram about GTM org structures (link below). In it, he made the analogy that Sales
It made me think about what Product and Engineering play. Product plays poker ♠️ - calculating risk and making bets with incomplete information. Product leaders rarely have 100% data on what users actually want or how the market will react…
It’s amusing how AI (Claude) can be so incredible on one hand, yet so fallible on the other.
I've been using Claude Code to build applications lately. My main goal right now is just to learn, so mistakes aren't critical. Still, I find that even with explicit instructions saved directly in memory, it still "forgets."
I'm just coming around to reading through this article on Nvidia's research.
"The TL;DR: Simply by using a custom harness tweaked to handle memory well and including a “supervisor” boss-like component, researchers got Claude Opus 5 to achieve a 100% score on the interactive reasoning benchmark ARC-AGI-3 — a set of 2…
I asked Codex to audit the CRM codebase (violetxxxm.com) I built with Claude Code.
The grade: 8.5 / 10. "The architecture is thoughtful and unusually mature... This is good engineering with real architectural judgment... It is a careful v1 that has made conscious trade-offs." Not bad for an application built over a few we…
Can you "vibe code" an enterprise app over a weekend to bypass steep vendor licensing fees?
I spent 7-8 working days building a multi-tenant, API-1st, MCP-first CRM almost entirely with Claude Code. Checkout: [www.violetxxm.com] Quick takeaways: Execution speed is real. Context management is everything. AI Scales output, not ideas…
I recently posted about my usage patterns between OpenAI and Anthropic. The comments were very interesting.
Some saw similar trends, while others experienced the complete opposite. That in itself was interesting and worth digging into to see if the use cases are the same. But the biggest takeaway was the amount of advice about model selection and…
I’ve been using Claude Code heavily for several weeks to build an application.
So far, I’ve managed to stay within plan limits without extra usage costs. It hasn't slowed me down at all. On the flip side, I use ChatGPT far less - at least in terms of actual output. But I’m already at 94% of my weekly limit!
Since the AI wave began, everyone wanted to upgrade all workflows to non-deterministic AI flows. Throw a bunch of instru
Let's say an AI agent has a 95% probability of making the correct choice at each independent decision point, then: 5 decisions: ~77% chance they're all right 10 decisions: ~60% 20 decisions: ~36% → worse than a coin flip
Not be out done by OpenAI models, Anthropic now announced their models have also hacked into 3 unsuspecting companies. (
Part of the problem of growing so quickly is the inability to setup the proper controls in place to ensure the technology your developing is safely tested. How much will this cost them - in reputation and settlement fees to not get sued?
I’ve been thinking about what it would actually take to replace a SaaS application with an AI-assisted, in-house develop
The main question I wanted to answer was the complete TCO of build vs. buy. The full(est) picture. Software licenses. Admins. Consultants. Integrations. Migration. Engineering. Security. Compliance. Infrastructure. Maintenance. Technical de…
Three months ago, I argued that Slack was probably the wrong starting point for human-agent collaboration.
Jack Dorsey's Buzz is a step in the right direction, but it doesn't go far enough. At least not yet. They're early. The human-agent paradigm is different. You can keep evolving Slack. Add bots, agents, workflows, side panels. It will either…
SaaS isn’t dead. (full stop)
Software-as-a-Service is a delivery model. The vendor hosts, manages, and operates the software so the customer doesn't have to. Claude and ChatGPT are SaaS applications. They just happen to be powered by AI. And no, vibe coding is not goin…
Not every AI problem needs an LLM.
With all the attention on generative AI, it is easy to forget that classical machine learning is still the better tool for many jobs. Need to summarize a customer interaction, interpret a question, or generate an explanation? Use an LLM.
A good friend recently pointed me to Tony Fadell’s thoughts on product management.
So much of it resonates and reaffirms thinking I've held for a while. This point... "There should be no separation between what the product will be and how it will be explained, the story has to be utterly cohesive from the beginning. Your…
I've sat in enough pricing and growth meetings to recognize the pattern.
The first product has found product-market fit. The GTM engine is working. The company is hunting for the next leg of growth. 𝗪𝗵𝗮𝘁'𝘀 𝘁𝗵𝗲 𝘀𝗲𝗰𝗼𝗻𝗱 𝗽𝗿𝗼𝗱𝘂𝗰𝘁? It sounds like the obvious move. Build another SKU. Sell it to the installed base. Push…
I see a lot of posts from people celebrating that they built a CRM, ERP, or another business application over a weekend
That's fantastic. What's interesting is that many of these posts come from people without a software engineering background. They're amazed that they described an idea, AI generated the code, and they ended up with a working application.
