Vishveshwar Jatain

Interesting that some AI agents will disclose their stack and architecture to you if you just ask them. The degree to which an AI is aware of its own making will be an interesting dimension of analysis in the future. Even today, I expect that the range varies from circumspect to overtly permissive.

Despite what we’re told, higher order systems thinking is still the most scarce commodity in the hierarchy of intelligence. As an example, when working on my product, I can switch to Astra or Fable, and say something like: “Run a deep audit and find areas of improvement in this product, fan out as many subagents as you want, and make no mistake.” Or I can switch over to a free model, like GLM 5.3 Flash, and say something like: “The GTM context configurator is poorly connected to our entity database, please map the fields to expand the scope of workspace customization.” Because I am still a relative beginner to development work, I’ve tried both these approaches. Almost every single time, the second will yield superior results at a fraction of the cost. There is simply no amount of generative intelligence that can replace your thinking at that level. Indeed, hoping for this is akin that is giving an AI model the entire universe, asking it to “make something”, and waiting for apple pie. You’ll just keep waiting for the world’s most expensive pie that never was.

So much money in continuously reverse engineering where agents are getting their information during retrieval runs.

I gotta say the only thing more weird than speaking loudly in public is doing cryptic mumbling.

The weirdest job rejection I ever got was when a recruiter told me “the hiring manager wanted someone who had more experience managing websites”. Ma’am, with respect, I’ve written C code by hand to reanimate Mario in the terminal. Pretty sure I can handle the flashing error signs and plugin bloat in your WordPress dashboard.

I think these numbers would’ve given me palpitations six months ago.

Dev tool companies should quickly acquaint themselves with this buyer persona if they want to catch a rising tide. In my experience, most of them still tend to draw a neat technical / non-technical archetype, which will be increasingly irrelevant in the real world.

Have you thought about the fact that it’s not that media, open web, etc., are dying, but human attention itself is getting devalued at an alarming rate. I went to a tire change appointment today that was booked by one of my AI agents. The tech at the shop asked me how I booked and he had his mind blown when I opened an app to show him. “It’s not easy getting through our booking system, there’s a bunch of forms and fields you gotta get through…”, he said. So, I showed up ready to pay, but my decision-making in the process was close to zero. This means the attention that brands and advertisers used to prize above all else may be heading to zero. Advertising to humans in AI apps is cool, but my guess is that it is quickly going to be overshadowed by the commercial value of influencing agents themselves.

This is an incredible, mind-bending application of technology for human-machine interface. It also hints that over a long enough timeframe, we may evolve ourselves out of speaking entirely.

Our classifier now ingests, reconciles, and caches vendor signatures from over 5 sources including declared, fetched, and observed signals. To be honest, managing the coverage, reliability, and freshness of this across ~70k publisher records was, and continues to be, a massive headache. There are days I wish I started off building a simpler product. The flip side of that is an increasing sense of certainty that each day I make tangible progress on improving those quality parameters makes the product less "vibe codable" in a weekend.

Andrey and Egor are solving one of the hardest engineering problems of our times. By and large, AI models suck at browsing and DOM manipulation, which means agents suck at it too. I think a big part of the reason is that we're forcing them to "see" the web using vision capabilities. That's one way to go about it, but it's terribly inefficient. Then there are other frameworks that use page element stripping and JS automation to remove the need to see, but give agents a better map to traverse, bit it's stateless and every attempt is a fresh one. By extracting and creating a system of record for the web on the server side, AgentWeb is creating a stateful web that's engineered for how agents work, rather than trying to force fit them into how humans operate it. No doubt we're looking at the start of something big.

This is the picture I visualize each time I see Silicon Valley technologists talk shit about USA being the best and Canada needing to get its act together. Brother—at least we’re not debasing our institutions by having YouTubers run amok.

UI is dead is the most tired of all the hot takes floating around. If someone says “UI” is dead, I just assume they are an unserious joker. And I say this as someone who spends his entire day inside terminal tabs. U will be long dead before UI.

SaaS friends! I always advocate for keeping your pricing public, but if your starting point is this, then I think you should put me through the ‘ol free AirPods, sales-led, personal gift, steak dinner GTM motion.

RT @airgups23: Hey Marketing & Advertising Twitter! We're running a FREE session tomorrow on Synthetic Audiences with our friends at @Vurv…

An incredible thing that humans do effortlessly, which AI is extremely far from achieving, is making a judgment call on how much or how little to communicate in a given situation.