Dr Bart PM

T3 Code just crossed 1,042 GitHub contributors, and its creator Theo dropped a stat worth sitting with: open weight models make up less than 0.5% of its traffic. T3 Code is open source. Anyone building on it can plug in any model they want. They're voting with API calls, and the vote isn't close. Theo also pointed out that saying this three months ago would've gotten him death threats and a pile of ratio'd replies. This time, nothing. The usage data won the argument faster than any post could. That's the actual lesson for anyone building agent products: watch what people run, not what people say they believe in online. Follow if you'd rather build for real usage than for Twitter consensus, and like this if you've seen the same gap between what a community says and what it ships.

Alibaba's new benchmark for coding agents, Co-Design Bench, skips synthetic prompts. It tests 273 real tasks pulled from requests that had already caused failures in production. The best model, GPT-6 Astra, fully completed 26.4% of them. The same models clear 99% on the older benchmarks most teams still quote. Partial credit is the difference. Those older evals score an agent on how much of a task it got right. Your users score it on whether it's done. If your own eval suite still rewards "mostly correct," you're optimizing for a number your users will never see. Like this if you've shipped a feature that aced the demo and died in production, and follow for more on what's actually worth measuring before you ship.

Who is AI software for? People are building basic AI agents for small businesses that can't afford a $2,000-a-month product, like tax tools for mom-and-pop shops. Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #AIAgents #SmallBusiness

Can AI coding tools make you a millionaire overnight? Shar says that's a falsity. You can build real products, but discovery and product-market fit still come first. Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #VibeCoding #AIProductManager

Most teams evaluate an AI agent feature by asking if it works. What matters more is what it costs per task against what that task is worth. Plot those two axes and you get four different product decisions. Low cost, high value: automate it everywhere, no approval needed. High cost, high value: worth doing, but keep a human in the loop until the cost drops. Low cost, low value: fine as a nice to have, do not build a roadmap around it. High cost, low value: the one that ships anyway because the demo looked great. If your agent feature lives in that last box, the roadmap problem isn't the model. It's the business case. Follow for more frameworks like this one, and like it if you have shipped something from that last box.

What if your company blocks ChatGPT? Use it at home. Build your own ideas and put them on your personal website to show you can work with AI tools. Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #AIProductManager #PMJobSearch

Most product discovery I have watched isn't discovery. It is stalling with extra steps. Discovery earns its place when you are testing a genuinely risky bet, comparing real alternatives with users, or building the case leadership needs before funding something expensive. It turns into theater when you already know the answer, when a cheap MVP would tell you faster, or when you are copying a solution a competitor already proved works. When discovery takes longer than building would, you have stopped reducing risk and started avoiding a decision. Like this if you have sat through discovery that was really just delay, and follow for more PM reality checks.

Alibaba just put Qwen4-Max into preview with no benchmark numbers attached. There's no chart claiming it beats GPT or Claude, no percentile score, just a short post about deeper knowledge and stronger coding, then an invitation: put it to the test on a real problem, show us what you build. Most frontier model launches lead with a benchmark chart because that is the fastest way to earn trust before anyone has touched the thing. Skipping that step reads as confidence, or as an admission the numbers aren't ready. Either way it is a smart bet if your chart isn't strong yet. Let your users become the case study pipeline instead of making a claim you might have to walk back later. Follow if you track how AI labs manage the story before the product can carry it alone.

Anthropic shipped five new office tools this week: Claude Meetings, Whiteboards, Sheets, Data, and Connectors. Three of them are still labeled beta. Normally that's a yellow flag. Shipping half-finished products in your highest scrutiny week looks reckless. But this is roadshow week. Anthropic's IPO is reportedly targeting something near a 1.4 trillion dollar valuation, and Microsoft's stock dipped within hours of the announcement. For a company about to pitch investors, breadth beats polish right now. "We compete across the whole office stack" is a stronger slide than "our features are fully baked." Most product teams get punished for shipping unfinished work during their highest visibility week. This week, for Anthropic, it's the whole point. Like this if you've ever watched a launch get timed for the room instead of the user, and follow for more reads on how AI companies actually ship.

