Božidar Benko

Codex playing chess in Mac Chess.app.

@borna_perak I'm already overbooked. :) I need to book a dedicated office for the next time. Plus, as you know, my X audience is huge. My post got whole 9 views. Might be the negative record!😂 Not sure what the X algorithm is.

@bbenko Love the format Boz! I'd encourage you to tweet it out at least a week before the meetup though. :) Wouldn't have made it either way, but would love to come to the next one!

Most AI events are either vague hype or thinly disguised sales. East Bay Tech #2 is trying to be neither. Tomorrow in Lafayette, we talk about what’s actually working with AI right now. ebt.fyi/meetups/2/

I went to a meetup in SF where someone from Anthropic was supposed to speak. Anthropic shut it down. The industry wants to look open, but a lot of these companies are still tightly managed, risk-averse, and scared of unscripted public conversation.

@ivanburazin What I find most interesting about OpenClaw is how easy it is to program an agent that combines multiple tools and data sources in a single workflow, then takes the action at the end. That’s when software starts to feel personal.

When people were using Claude Code with native tool integrations, there were always missing integrations, partial data, and frustrating reports. Now OpenClaw with full credentials has access to all the tools a remote knowledge worker would. It can: - log into websites - fetch data itself - use 2FA on the phone - access any system we can access I've been playing around with it for the last couple weeks, and it's literally a remote person working full-time for you.

Your startup may still be building something impressive, but the business around it could already be obsolete. Sunk costs kill more startups than bad ideas. steveblank.com/2026/03/17/you…

Most startups have always been dead on arrival. AI is just going to make that more obvious. When building gets cheaper and faster, bad ideas, weak distribution, and fake moats get exposed a lot sooner.

@borna_perak Here's a good list of most AI writing tropes: tropes.fyi/tropes-md

"NEVER use em dashes (the long dash character)" must be the most commonly instructed line to any AI agent

I had a joke, but it got lost in the context window.

@borna_perak You lucked out because you are not afraid to try and use the best tool for the job, at any time. All these tools are/were great at their moments in time. youtube.com/shorts/GBX-3cA…

@bbenko I lucked out and have started my career with SVN (subversion) which was tad better than SourceSafe I think

Vibe coders generation will never know that git merging was once a nightmare

The frustrating part of this next wave will be all the people who vibe-coded an app with Claude, never shipped it, and now feel weirdly qualified to dictate the technical roadmap. Demos make people confident. Production makes people humble.

⁠Spent my morning debugging why OpenClaw broke after an upgrade. Logs, config, version diff, second-guessing reality, and of course, it was the cache.

I love AI software and I’m building on AI startup, but there are real traps founders should plan for early: 1) Customer experience: people tolerate AI when it’s fast and useful, they hate it when it feels like a maze. 2) Trust: if customers think the system is hiding, guessing, or dodging, they bail. 3) Technical edge cases: accents, noise, interruptions, and weird requests still break flows. 4) Human handoff: if escalation is clunky, one bad call can burn the relationship. 5) Legal/liability: AI should never make promises, quotes, or commitments outside policy. 6) Market fit: some operators don’t want more leads, they want better-qualified jobs. 7) Brand risk: for premium services, “AI-first” can feel cheap if not done carefully. 8) Integration reality: no CRM/calendar/dispatch sync means no operational value. 9) Maintenance cost: launch is easy, reliability and continuous tuning are the real work. 10) Competition: you’re not just competing with old software, you’re competing with solid human service. AI can absolutely win here, but only if reliability and trust come before novelty.

OpenClaw Pam is now the most useful member of our tennis group chat. One guy wanted to move the match later, another pushed back, nobody could agree, so they told Pam to pick the time. She did, and everyone accepted it. We are so cooked 😂

@tmrcv I'll try github.com/comet-ml/opik-… . 🤞@vincent_koc also mentioned OPIK on a recent meetup.

@bbenko When I worked on agentic video classification, Opik was 🔝 for tracing these kind of issues.

I asked the same question twice using OpenClaw + Codex + my custom skill and got two different answers. First run it picked the wrong day for today, so the output was for the wrong date (tomorrow). Second run, same query, and it was correct. It picked the right date for today. I then tried to reproduce the error multiple times so I could fix it, but couldn’t reproduce it. That’s the reality: custom skills help, guardrails make LLM automation dependable.

Just vibe-coded a site for our meetup ebt.fyi Small, in-person AI discussions in the East Bay. Real workflows, model releases, and what’s actually changing in product work. If you’re in the Bay Area (especially East Bay), come join us this Friday.

OpenClaw on a VPS is easy to set up, safe when you harden it properly, and free to try on AWS Free Tier and Codex. I wrote the full step-by-step with real commands: sidegem.com/blog/openclaw-…

OpenClaw’s superpower is also its risk: it acts first. That’s exactly why I use it, and exactly why I don’t give it broad access. Use a sandbox, least-privilege permissions, and hard backups, then it’s a beast.