Blaz Fortuna

@ccccjjjjeeee Had the same realization recently. Threw together an MRP engine for manufacturing companies in 2 hours at a hackathon. When building gets this fast, it shows that code was never the real bottleneck, getting up-to-date, reliable data and understanding the problem is.

It actually worked! For the past couple of days I’ve been throwing 5.3-codex at the C codebase for SimCity (1989) to port it to TypeScript. Not reading any code, very little steering. Today I have SimCity running in the browser. I can’t believe this new world we live in.

Thanks @Rippling for hosting the game!

Enjoying #SuperBowl with @Rippling @SoPacBrewing

We’re building at a @Stanford hackathon, supported by @GoogleDeepMind, and in 3 hours we have to ship a working MRP engine— something manufacturers usually pay millions for. That wasn’t the hard part. The hard part is data that captures intent and real business logic. Most ERP systems are just expensive databases that rely on people to type reality into software. We’re here to build systems that understand how manufacturing actually works — not how it’s documented. #Stanford #DeepMind #Manufacturing #AI #Hackathon #ERP

Went go-karting with other founders. Highly recommend. @12founders

Startups run on caffeine, conviction… and perfectly timed shrimp. 😄🍤 Would ship again.

Reminds me of frustrations from 2 years ago — people tried free ChatGPT, got bad results, held it as proof the whole thing is useless. Often just running the same question through a better model produced correct result. But agree with the conclusion — AI for acceleration, humans for architecture. Reality is you can now do in hours what was a multi-week project a year ago. Stuff you either didn't attempt or needed a bigger team for. Both bigger teams and AI increase the chance of code becoming unmanageable. With bad architecture, AI is just a multiplier that gets you there faster. Blaming AI for that is just shifting responsibility.

AI-Induced Code Entropy: The Hidden Cost of Unchecked AI Acceleration In the rush to adopt AI coding tools, teams celebrate 5×–10× faster feature delivery and skyrocketing velocity metrics. Yet beneath the surface, a subtler but more dangerous phenomenon is emerging: AI-induced code entropy - the rapid, compounding disorder in codebases driven by heavy reliance on generative AI without sufficient human architectural oversight. Code entropy refers to the gradual (and often exponential) decay of structural coherence, readability, and maintainability over time. While all codebases naturally trend toward disorder without active maintenance, AI tools accelerate this process dramatically when used as a primary writer under velocity pressure. What begins as “vibe coding” bliss ends in deep spaghetti decomposition, where the AI itself can no longer reliably extend or modify the system without injecting new bugs. Key Stages of AI-Induced Code Entropy Early Phase (Velocity Mirage) AI generates boilerplate, prototypes, and isolated features at astonishing speed. Metrics soar: more PRs, faster demos, higher “AI adoption” scores. Management pushes harder; reviews and refactoring feel like unnecessary friction. Mid Phase (Subtle Accumulation) Patterns emerge: duplicated logic in subtly different shapes, one-off conditionals for edge cases, “symptom fixes” instead of root-cause changes. Inconsistencies compound across files/modules (incoherent naming, fragmented domain rules, missing invariants). Duplication explodes (reports show up to 4× increases); refactoring actually declines as teams chase features. Late Phase (Tipping Point – Deep Spaghetti) Global context fragments → small changes cascade into unrelated breakages. Bug rates spike; “almost-right” AI outputs require more debugging time than they save. The model hallucinates or proposes fixes based on corrupted patterns, accelerating decay instead of reversing it. Feature velocity flattens or reverses: adding capabilities becomes high-risk surgery. Why AI Accelerates Entropy So Effectively * Optimizes for local correctness and speed, not long-term elegance or invariants. * Lacks “human discomfort” - no instinctive aversion to redundancy or weird layering. * Produces verbose, pattern-heavy, subtly wrong code that “looks correct but isn’t reliable.” * When chained (add feature → fix bug → add another), entropy grows like a bacterial culture in a petri dish. * Under management pressure for velocity, safeguards (mandatory reviews, quality gates, debt sprints) get cut first. Real-World Signals (2025–2026 Insights) * 66% of developers spend more time fixing AI-generated code than they save writing it (Stack Overflow 2025). * 61% report AI frequently produces unreliable-looking-correct code requiring substantial review effort (Sonar 2026). * Teams observe exponential debt: what took 18 months to become unmaintainable now happens in weeks (industry anecdotes). * Once deep entropy sets in, escape routes shrink: big-bang rewrites, painful human refactoring, or stagnation. Breaking the Cycle Treat AI as an accelerator for drafts, experiments, and low-stakes work - not the primary maintainer of long-lived systems. Enforce hard boundaries: mandatory human architectural review, strict quality gates (static analysis, tests, security scans), ring-fenced refactoring capacity, and balanced metrics (velocity + churn + duplication trends + bug rates). Without deliberate discipline, the initial productivity gains are borrowed from the future - often at ruinous interest. AI-induced code entropy isn’t inevitable, but ignoring it turns today’s speed into tomorrow’s multi-year tax on engineering health and product evolution.

Spent some time working on classification into large taxonomies. Simple combination of embeddings to narrow targets + prompts for final selection works miracles compared to state-of-the-art 5-10 years ago. And super simple to incorporate feedback. #blastfromthepast

Recently re-read The Caves of Steel. Asimov imagined 8 billion people as a crisis based on realities of 1953 when population was around 2.6 billion. 70 years later we're at 8 billion and somehow humanity figured it out. Progress is easy to miss when you're living through it.

shipped a new api connector today, few hours of work. just gave claude code the docs url + pointer to a different integration from our codebase. it figured out auth, endpoints, edge cases. this used to take a week

Recorded a quick tutorial on connecting Nordoon to Outlook. If your POs come in via email (most do), this saves a lot of copy-paste. youtu.be/ndnPdE7rGlE

Asked Claude Opus 4.5 to make me a Twitter header inspired by The Great Wave of Kanagawa 🌊🤪

I'll be at AI Engineer Summit over next few days. DM me if you are around!

Happy to share the free online tools developed by our team at #Quintelligence to explore the published biomedical science and better understand the #COVID19 disease from the #research point of view. This work was developed in the @MIDASProjectEU midas.quintelligence.com/covid19-medline

Internet of Things - cutting edge technologies for extracting insight from your data! #IoT #iotplatform #DataAnalytics #IOTA #edgecomputing #5G #ArtificialIntelligence #books @johndav amazon.co.uk/Internet-Thing…

With @marko_grobelnik at #AI4M workshop in Brussels. Discussing @qlectorAI approach of using AI to help manufacturing optimize their operations.

Za vse ki greste jutri na #StartupCrawlsi smo pripravili zemljevid, da se ne izgubiš: - bit.ly/2CZtEQU In na pletki - startupjob.si/startup-crawl #Slovenija #Ljubljana #Kranj #Koper #Celje #Ptuj #Crnomelj

Naslednji petek, 11. Januarja, lahko preverite kakšen je utrip na slovenski tehnološki sceni dnevnik.si/1042855399

Najdena papiga Lj, Murgle/Trnovo. Če kdo pogresa, naj se javi? RT zaželjen #papiga #najdenazival #LJ #Murgle #Trnovo

Za naslov slovenskega startupa leta 2018 se bodo potegovali @event__registry @gotoky_gt #beeping @SmartOptometry @MaglevAudio krog.sta.si/2511553/znana-… #startupsi @StartupSlovenia