Boris Djordjevic

Google switched off my site in one hour. $50,000+ of my own money. Solo founder, self-funded. No warning. No explanation. No one to ask. 27 Oct, 22:00 UTC, 25 days after launch: ~360 Google impressions an hour, then under 50. No manual action. Just a black box.

LLM providers financial incentives will steer their models toward creating this dopamine hole; your output may skyrocket but output does rarely mean you are moving the right needle as far as your goals are concerned.

My agents kept dying mid-run on "You've hit your session limit." So I built Headroom. It's a Mac menu bar app: when one Claude account runs low, it moves Claude Code and Claude in Chrome to an account with room. Your agent keeps going on its next request. While an agent drives Chrome, or you've picked an account yourself, it holds. Only running out moves it. 7 accounts. 188 switches in 4 weeks. Free and open source. Tell your agent: install headroom from github

Fixed the iPhone Duo screen line. You're welcome, Apple.

And the total market currently addressed for GLP-1s is still in the single digit!

Your AI agent can use X now. Properly. I built xmaster so my agent does my X admin, not me. It reads, posts, replies to any thread, and finds posts worth replying to in about 3 seconds. Free, through your own browser session. Tell your agent: install xmaster cli from github

Because I have 10+ subscriptions, because I like minimalistic and easy to visualise, please enjoy the most advanced LLM Switcher yet! github.com/paperfoot/claude-…

Just built an app to get my son to practice his times tables in a Duolingo-like gamified manner - initially built with Codex but then Opus 5.5 has improved many aspects of it. There's a lot more that can be done, and I'm currently applying the same template on more advanced things he can start learning! apps.apple.com/us/app/simply…

Made with AI? It's SI now. Time to update the label @elonmusk @benjitaylor

I've been steadily using more and more Grok 4.5 high fast, and it's been doing multiple workflows (in automation, logic, engineering and using skills) where its output was comparable to Fable 5. The speed and efficiency was way beyond. I'm very excited for Grok 4.6/4.7!!

Would you trust an AI to tell you if your dog or cat needs a vet? Building petus.ai a free pet health, longevity, and digital co-pawrent.

Petus is on Product Hunt today. It's the thing I've spent the last year building: a free AI vet for your dog and cat. Learns your pet, keeps their whole health record, answers the 2am questions so you're not guessing. Free, no card. Quietly proud of this one. A look or an upvote helps a lot: producthunt.com/products/pet… What would you ask it first?

⚪️ New app: Petus, a second brain for your dog or cat. One AI companion for their health, history and everyday life, made to make your life easier and theirs longer. Free, no card: petus.ai

My review got GPT5.6 sol xhigh/ultra: I'm thoroughly disappointed, in how unreliable it is - it suffers from some of last years models issues e.g. hallucinations, ADHD (not completing or wiring tasks) not orchestrating things correctly and not remembering details or instructions or just not implementing what it has been explicitly told to do. I've ran 4 large projects (and a couple of small ones) in the last 36 hours that have been running pretty much constantly, and all of them suffered from large mishaps, where GPT 5.6 went an did things that were unnecessary, dangerous, took actions far out of scope of its assigned job and also while being completely misinformed. 1. It went on to my Twilio account and tried to reset passwords when I use social login, autonomously - and there was nothing that was related to its access of Twilio in the instructions I've provided. It just went to hallucinate tasks and improvements out of scope of its job. 2. It assumed I owned domains for a website I've asked it to analyse, and tried to access my GoDaddy account. 3. In another example, GPT 5.6 spend roughly 6 hours in a loop trying to flash a Rabbit R1 wholly inappropriately trying over the same way over and over again (which is the definition of stupidity as per Albert Einstein) - it's like a dog running after its own tail, but I feel I'm insultingh the dog comparing it to GPT 5.6. 4. For a 1M codebase platform project, as I kept getting feedback, I would add more prompts to queue/steer the conversation, and while the prompts were aknowledged and its intention to carry the work was written, it completely forgot to implement about 60 percent of those tasks. 5. It's making too much junk in my local folders, screenshots, MD files, that it does not clean up properly. 6. It misses the obvious; e.g. I'm asking it to review the CSS code and ensure there are no inconsistencies; it goes on to work on the copywriting, changes the layout drastically, and leaves the small details that I was hoping it would catch (obvious to me reading the code or inspecting visually) completely untouched. 7. I've seen signs of its destructive capabilities, where GPT 5.6 went on my production databases and prepared plans to modify or take wholly unnecessary and inappropriate actions, in ignorance, and without inspecting and looking at what was there. Luckily, nothing happened, but I did feel goose bumps as I inspected logs and tool calls. 8. It's not able to educate itself well - when I ask it to research on a stack or technology we plan to use in a project, it looks very superficially at only a few sources, despite my prompt asking it to search extensively official documentation, user feedback, issues, to search on X, github for example projects, etc. which means in many cases it's working on old or broken assumptions; I cannot trust it to implement the SOTA without my explicit interventions - and sometimes prompting and steering the conversation toward that outcome does not help as my prompts are ignored despite being acknowledged. For the positive: it has done some things well in term of user interface for a macOS app, at times its reasoning is really good and catches unexpected unknown unknowns. I'd still overall rank it (much) lower than Opus 4.8 for working on complex projects. Where GPT 5.6 shines is doing the first 70 percent of a website or iOS app - it does things beautifully, not as well as Fable 5, but getting close to. For any work that requires less than >1 hour, on a new codebase, or empty project, GPT5.6 is nice and more 'fluid' to get early work done, while my Fable 5 workflow for equivalent quality requires more prep time. As a disclaimer: I've used Codex MacOS app. Same prompting and workflows as I use for Claude Code (in Terminal). Over time I'm sure that I can adapt and compensate for its shortcomings with more precise prompting, more patience and micromanagement, but that's not a direction I enjoy having to take for a mid-2026 LLM model.

I spent the last year building an AI that helps you decide if your dog or cat needs a vet. It's called Petus. It's free, no card. I need pet owners to try it and tell me what's wrong with it. The first 100 people who send real feedback get a year of the full version free. What would you ask it first? petus.ai

OpenAI built GPT-4b with @RetroBio_ to optimise OCT4 reprogramming. 199 Bio is working on the same problem. @sama, would you open it up more broadly?

Digital creators are pivoting from luxury into AI and tech. @xeniatchoumi at Cannes this week is one of the cleaner examples.. buried in this piece is her move into AI and digital entrepreneurship. Who else is doing both lanes at once?

The thing I love most is going deep into many fields at once, then watching the connections appear between them. AI has just made that possible at a scale we never could have imagined