Sandeep Singh Saini

Anthropic in the last few weeks: Claude for Legal. An assistant app for small business owners. Bought a one-person biotech startup to run its own drug programs. A chip team designing custom silicon. All of it right before an IPO at up to $2 trillion. I don't think that's random. Selling model access is turning into a race to the bottom. Every few months someone ships a model that's close enough and cheaper, and your customers can switch providers by changing one line of config. Tokens don't justify $2T. So you stop being the engine and start building the car. Legal, small business, drug research. Go where the money actually is. The awkward part: a lot of the companies paying Anthropic are startups building those exact products on Claude. Your supplier just became your competitor. If you're building on an LLM API right now, how are you thinking about that?

Open source doesn't mean “pretend you wrote it.” Use the code. Build a business on it if the license allows. That's part of why people choose permissive licenses. But running someone else's work through an AI doesn't make you the person who figured it out. Someone spent evenings chasing that bug, reading undocumented hardware behaviour, and finding out which obvious solution doesn't work. You got to skip that part because they shared it. Whatever the license requires, why is giving them credit such a difficult thing to do?

Some AI coding demos look like they hired an entire department to change one button. One agent makes a plan. Another reviews the plan. A third tells you the first two are making progress. Cool. Where's the diff? I'm all for tools that save developers time. But if I have to supervise five agents, repeat the requirements, and check whether anyone actually ran the tests, that's another job. Show me what changed, what you tested, and what still doesn't work. I can do something with that. When you're choosing a coding tool, how much does speed matter if checking its work takes longer than doing it yourself?

Google is now offering voluntary exit packages to employees in Canada, France, Germany, Japan and Switzerland. Read the exclusions though. Google Cloud and DeepMind aren't eligible. That tells you exactly how big tech sees engineers right now. If you're on AI or the infra that runs it, you're the strategy. Everyone else is a line item they'd like to shrink politely, with a severance cheque so it doesn't look like layoffs. Meanwhile Alphabet is planning around $200B in capex this year. Experienced developers who are good at their jobs are being nudged out of companies that are swimming in money. If you're one of them, are you going after an AI role, or leaving big tech altogether?

Watching tech CEOs the morning after the US midterms is better than any keynote. Musk spent about $445M and is now joking about "data centers in space." Calacanis moved to Texas to get away from the left, and Texas just went blue. And on election night Sam Altman suddenly posted that people should stay in control of the future and that this should constrain AI progress. Same people who spent two years saying regulation kills innovation. Now Congress wants AI CEOs to testify under oath, and the tone is already changing. I'm not picking a side here, I'm in Ottawa watching with popcorn. But it says a lot that the "principles" seem to move with whoever holds the subpoena power. Over the next few months, which CEO do you think changes their tune the fastest?

Apple shipped a watchOS beta yesterday that listens to the conversations around you and turns them into a daily summary. It's opt-in. It's on-device. Summaries delete after 7 days unless you save them. Fine, I'll believe all of that. What I keep thinking about is the other person. I can opt in for my wrist. The guy across the table at lunch didn't opt in to anything. He just got turned into bullet points with a title. As a developer I get why this is fun to build. As a person I'm not sure I want to be the one wearing it in a meeting. Would you turn it on, or would you ask someone to take theirs off?

V2 coming soon

Jensen was right we dont need to learn anything in the future, AI cant teach us everything 🤡

Big Tech has signed roughly $3.3 trillion in future AI obligations. Alphabet alone is at $1.06 trillion, up about 700% in a year, and its free cash flow just went negative. Those leases don't care if the demand shows up. If it does, compute gets cheap. If it doesn't, someone pays for empty racks, and I'm pretty sure it's us through higher API and cloud prices. Which comes first: cheaper AI or a bigger cloud bill?

Alphabet now has $1.06 trillion in off-balance-sheet commitments for AI data centers and equipment. That's up about 700% in one year. Add Meta, Microsoft, Amazon and Oracle and Big Tech has signed up for roughly $3.3 trillion in future AI obligations, mostly long leases and hardware deals. Alphabet's free cash flow just went negative, and the stock fell 12% on it. As a developer, here's the part I keep thinking about. Those leases don't care whether the demand shows up. If it does, compute gets cheap and we all win. If it doesn't, someone pays for empty racks, and it usually lands on customers as higher prices for the stuff we already depend on. Which happens first: AI prices drop because of all this capacity, or cloud bills go up to cover it?

A guy in Kitchener drove a forklift for 20 years, started vibe coding in April, and the news is calling him a tech founder. Good for him for trying something new. But can we be honest about what this story is selling? He told CTV he doesn't really know anything about coding. He's paying for AI plans and trusting whatever comes out. That's not a knock on him. That's how most people are building right now. The AI writes it, it runs, ship it. Nobody's asking who fixes it when it breaks. Who checks the security. Who understands why it works at all. Meanwhile developers with years of real experience are sending hundreds of applications and getting ghosted. Juniors can't get a first job. And the headline everyone shares is "anyone can be a tech CEO now." Running code and understanding code are not the same thing. We're about to find out how big that gap is. Developers, is this a door opening for more people, or the reason your job got harder to find?

