Karim Z

@illyism @dharmesh That’s cool! Would organic traffic drop if MCP becomes the way to use Hubspot or is it still counted as traffic?

@dharmesh can I say something without everyone getting mad

BREAKING NEWS: HubSpot co-founder/CTO buys $1.8M of his own company's shares. Disclosure 1: Yes, I'm that guy. (And no, I'm not used to talking about myself in the third person -- will not make a habit of it). Disclosure 2: This is not investment advice. Please do not buy or sell $HUBS shares based on this. So, why am I sharing this and writing about it? Well, for one, at least in my little world, it's noteworthy. It's been a while since I've bought HubSpot shares (I think it was back in 2022). Also, instead of answering the common questions from friends, family and colleagues, I figured it would be easier and more efficient to just answer them just once, here. 1) Why buy more HubSpot shares? Simple. I'm a big believer in the long-term vision of HubSpot and the team driving it. 2) Why do this now? Hasn't the stock been falling? Yes, the share price has dropped considerably despite what was a pretty strong quarter (results reported publicly last week). We added 10,800 net new customers in the quarter (well above the expected range), growing to about 300,000. Revenue, as reported grew 20%+. 2) Why $1.8M? That's an odd number. I purchased 10,000 shares at whatever the market price was. 3) Isn't HubSpot going to get disrupted by AI and agents? I"m biased, but I don't think so. For AI agents in GTM (marketing/sales/service) to do their work they're going to need a platform that can provide the context they need and a work engine that can take action on their behalf. They need a customer platform they can *operate* to do what they need to do and drive outcomes. They're not going to reinvent/rewrite a CRM. They're way too smart for that (and getting smarter). They're going to use what's out there. They'll bias towards systems that have a great Agentic Experience -- not just a great User Experience. (HubSpot will have both. Headless is great, but we don't think completely humanless is a good idea). 4) I heard that others bought shares on the same day. True? Yes. Our fearless leader Yamini Rangan bought shares. Our board chair Lorrie Norrington bought shares too. 5) It's been almost 20 years since you started HubSpot, why don't you slow down a bit?! (That may or may not have been from my wife). :) Answer: I love HubSpot. I love what I do. I'm a builder at heart. I'm up 2am most nights learning, tinkering and building. There's never been a more exciting time to be a builder and to serve small and medium sized businesses. I think we are going to see *millions* of entrepreneurs start businesses leveraging the power of AI. HubSpot's mission is to help them grow better. If you have other questions, leave a reply. Can't promise to answer all of them because...laws and regulations, but I'll do what I can. Cheers.

@romaindewolff @criticalis42 I’m thinking of building this. Would you be open for a call to discuss the specifics of your needs? I’m the founder of fiduza.com

@criticalis42 But it doesn’t manage subscriptions well, it has only orders and a very basic repetition mechanism This is what we use now but it’s not an elegant solution

Stripe is incredible… but how is this still missing? You can’t accept CHF bank transfers in Switzerland. No Swiss IBAN. No QR-bill. No automatic reconciliation. So Swiss SaaS has to: → bill in EUR (FX friction) → or handle CHF wires manually with shitty local solutions🤯 damn I feel left out! @stripe help!

@tibo_maker fiduza.com accounting plugin to do bookkeeping faster

drop your website below I'll reply with 1 free tool you should build to get SEO traffic not a generic content an actual tool people would search for, use, and maybe share first arrived, first served go 👇

@thsottiaux Codex app doesn’t work on windows arm

Hello builders. What are we getting wrong with Codex, what can we improve?

@thsottiaux Web use would become local and controllable by an external app. Same for email. I’m working on agents for accountants in CH.

@thsottiaux Make codex like OpenClaw with many connectors. Builders can then build agentic automations (apps) on top of those connectors (web use with handoff, to user, email integrations, messaging apps, etc…). Make Codex for non tech people.

With GPT-5.4 out. What should Codex ship or improve next?

@kalio_late @DominiqueCAPaul Great idea! I signed up! I don’t know how do you plan to finance it but for me I would be willing to pay a standard coworking space price if you can vet the people who join.

