Chinmay Nivargi

Find myself preferring Opus 4.6 over 4.7 for PM artifacts. 4.7 feels like a smarter ChatGPT o1 or o3 - very smart, but not really readable. Anyone else have the same experience?

Of course I had to try to create the Mt. Rushmore of tech companies discussed by @benthompson and @gruber. Nano Banana and ChatGPT were quite good - and Grok and Meta AI came nowhere close. Same prompt for each. Guess which is which.

@andrewchen A coworker has agency not only to break a goal down into steps to achieve them, but also to set them in the first place. Agents are getting good at memory and integration, but not yet on true agency. Agents working together, challenging each other are paths to RL our way there...

ai is shifting from “chatbot UX” to “coworker UX.” it’s not just about having smartest model, they’ll have the best handoff between human judgment and autonomous execution I think this is what we see in explosive hype around openclaw. What's magical about it is that it can be proactive, it can self-improve, it can link into your accounts so that it’s triggered. These are the things you need to be a coworker, chief of staff, colleague, etc rather than just something that is a smarter google search. What people underestimate is that the interface paradigm itself is changing: - Chatbots assume every task begins with a prompt - Coworkers don’t wait for prompts They watch the environment, notice patterns, and surface things before you ask. The best AI systems will feel less like “ask me anything” and more like “I’m already working on it.” that requires three things that chatbot systems historically lack: memory, agency, and integration. Memory so the system understands your projects and preferences over long running periods of time. agency so it can break goals into steps and execute them. integration so it can touch real systems, like email, docs, repos, finances, calendars, APIs. Once those exist together, the model stops being a tool and becomes a participant in the workflow. we are sooooo close to having all this, but not yet... the claws show a glimmer of the future. So the question is, which agentic systems will know what you’re trying to do? Which ones can take partial direction and move the ball forward? Which ones learn your style and anticipate the next step? Feels like we're almost there, and likely to figure this out in 2026. am very very pumped this is about to happen.

Now I know how @Spotify characterizes my music taste 😅

@lennysan @amankhan and @talraviv have reinvented the technical tutorial here - and Cursor seems to be a perfect substrate for it. This is way more approachable in the age of coding agents, and improves upon Jupyter notebooks. Although, maybe we are trading off some comprehension for personalization and vibes!

@matt_elevenlabs 11B

We just raised $500M at an 11B valuation 🎉 To celebrate, we’re giving away 1,000 free credits so you can test our platform. For the next 6 hours, comment “11B” below and we’ll DM you the credits (must follow) 👇

A retcon that fixes everything! You’re welcome.

@TrungTPhan @RampCapitalLLC Perhaps it’s possible to set up an AI workflow that asks Grok daily and populates this list

@RampCapitalLLC last few years stopped by third week of jan, need to finish the spreadsheet this year!

started the X Main Character tracker for 2026

@pitdesi Sad ending - it looked so good!

I really wanted to love this as a minivan enthusiast and cute vehicle lover, but nobody wants a van with less than 300 miles of range All of the bad EVs (most of them) are getting flushed out of the system with the end of the EV tax benefit

@devangbaheti @patrickc @stripe I didn’t read it as loans rejected for business that accepted them. It is loans not offered at all, which may not have as much of an impact on affinity and volumes.

I'm very excited that we now have results from an almost two-year randomized controlled trial that we ran across thousands of businesses on Stripe: those that accepted a loan from @Stripe Capital grew annual revenue around 27% faster. (Two years isn't that long in the scheme of things, but it is when you're waiting for the results of an experiment you're interested in.) We had two kinds of controls: businesses to whom we offered loans but didn't accept them, and a holdout group of businesses to whom we randomly did not offer loans but which were otherwise identical to those to which we did. As such, we feel confident in the causal nature of this conclusion. While this might not sound like news ("capital increases growth"), I think the finding is a good reminder that many businesses are still quite capital-starved (this effect is on top of all of the other sources of capital that businesses have access to), and it is consistent with what we hear directly from businesses in surveys. Beyond Stripe, inefficient capital allocation at economy-wide scale is likely a major bottleneck to growth around the world. We have an ambitious roadmap planned and we're very much looking forward to expanding worldwide access to growth capital.

@carlvellotti @shobhitchugh CC

Every PM should be using Claude Code. So I built a HUGE course for you to learn Claude Code... IN Claude Code! 🔹 Complete guide 🔹 Make PRDs, analyze data, create decks Soon, I'll sell it for $149. For the next 24h: FREE! Follow + RT + comment "CC" & I'll DM it.

@lulumeservey For those wondering- benning.army.mil/infantry/199th…

Everyone working at a startup should read A Message to Garcia

@ValerieTetu A/B testing will still continue to answer the question - On average, did the dynamically adapted, personalized feature lead to better conversion / engagement / revenue? And by how much?

I question the future utility of A/B testing for products and websites. Why would we optimise for the "best average for all" in a world in which interfaces, content, and features can dynamically adapt to each individual user?

@ankrgyl IMO Evals ≠ Experiments. They serve different layers. Pre-launch evals = input metrics → ensure the AI feature itself is robust. Experimentation = output metrics → measure actual customer impact. Evals test the model in isolation; A/B tests test the whole system in the wild.

This is the exact opposite conclusion to draw from the acquisition. AI enables you to build much more dynamic products that evolve faster (and even automatically). The foundation underlying that is good evals. A/B tests are officially the way of the past.

@thomyorke channeling his inner @bjork here open.spotify.com/track/4cug7RkM…

@sariazout For me the key is to actively inject a lot of my own thoughts in the prompts and followups vs. giving into the vibes. It still may be a distraction - but I find that LLMs help me get over the hump of creating the mental space and focus to sit with my own thoughts.

I’ve been spending 4+ hrs a day talking to LLMs and I love it, it feels fun. But I’ve been having this feeling lately that it may just be another distraction and maybe if I just sat with my own thoughts longer, I’d get somewhere better AND faster. Not sure. Curious if others feel this too.

@jspujji 🚀

The fastest way to waste a decade is to optimize for goals you never chose. Every week, I meet operators smashing revenue targets yet privately admitting they have no personal North Star. That mis-alignment bleeds into hiring, roadmap…even family time. My fix is a one-page Dream Future State prompt. 15 mins with it drags your ACTUAL ambitions into daylight. Comment “🚀” and I’ll DM it.