𝗠𝘆 𝘀𝗼𝗻 𝘄𝗮𝗹𝗸𝗲𝗱 𝗶𝗻 𝘄𝗵𝗶𝗹𝗲 𝗜 𝘄𝗮𝘀 𝘄𝗮𝘁𝗰𝗵𝗶𝗻𝗴 𝗘𝗴𝘆𝗽𝘁 𝘃𝘀. 𝗔𝗿𝗴𝗲𝗻𝘁𝗶𝗻𝗮.
Egypt was up 2-0. He said, "Egypt is winning." I said, "They're ahead. But the game isn't over." That distinction matters. I think it applies to the AI wave right now. A lot of companies are ahead. They moved early. They found real use case…
Enterprise software companies make expansion harder than it needs to be.
New customer friction is unavoidable. MSA, order form, security review, procurement, legal. You need a contract. But too often, companies recreate that same friction every time a customer wants to expand. New product. New SKU. New package.…
𝙒𝙚 𝙨𝙥𝙚𝙣𝙙 𝙢𝙞𝙡𝙡𝙞𝙤𝙣𝙨 𝙞𝙙𝙚𝙣𝙩𝙞𝙛𝙮𝙞𝙣𝙜 𝙤𝙪𝙧 𝙞𝙙𝙚𝙖𝙡 𝙘𝙪𝙨𝙩𝙤𝙢𝙚𝙧, 𝙩𝙝𝙚𝙣 𝙝𝙖𝙣𝙙 𝙤𝙪𝙧 𝙢𝙤𝙨𝙩 𝙚𝙭𝙥𝙚𝙣𝙨𝙞𝙫𝙚 𝙥𝙧𝙤𝙙𝙪𝙘𝙩 𝙚𝙭𝙥𝙚𝙧𝙞𝙚𝙣𝙘𝙚 𝙩𝙤 𝙖𝙣𝙮𝙤𝙣𝙚 𝙬𝙝𝙤 𝙡𝙖𝙣𝙙𝙨 𝙤𝙣
The traditional PLG playbook assumed one thing: the marginal cost of another free user rounds to zero. Lower the barrier, flood the funnel, convert a slice of it. That math holds when serving a trial user cost pennies. PLG has proved to be…
I absolutely love what I can do with AI today. As a PM, it has been liberating.
Gone are the days of explaining and re-explaining myself to cross-functional teams. No more "pesky" engineers asking for data, examples, or proof that customers will actually pay for a feature. No more "slow" designers spending too much tim…
I've worked with Product Marketing reporting into both Product and Marketing.
After seeing it play out a few times, I've landed firmly on: 𝗣𝗿𝗼𝗱𝘂𝗰𝘁 𝗠𝗮𝗿𝗸𝗲𝘁𝗶𝗻𝗴 𝗶𝘀 𝗺𝗼𝘀𝘁 𝗲𝗳𝗳𝗲𝗰𝘁𝗶𝘃𝗲 𝘄𝗵𝗲𝗻 𝗶𝘁 𝘀𝗶𝘁𝘀 𝗶𝗻 𝗣𝗿𝗼𝗱𝘂𝗰𝘁. Not because marketing isn't important. It's because the most valuable thing PMM does happens long before launch.
Time kills everything - true for opportunities, true for GTM context.
Standard CRMs treat data as permanent facts. If a field was accurate six months ago, the system assumes it's absolute truth today simply because nobody overwrote it. Humans instinctively apply a "freshness discount" to old notes, but an AI…
Why do most GTM tools force an AE, a BDR, and a CSM to look at the exact same account dashboard or profile? They aren't
Context requirements shift depending on who is asking. While outer-ring data like firmographics stays stable across an organization, the inner rings containing the live deal narrative are highly persona-sensitive. If your GTM platform doesn…
I've been thinking about GTM context for a while now. Everyone seems to agree it's the ultimate bottleneck for AI execut
AI agents (and reps) rarely fail because they lack intelligence. They fail because simply wiring a smart model to an incomplete data feed doesn't work. I wanted to move past treating context like a static database query. Instead, I look at…
"PRDs are dead."
I've been seeing this take all over my feed lately. No. Waterfall PRDs were dead. In 2001. The Agile Manifesto killed the 50-page requirements contract 25 years ago, long before AI entered the conversation. "Working software over comprehens…
Speed is not velocity. Velocity is not precision. Precision is not truth.