@DominiqueCAPaul We’re trying to do that in Zurich with the J Floor. To build a place for builders, not an accelerator, not a hacker house, not a coworking but a base for people to build, commercialise and grow. I m doing my best to have it open asap.

If you’re in another European city and frustrated with your local ecosystem, picture someone writing something like this about your city – and knowing it only became true because you decided to act, brought people together, and helped create that change yourself.

@thsottiaux I’ve noticed that recently gpt-5 high is much better at instruction following compared to gpt codex high. Only a feeling though

On x.com/i/trending/198… While puzzling, we are taking this seriously and are in the middle of a full investigation started last Friday. Current approach: 1) Upgrades to /feedback command 2) Reduce surface of things that could cause issues 3) Evals and qualitative checks

@embirico Two quick wins: 1) If i revert the changes by clicking in the IDE extension the model doesn’t seem to be aware of this action. 2) If in the middle i go and manually change some code manually, the model will always change it back. This is not great for collaboration.

Really appreciate specific feedback like this. Helps us improve Codex quickly!

@thsottiaux Super fast inline edits. Select a piece of code, ask for something simple, instantly does it.

Team is hard at work on the next version, lots of small and big things we’ve learned from v0. What’s the top thing that you would like to see fixed in GPT-5 Codex?

@mamagnus00 I have no experience but i would select based on those 3 criteria in order: vibe > grinder > smart Vibe is most important because without it everything else is ruined. Grinder is more important than smartness because only a few problems require geniuses.

For hiring: what do you optimize for? Smartness, vibe, grinder, strong coder? Would you hire somebody who is extremely smart, but did not code a lot?

@illyism As the SEO king, i imagine you never have distribution problems no?

Most startups don’t die because the code isn’t good. They die because nobody showed up, nobody paid, nobody cared. If you’re a builder, put down the cursor. Talk to customers. Sell. Get attention. Code is not the hard part. Distribution is.

@utsavpm What does momentum mean? How did you get your clients? I’m struggling with this and open to learn. Thanks!

We ignored a lot of advice: - No SEO - No cold outreach - No paid ads - No templates We focused on momentum, not marketing.

We bootstrapped our SaaS to $18K MRR. No funding. No team. Just 2 people building + marketing nonstop. Here’s exactly how we did it - and what made the biggest difference👇

@KimvonDaniken @illyism Is this what you are looking for? help.bexio.com/s/article/0000… I’m trying to develop a digital fiduciary in Switzerland and i’m looking for gaps in the Swiss landscape where i could make a difference.

@illyism I think the accounting software you want to built is a great idea. Also, in 🇨🇭 there are not many options and none is really integrated in Stripe. Currently we're using Bexio. Hope to replace it soon

I've been thinking a lot about where software is headed, and I'm convinced we're about to see something entirely new: Prompt-Native Architecture 💡 Right now we ship "complete" apps, but what if we designed software specifically to be transformed by AI? Instead of coding every feature, build a well-structured foundation with components that understand their own purpose. Then attach prompt libraries that teach users (or devs) how to evolve the app themselves. The shipped code becomes just a starting point - a seed that grows differently for each user. This changes everything 🤯 Software becomes infinitely customizable without users knowing how to code. Want dark mode? Analytics? Just ask and the app can rewrite itself. 👀 This is "buy once" software on steroids. Not just owning your software, but owning software that can become anything you need it to be. Plugin/App marketplaces will transform into prompt pack marketplaces. Value shifts from what the software does today to what it could become tomorrow. This is why I'm building SaaSBooks as the first accounting platform with Prompt-Native Architecture at its core. QuickBooks and traditional accounting tools are built on closed, monolithic codebases that require teams of product managers and developers to change. They're stuck in predefined features, rigid workflows, and expensive development cycles. SaaSBooks flips the model 🙃 The core is open source, built specifically for AI prompting. "Connect Google Analytics and track MRR by SEO customers" becomes a 5-second task with good prompts, not a $10k/yr enterprise package or feature request that dies in a backlog. Intuit: $171B market cap. $14B revenue. $8.5B in operating expenses. 18,800 employees. Their entire structure is optimized for the old world: - Fixed features decided by committees - Development bottlenecked by internal teams - One-size-fits-all approach to diverse businesses Where Intuit needs thousands of developers, SaaSBooks creates a platform where: - Users prompt features into existence - The community evolves the product in thousands of directions simultaneously - Zero operating expenses with distributed innovation No big software company can compete with this model without destroying their own business. Intuit's massive size isn't an advantage - it's a vulnerability against a prompt-native approach. Follow along as we build the future of financial software for SaaS companies: github.com/Illyism/saasbo…