Addy Osmani's piece on "cognitive surrender" landed a little too close this morning. I caught myself doing it last week. An agent handed me a 300 line diff for what I asked to be a small fix. Tests passed, it ran fine, and I approved it after skimming maybe a third. Two days later I couldn't explain to a teammate why one of those functions existed. The code wasn't the problem. I was. I had stopped being the reviewer and become the person who clicks merge. So I'm trying a few rules: if I can't explain a change out loud, it doesn't ship. Big diffs get split before I read them. And once a week I write something by hand, start to finish, just to keep the muscle. Fellow devs, what's the one habit that actually keeps you sharp when the agent writes most of the code?

Funny how fast "I don't trust a woman flying my plane" disappears everywhere else. Nobody says it to the surgeon about to put them under. Or the cop who shows up when they call 911. Or the firefighter carrying someone out of a burning building, the engineer who signed off on the bridge they drive over, the soldier on the front line, the judge deciding their case, or the teacher they trust with their kids eight hours a day. Women do all of it. Same training. Same exams. Same licenses. Same standards. The doubt only shows up when something goes wrong and someone needs an easy target before the facts are in. A pilot's license doesn't come with a gender checkbox. It comes with checkrides, medicals, simulator hours and instructors who fail you if you aren't ready. If you trust women with your life everywhere else, what exactly changes at 35,000 feet?

Nine people died in the Learjet crash in Florida on Sunday, and half the replies I'm seeing are blaming the co-pilot because she was a 20-year-old woman. Here's what's actually known. She held a commercial pilot license, was a certified flight instructor, and had around 1,300 hours of flying. She had no prior incidents on her FAA record. The NTSB says the crew reported clear ice and moderate turbulence at 35,000 feet, then the jet went into an overspeed condition and dropped fast. Investigators say there's evidence the plane started breaking up before it hit the ground. They haven't named a cause and won't for months. The audio people are sharing is her calling mayday and declaring an emergency. That's a pilot doing her job in the worst seconds of her life. Every crash gets a quick scapegoat online. The real answer usually comes much later from the people reading the data. Why do we decide who's to blame before the investigators even leave the crash site?

Most AI features built in Canada today make a call to an API in the US. That's fine until it isn't. One pricing change, one export rule, one policy update in a country we don't vote in, and a lot of Canadian AI products stop working on a Tuesday. Open models are good enough now for most real work: summarizing, classifying, extracting, drafting. What's missing isn't the models. It's Canadian compute you can rent by the hour without a six-month procurement process. The talent is here. Cohere is here. Cheap power and cold weather for data centers are here. So why does running a model in Canada still feel harder than running it in Virginia?

Starting in December, ChatGPT and Codex on Plus, Pro and Business will pull from one shared usage limit. So every time I ask ChatGPT to rewrite an email or explain some random thing, I'm spending tokens I might need at 11pm when a build is broken. Right now I treat them as two different tools. Chat is for thinking out loud. Codex is for work. Once they share one pool, every casual question has a price, and I'll probably start rationing without noticing. OpenAI says most people will see similar reset waits. Maybe. But the people who feel it first will be developers who live in Codex all day and still use ChatGPT for everything else. When the pool runs low, which one gets your leftovers, chat or code?

Apple's "walled garden" isn't a garden anymore, it's a public pool. Five overlapping laptops, a $2,899 touchscreen Mac they swore they'd never build, bolted-on AI, and a new SKU every $200 so nobody ever gets a clean upgrade. We used to pay the Apple tax because they had the taste to say no. Now they say yes to everything as long as they can charge an upsell for it.

Apple didn't build a $2,899 touchscreen MacBook Pro for you. They built it to sell more $19 Polishing Cloths. 🤣

Apple just launched a $2,899 touchscreen MacBook Pro, and their lineup officially makes no sense anymore. Right now you can walk into an Apple Store and buy: 1. An iPad Pro with an OLED touchscreen and a $349 keyboard 2. A MacBook Neo 3. A MacBook Air 4. A $1,999 MacBook Pro with the new M6 chip and no touchscreen 5. A $2,899 MacBook Pro with a touchscreen and last generation's M5 chip name Five different ways to buy a screen attached to a keyboard that all do 90% of the exact same thing. When Steve Jobs came back to Apple in 1997, the first thing he did was draw a four-box grid on a whiteboard: Consumer or Pro, Desktop or Laptop. Four products total. Everything else got killed. Today's Apple never kills a product when it can just invent a new price tier between two existing ones. If you had to cut half of Apple's lineup tomorrow, which one goes first: the iPad Pro or the touchscreen MacBook?

Apple lost their Identity 💀