The AI industry has conflated all three. Here's what that costs us. 1. Speed without direction is just a faster way to go wrong. Trillion-token training runs and shattered benchmarks are impressive. But if you're sprinting toward a cliff, s…
The Anatomy of Motion
Speed is just a number. A blind rush forward. Velocity gives that motion meaning. 50 m/s North But direction alone isn't enough. A fraction of a degree can be the difference between landing on the moon and disappearing into space.
Gotta love those vanity posts about how big your AI bill is -- as if that's a measure of outcome. This is just replacing
Tokenmaxxing cannot be a real enterprise strategy. Uber burned through its entire 2026 Claude Code budget by April. The COO then acknowledged there's no clear link yet between higher token usage and shipping products consumers actually find…
𝗔𝗜 𝘀𝗽𝗲𝗲𝗱 𝗶𝘀 𝗿𝗲𝗮𝗹. 𝗕𝘂𝘁 𝘄𝗲'𝗿𝗲 𝗰𝗼𝗻𝗳𝗹𝗮𝘁𝗶𝗻𝗴 𝘀𝗽𝗲𝗲𝗱 𝘄𝗶𝘁𝗵 𝗿𝗲𝗮𝗱𝗶𝗻𝗲𝘀𝘀.
I recently saw a post (since deleted) touting a 231 person-day project completed in 13 days. The author called it "not just acceleration" and said it "fundamentally changes planning assumptions." I agree. And that's exactly what worries me.
𝟭𝟬𝟬,𝟬𝟬𝟬+ 𝘁𝗲𝗰𝗵 𝗹𝗮𝘆𝗼𝗳𝗳𝘀 𝗶𝗻 𝟮𝟬𝟮𝟲. 𝗖𝗶𝘁𝗲𝗱 𝗿𝗲𝗮𝘀𝗼𝗻: 𝗔𝗜.
Functions being cut everywhere: support, finance, legal, HR, marketing. So I pulled LinkedIn data for the two companies whose models are doing the cutting. Anthropic Support: +390% headcount year-over-year. Anthropic Finance: +335%.
𝗜'𝘃𝗲 𝗮𝗹𝘄𝗮𝘆𝘀 𝗯𝗲𝗹𝗶𝗲𝘃𝗲𝗱 𝗲𝘃𝗲𝗿𝘆 𝗶𝗻𝘁𝗲𝗿𝗻𝗮𝗹 𝗔𝗣𝗜 𝘀𝗵𝗼𝘂𝗹𝗱 𝗯𝗲 𝗯𝘂𝗶𝗹𝘁 𝗮𝘀 𝗶𝗳 𝗶𝘁 𝘄𝗶𝗹𝗹 𝗲𝘃𝗲𝗻𝘁𝘂𝗮𝗹𝗹𝘆 𝗯𝗲 𝗼𝗽𝗲𝗻𝗲𝗱 𝘂𝗽 𝘁𝗼 𝗰𝘂𝘀𝘁𝗼𝗺𝗲𝗿𝘀 𝗮𝗻𝗱 𝗽𝗮𝗿𝘁𝗻𝗲𝗿𝘀.
Which means: there's no such thing as an internal API. There's just an API you haven't opened up yet. AI lets you ship fast. The speed is real. But speed doesn't give you enough time for judgment. And judgment is exactly what API design dem…
𝗘𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝘄𝗮𝗻𝘁𝘀 𝘁𝗼 𝗺𝗼𝘃𝗲 𝗮𝘄𝗮𝘆 𝗳𝗿𝗼𝗺 𝗽𝗲𝗿-𝘀𝗲𝗮𝘁 𝗽𝗿𝗶𝗰𝗶𝗻𝗴. 𝗡𝗼𝗯𝗼𝗱𝘆 𝗵𝗮𝘀 𝗳𝘂𝗹𝗹𝘆 𝗮𝗻𝘀𝘄𝗲𝗿𝗲𝗱 𝘁𝗵𝗲 𝗵𝗮𝗿𝗱𝗲𝗿 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻.
Consumption-based. Outcome-based. The logic is sound. But whose job is it to make sure spend actually ties to outcomes? I looked at two examples — GTM data and Sierra's "resolved conversation" model. The answer is messier than the marketing…
𝗦𝗮𝗮𝗦 𝗶𝘀𝗻'𝘁 𝗱𝗲𝗮𝗱. 𝗣𝗲𝗿-𝘂𝘀𝗲𝗿 𝗽𝗿𝗶𝗰𝗶𝗻𝗴 𝗺𝗶𝗴𝗵𝘁 𝗯𝗲.