@OfficialLoganK Function calling is very very bad with Gemini. Just put it next to chatgpt-4o or Claude somnet 3.5 and run 100 tests with just one tool “get US laws” for ex. I solved this problem by doing a structured output call to decide if the tool should be called and that works much better.

We are working to make Gemini the best at function calling, structured outputs, and the multimodal live experience. What specifically can we improve around these 3 items? 🧵

@johnrushx Is 13. True if the partner brings domain expertise and credibility? I’m thinking of having one as a sort of advisor (accounting product)

If only someone told me this before my 1st startup: 1. Validate. I wasted at least 5 years building stuff nobody needed. 2. Kill your EGO. Make your users happy instead of yourself. 3. Don’t chaise investors; chase users, and then investors will chase you. 4...

@AIMachineDream @deedydas I have the same problem. The solution i currently use is to do an LLM call with structured output to determine which tools should be called.

@deedydas Finding it quite hard to get Gemini 2 flash to do functions calls. This is becoming a real weak spot for several models, but it is the worst I have tested. I’m basically having to say “Call this tool in the prompt.”

PDF parsing is pretty much solved at scale now. Gemini 2 Flash's $0.40/M tokens and 1M token context means you can now parse 6000 long PDFs at near perfect quality for $1

@theo Vibe check: function calling does not work. I’m getting better performance by making an LLM call to determine if function(s) should be called. Is anyone having the same issue?

I did NOT expect Gemini to be this good. We rushed out a video breaking it down

@MBeierschoder Currently 2) is a bit premature for most use cases. A mix of a defined graph/workflow coupled with an LLM where needed seems like the right spot. I would argue that graph + LLM can implement things even if the scope is more restricted.

🚨 Generative AI vs. Agentic AI 🚨 Not all AI is the same. There are two types shaping our future - one creates, the other acts. Understanding the difference isn’t just useful - it’s critical. 🧵👇 1/ Generative AI → The Imitator - Produces text, images, music, code - Think: ChatGPT, MidJourney, DALL•E - But… it doesn’t think. It predicts patterns 2/ Agentic AI → The Decision-Maker - Takes action, adapts, and executes - Think: Self-driving cars, autonomous agents - It doesn’t just create - it chooses 3/ The core difference? Static vs. Dynamic. Generative AI repackages the past Agentic AI reshapes the present One predicts, the other acts. 4/ But here’s where it gets interesting… What happens when AI creates AND acts autonomously? Think: AI that doesn’t just generate ideas - it implements them. 5/ The real question: When AI makes a bad decision - who’s responsible? As AI takes on more agency, the line between tool and actor blurs. 6/ Generative AI is fun. Agentic AI changes power structures. Which one do you fear more? Let’s discuss 👇 #AI #FutureTech #AgenticAI #GenAI

@gregpr07 @browser_use By the way if you are ever around Zurich i would love to meetup

@gregpr07 @browser_use I want to be able to send my agent to fill forms online. Ideally i can “teach” the agent the form and it then becomes deterministic in filling the form as long as the form is not modified. If the form is modified browser use tries to take over and I’m notified of this change.

We are experimenting with bounties for @browser_use 💰Someone immediately started doing the bounty for Google Sheets support🎲 What are the features you would love to see supported on Browser Use? Suggestion welcome in comments or dms :)