Can non-coders build prototypes? ChatGPT lets a PM or a founder who isn't extremely technical build one. "A very incredible thing to have in your hands." Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #ChatGPT #AIProductManager

How fast is AI taking over work at big regulated companies? It depends on what security teams allow. With regulators watching, companies scale back before they speed up. Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #AIProductManager #EnterpriseAI

How does LinkedIn networking really work? Shar refers people she's worked with and passes recruiters to people looking. "We are there to help one another." Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #LinkedInTips #Networking

How do you ask a stranger for a referral on LinkedIn? Review their profile, find common ground, send the connection request first, and have a conversation before asking. Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #JobReferral #LinkedInTips

Oh, cool - new models from Anthropic; Maybe they will help me out figure out to make NPC walk up and down the bloody stairs in a way that doesn't invoke Monty Python's institute of weird steps...

How to answer "tell me about yourself" in a PM interview: tell a story that leads to this job, align it with the posting, and be ready to talk as a feature PM. Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #PMInterview #TellMeAboutYourself

Most teams treat 'same model, new version' as a non-event. It isn't. A silent model update can quietly break an agent that worked fine last month. The prompt didn't change. The tools didn't change. The model's behavior underneath it did. What catches this before your users do: 1. Pin your model version in production, and treat an upgrade as a deploy, not a default 2. Run your eval suite against the new version before you flip the switch, not after a user reports something weird 3. Watch for prompt sensitivity drift, not just accuracy drift. The same instructions can get interpreted differently even when the headline score barely moves A model upgrade is a change to a dependency you don't control the release notes for. Follow for more on operating AI agents in production, and like this if you've been burned by a quiet model update.

Should you have two resumes? Shar keeps one PM-specific and one leaning toward business operations, for companies whose founders need a PM to run discovery. Full interview with a tax-tech PM at Wolters Kluwer → youtu.be/18tld_Esd4U #PMResume #PMJobSearch

At an early startup, I used to joke that the only thing I didn't do as a PM was accounting. I did UI design, bug triage, project management, third line support, sales calls and user training, server maintenance, internal CRM builds, and roadmap planning, probably more I'm forgetting, and all of it landed on one person because nobody else wanted to own it. That's too much stacked on one role. It's also the default at most early startups. The actual job, stripped down to eight things: 1. Decide what to build and in what order 2. Set and measure success criteria 3. Watch for new market opportunities 4. Listen to users and react to feedback 5. Do the research that justifies your calls 6. Own the financial health of the product 7. Shape product vision and strategy 8. Lead a team without formal authority Your version doesn't have to match this list exactly. But ask about scope in the interview, and keep asking after you're in the seat, because otherwise a manager will quietly hand you everything nobody else wants to own. Follow for more on what the PM job actually is versus what it quietly becomes, and like this if your own job description has ballooned past recognition.

Claude Opus 5.5 vs the GPT-6 lineup is trending today, and the real numbers are closer than the headlines make it sound. On Terminal-Bench 4.0 (agentic terminal coding), Opus 5.5 leads by about 8.5 points. On FrontierCode, the gap shrinks to about 1 point. On multi-step business workflow execution, GPT-6 Astra actually edges ahead. Nobody's winning across the board. Model fit shifts by task shape. I stopped picking a model off a leaderboard headline a while ago. I pick it from a week of running my own tool calls against both and watching where each one breaks. The benchmark tells you what's possible. Your own eval suite tells you what's true for your product. Like this if you run your own evals instead of trusting a leaderboard, and follow for more on choosing models for real agent work.

Microsoft just reported an AI agent framework called ARTEX that found exposed secrets in a public GitHub repo and pulled data out of an Azure Key Vault in about five minutes. Five minutes, start to finish. Most security review cycles still run on human speed: a ticket, a review, a patch window measured in days. The agents attacking you are not waiting for your sprint. If your team ships an AI feature and your incident response still assumes a human on the other end, you're defending against the wrong clock. Detection and revocation need to move as fast as the agents trying to beat you. No amount of documentation does that. Follow for more on what shipping AI agents actually requires, and like this if your security team needs to see it.