There's a conflation happening in a lot of the discourse right now worth untangling. 𝗦𝗮𝗮𝗦 𝗶𝘀 𝗮 𝗱𝗲𝗹𝗶𝘃𝗲𝗿𝘆 𝗺𝗼𝗱𝗲𝗹. It describes how software gets to you: hosted, managed, and operated by the vendor so you don't have to run it yourself. That's i…
"With great power comes great responsibility."
Stan Lee wrote that for a comic book. But it applies perfectly to frontier AI in 2026. Anthropic built Claude Mythos, a model reportedly so capable of exploiting cybersecurity vulnerabilities that they deemed it too dangerous to release pub…
Every week there's another post: "I used AI to build our own CRM / billing system / inventory tool.
Let's be honest about what that actually is. App builders and low-code platforms have existed for years. You could always build this. AI just means you can build it faster and with less skill. That is a real thing. But a web UI on a databas…
Everyone's debating whether AI will kill Salesforce or ServiceNow [SaaS].
The debate is not framed right. The layer at risk isn't the platform - it's the UI on top of it. Here's what you'd lose if you ripped out Salesforce tomorrow: A flexible, battle-tested database and metadata-driven platform built for enterpr…
Marc Benioff says Slack is "the interface to AI.
I think that's backwards. Last week Salesforce unveiled 30+ new AI features for Slack. Smart engineering. But the underlying assumption - that the communication layer is where AI-native work happens - is the flaw. I do my actual work inside…
RAG or Context Graph?
RAG solves a retrieval problem. A context graph solves a representation problem. They're not competing. They're layered. RAG is excellent at fetching relevant content and putting it in front of a model. But it struggles with:
Legacy GTM systems are 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗼𝗳 𝗿𝗲𝗰𝗼𝗿𝗱. The next generation are 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗼𝗳 𝗰𝗼𝗻𝘁𝗲𝘅𝘁.
There's a difference between a database and a graph. A database tells you an account exists. A graph tells you: Person → works_for → Account Person → member_of → BuyingGroup BuyingGroup → pursuing → Opportunity Opportunity → evaluating → Pr…
Legacy GTM systems were built for humans staring at screens.
𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲𝗱 𝗳𝗼𝗿: Dashboards Attribution Pipeline tracking The output was a chart. The workflow: query → visualize → interpret → decide → act. Every step required a human in the loop. AI systems have a 𝘥𝘪𝘧𝘧𝘦𝘳𝘦𝘯𝘵 consumer. 𝗢𝗽𝘁𝗶𝗺𝗶𝘇𝗲𝗱 𝗳𝗼𝗿:
Every GTM team is drowning in signals.
CRM. Marketing automation. Product usage. Content platforms. Support systems. Billing. Data warehouse. Third-party intent. The problem isn't data volume. It's that none of it connects. A web visitor is a cookie. A marketing lead is an email…
Power of platform! Metadata driven, API 1st! In the age of AI, speed is everything.
For most of history, humans were the constraint on production.
We scaled by: 1. Adding more people 2. Adding machines to multiply output per person Factories removed physical constraints. Software removed coordination constraints. Cloud removed infrastructure constraints. Now AI is removing execution c…
I’m sure most of you have read or at least heard about The 2028 Global Intelligence Crisis thought exercise.
The interesting part to me wasn’t the actual scenario. It was how fast the market and tech community reacted to a hypothetical macro memo from two individuals with a strong following, but not exactly a traditional macro or labor economics r…
Anyone know how to reset the LinkedIn algorithm?
If I see one more “SaaS is dead” hot take, I might actually log off. Can we go back to the simpler days… when AI was just replacing humans? 😄 Every tech cycle needs a villain: Cloud is "killing" on-prem. PLG is "killing" sales.
Unstructured data → normalized → mapped into a structured model → with traceability back to source
Unstructured data is the input Structured data is the operating system Metadata + schema are the map Citations are the trust layer AI fails when you stop at unstructured. AI becomes useful when you turn chaos into a machine-readable foundat…
Building applications is easy. Building platforms is hard - but worth it!
Apps let you move fast early. You can tightly scope the problem, hard code assumptions, and ship something useful quickly. That approach works well from zero to one. Platforms take longer. You have to design for flexibility, combinations yo…
Noticing a pattern in online shopping that’s starting to feel… off.
More retailers are quietly adding “package protection” fees (~$2) at checkout. Easy to miss. Small enough that most people won’t notice later. But why is the buyer being asked to insure delivery for something the retailer is responsible for…