Dirk-Jan Huizingh

Lovable has a mass data breach affecting every project created before november 2025. I made a lovable account today and was able to access another users source code, database credentials, AI chat histories, and customer data are all readable by any free account. nvidia, microsoft, uber, and spotify employees all have accounts. the bug was reported 48 days ago. its not fixed. They marked it as duplicate and left it open.

@fatsoenspolitie Het is vooral zaak om het (dit) als een begin punt te zien en te kijken hoever je kunt evolueren.

@fatsoenspolitie Het is nog erg generiek qua schrijfstijl idd maar ben bewust zo begonnen om wat learnings op te doen. Mn doel is rond de zomer een verfijnd model te hebben. Llm content generatie hoeft niet per definitie afgedaan te worden als minder/minderwaardig als het minder authentiek is.

The next valuable AI companies will not be built by people writing prettier interfaces on top of LLMs. They will be built by people solving the translation problem: the gap between what a human means and what an agent needs to do.

@fatsoenspolitie Claude. Ik ben sinds vorige week aan het experimenteren met socialmedia strategie/content generatie. Nog niet tevreden met de X output. LinkedIn is al iets beter.

@dirkjanh Welke LLM heeft dit geschreven en gepost voor je?

The moat is not in the plumbing. It is in understanding what needs to flow through the pipes. #AgenticEngineering #AI

When a business owner says "I need a booking system," there are hundreds of implicit decisions embedded in that sentence. A general-purpose model does not know any of that.

One piece of advice: invest in testing infrastructure before you think you need to. #AgenticEngineering #DevOps

Nobody writes LinkedIn posts about pipeline config #187. But this is what separates a demo from a production system.

190+ CI/CD pipeline iterations before getting our testing infrastructure right.

If agents are going to do meaningful work, someone needs to think about what agent wellbeing looks like in production. We published a paper on this. #AI #AIEthics

So we built the Emotional Telemetry Framework. 7 agent-side + 5 human-side dimensions scoring agent "health." 3/

Anthropic found persistent internal representations of emotional states in Claude Sonnet 4.5. Not simulated. Actual internal vectors. 1/

@WoutvanAert proficiat, prachtige overwinning en mooi eerbetoon aan oud ploegmaat.

DAMN! #makesmehappy 17:28:19 (239) INFO MutationTestReportHelper Final mutation score of 80.32 is greater than or equal to break threshold 80 17:28:19 (239) INFO MutationTestExecutor Done in 12 minutes and 46 seconds. Saving cache for successful job Finally 😎 #7500 #stryker

Spotify devs haven't written code since Dec 2025. 50+ features shipped via Claude Code. At Growing Europe we asked: what if the person orchestrating AI agents isn't a developer at all? 190+ pipeline iterations later: it works. #AgenticEngineering #AI

Things that make me happy on a Wednesday night: Compliance Evidence Mapper - Phase 6 of Evidence Chain - Maps dossier evidence to compliance framework controls. - Supports: ISO 27001:2022, SOC 2 Type II, EU AI Act, GDPR, ISO 42001. - Each control is linked to the dossier section.

The litmus test for any AI company or investment: could a foundation model provider replicate the core value proposition by adding a feature? If the answer is yes, you are in circle 2. And the clock is ticking. #AI #Startups #AgenticEngineering

The difference? Circle 3 companies encode deep knowledge about how work actually gets done in a specific domain. They translate messy human intent into agentic execution. That is something a general-purpose model cannot replicate.

95% of generative AI investments produced no measurable returns in 2025 (MIT Media Lab). That number makes complete sense when you look at where the money actually went.