I thought I have prompted AI to do some creative and unique design. Now I see the same tricks and colors being used on half of websites. Prompt harder or learn design? 🤣

@dhh I have wrote you a letter 🤣 blog.reqproof.com/p/dhh-this…

When you are measuring how intelligent model is, in fact you mostly measure how good it was following the plan, or question, based on your own intelligence. Difference is subtle, but crucial. And it is so easy to be confused, while measuring small model performance, like various Flash variants or local models. Smaller models became quite good at doing the work under the well specified goal, like fixing the well specified bug. But can it find the bug, without you giving it any clues? In my benchmark the answer is mostly No. Being able to invent the questions - thats the real intelligence - and thats where the big models show true strength. Understanding this nuances - is key to efficient orchestration. Full report below

I used to delete a requirement file and call it done. git rm takes it from covered by a dozen traces to does not exist in one commit. The traces that pointed at it just go quiet. The audit cannot tell whether the rename was intentional. Now I supersede it. The old id stays, with a pointer to the new one. Otherwise the audit treats a rename as if the requirement never existed.

Grok SuperHeavy subscription is the best value you get on the market now! It comes with almost unlimited Grok usage, X premium, and $200 Cursor Ultra package included! The last one gives you near unlimited Grok Bot usage too - which quite a powerhouse hidden behind simple UI!

I had a quite bizarre experiment, with project call Xerj, which challenged the way I think about the coding, but also challenged the ethical part of it. It helped me save tokens and time on some some workflows up to 5 times, while having more quality code. Lets be honest, majority of projects has nothing novel. Or novelty achievable though the combination of the well known blocks, and packaging them in a new way. What if you just say to the agent like "Here is the idea - find top 10 OSS projects, which has parts of this story, and just do similar". And it kind of did it... Very well actually. Xerj, which I mentioned earlier, is basically local search engine with ES compatibility, which natively index the codebases, and build for AI agents. It downloaded 10 projects, put them to index, and agent used it for the reference while coding. As for ethical part, if we think about it LLMs are already big copy pasting machines, what I described above just brings it to the local machine, more accurate, and helps you save the tokens, and it kind of works... This reference projects does not even have to be in the language you use, it will rewrite it just for you, only the parts you need, and it really challenge all the licensing ideas we had over the years. But it solve some real corporate issues too. As engineering leader, it actually solves the issue where each team re-implements similar patterns in different projects but slightly differently, and you wanna keep the same architectural patterns across the org. Another use case I want to try soon: do I really need ANY dependencies? In the majority of cases I use 10% of a library's capability, while the rest is just CVE surface. So why not copy only what you actually need from the dependency? And if you hit a bug, you just copy more... So I guess thats where engineering is going?

Can software engineers even be called “engineers”? Have we really earned this title, or just inherited it?

If AI writes the code, what exactly is my job? The question reveals more than we want it to. If removing the keyboard also removes our engineering identity, perhaps coding was the identity all along.

This. But single markdown file is too limiting for big projects. I found elegant solution, where you can turn docs of any size to mcp, and chat with them locally. It actually works very well with code chat too! Inside it elastic search like engine, working locally. Link below.

For 100% reliability, limit to 1 tool use per ai call, and make it run in agentic loop until it complete task, each time calling one tool, and in the end special complete_job tool. Plus own tool calling format, some simple xml syntax. Code editors like Cline/Roo do it like that.

Current MCP trend is killing seat-based SaaS economy. Why pay of seat, if I can have amazing UX via MCP. We moving to the future of limiting public API access, value based pricing, and official "virtual employee" agents, from every major SaaS vendor.

The recent agent-to-agent protocol more than just a technological solution, it is attempt to "fix" the SaaS economy and API market, and moving to value based pricing. First of all, the recent agent-to-agent (A2A) protocol that Google introduced isn’t really some massive tech breakthrough in itself. It’s more like an attempt to fix the very shaky SaaS economy, which is getting hammered by “shadow API usage.” Think about it: why pay for an extra seat on a SaaS platform - like Zendesk - if you can just integrate its API into your chatbot and let everyone in your team access it? Vendors often can’t even tell if your requests are coming from AI or human users. Historically, API usage wasn’t a big threat, because sure, you could export data, crunch numbers, or build some custom scripts, but replicating a polished user interface or full product experience was too much hassle. Now, with MCP and agentic stuff, you can deliver an experience so good it might actually beat the vendor’s own UI. That’s alarming for SaaS providers because you basically get 90% of the value at 10% of the price - and they lose control. Google’s new A2A protocol helps address this issue. From a pure technology angle, I like it. You can schedule tasks, define “skills,” and basically treat these agents like virtual employees. It’s probably the direction the industry is heading. We’re also seeing major companies sign up for this kind of format, because it preserves their revenue streams. I fully expect more vendors to limit or even block their existing APIs, raise prices, or make them available only through premium tiers. (Twitter basically did this recently, so maybe they just saw it coming.) Why would they do that? Because most SaaS vendors want total control over their products and how they’re used. With these “shadow” MCP integrations, they lose the ability to track usage or enforce pricing. So their plan is to restrict API access and push official “A2A agents,” selling them as virtual employees. In the future, you might just visit some developer portal and rent as many virtual workers as you want, each capable of parallel tasks, etc. Over time, I see public API usage dropping off, but total API usage skyrocketing - these agents will make a ton of calls behind the scenes. MCPs are still in a technical adoption phase right now, but they’re going to evolve with more user-friendly experiences, and vendors will likely shift to charging for the actual value they deliver rather than billing per API call.

This. I have UX and PMs sending PRs and fixing bugs, all with tests and security checks. And its not some vibe coding, we are building serious enterprise software. Thats one of those magic AI moments when we introduced it (integrated via github actions), you just create a ticket and you get PR with fix, passing tests, in 5 mins. Quality is always a tricky. All those low/medium priority issues, which in past required creating ticket, waiting for a month for engineering resource to fix (and to be frank usually dying in the backlog), now just get fixed right away 🤯 And you do not need any special stuff for it. You can achieve it with OSS tooling, and standard Claude models. probeai.dev/integrations/git… Your also can see how to run it outside of github with it probeai.dev/cli-mode#enablin… My DM always open if you need help configure it, or show some latest cool stuff.

Here is the guide on how you can add super capable AI agent triaging your Github issues and reviewing PRs, for free. It leverage new Quasar Alpha model from OpenRouter, which is very capable! 👇

Now you can chat with documentation or code, via MCP, just by specifying github url - all search happens locally! More over you can create your own MCP servers, with pre-baked data! The easiest way to create MCP so far! All free and Open Source. Link below 👇