Europos dirbtinio intelekto paradoksas: kodėl naujoji DI ekspertų karta jau tarp mūsų
The false choice that dominates European AI debate
We live in a time of self-proclaimed AI specialists. They fall into two camps, shouting past each other. One camp declares we cannot survive without AI. The other insists we must reject it entirely.
Between these extremes? Silence. Virtually nothing.
The first group is probably unaware of what existed before current AI. They don't know AI was already here before it was called AI. This group needs to understand that current AI is not what it should be. We moved too fast. We cut corners. The result, coming soon: unusable AI.
Why does this group shout so loudly that we can't survive without AI? They've attached an easy revenue model to it. It's not scalable. They're holding onto the trick they know.
The second group knows what came before. They understand the individual components. They've been specialists in these areas for years. This group needs help seeing the possibilities and the boundaries. They will likely be the first to recognize what AI should be, and what can be achieved the right way.
Their knowledge fits into the AI architecture. In the next generation of AI, built for humans, these people will find their preferred place. Here lies an enormous opportunity for rapid reskilling and contribution. The revenue model they currently risk losing gets new momentum.
Can we extract a challenging polarization from this, then connect it to a nuanced conclusion?
The current AI landscape in Europe
Let's be specific about where Europe stands. The numbers tell an important story.
Europe faces a documented AI talent shortage. The UK has 168,000 AI vacancies. Germany has 102,000. France has 88,000. The ratio of demand to supply for AI skills globally is 3.2 to 1. Only 10% of the world's top AI researchers live in Europe, and that percentage drops annually.
Enterprise AI adoption in Europe stands at just 13.5% as of 2024. AI talent represents 0.41% of the EU workforce. Meanwhile, the US has 240+ tech firms worth more than $10 billion. Europe has 14. US AI VC funding reached $68 billion in 2024. The EU managed $8 billion.
These statistics create panic. They drive the first group's narrative: Europe must import AI, buy AI, adopt AI quickly or fall behind forever.
But this narrative assumes current AI is the only AI. It assumes black-box deep learning is the final destination. It ignores what Europe built before the hype.
What Europe had before AI was called AI
While American companies were scaling neural networks, Europe was building something different. Something foundational. Something that current AI desperately needs.
Formal methods and verification: Europe has 25+ years of documented expertise in proving software correctness. The University of Freiburg, Inria, Formal Methods Europe, companies like Prover Technology, CEA List, Axiomise, PQShield. These institutions and companies didn't disappear. They're still here. Their expertise is more relevant than ever.
Constraint programming: The Handbook of Constraint Programming was published in 2006. The CPAIOR conference began in 2004 in Nice, France. Research groups like VeriDIS at Inria, KU Leuven, RISC at Johannes Kepler University have been developing constraint-solving techniques for decades. This isn't theoretical. It's the foundation of optimization, planning, and decision-making systems.
Logic programming and Prolog: Born in Europe. University of Aix-Marseille, 1972-1973. Creators Alain Colmerauer and Philippe Roussel. Fifty years of European research. In 2022, the community celebrated "Fifty Years of Prolog and Beyond." This isn't abandoned history. It's living expertise.
Semantic web and knowledge engineering: SWAD-Europe was a major European initiative for the Semantic Web within W3C. OWL became a W3C recommendation in 2004, led by European researchers. Today, Nature publishes research on semantic knowledge graphs. The expertise never left.
This is what the second group knows. This is their specialization. And they're watching the first group reinvent wheels poorly, ignore decades of research, and create systems that fail precisely because they lack these foundations.
The problem with the first group: unsustainable hype
The AI consulting market will reach $630.7 billion by 2032. This explains the shouting. Consultants, cloud vendors, service providers have attached their revenue model to current AI. They need to convince European companies they must adopt AI now or die.
But the model isn't scalable. Here's why.
Current AI suffers from exponential error accumulation. Systems that are 98% accurate per step become 13% accurate after 100 steps. Hallucinations snowball. Multimodal systems lose 31% performance when encountering generated errors. The mathematics doesn't care about revenue models.
The first group's solution? More training data. Better prompts. Verification steps that add more steps to the problem caused by having too many steps. They're trying to fight mathematics with natural language.
This isn't sustainable. European companies adopting these approaches will face the 95% production failure rate documented by MIT and Fortune research. They'll spend millions on systems that are mathematically guaranteed to fail in multi-step scenarios.
The first group knows this. They can't admit it. Their revenue model depends on maintaining the illusion that current AI is working. So they shout louder.
The opportunity in the second group: overlooked expertise
Now consider the second group. They watched the AI hype from the sidelines. They saw companies reinventing techniques that Europe solved decades ago. They saw neural networks struggling with problems that constraint programming handles elegantly.
They kept quiet because the market wasn't rewarding their expertise. AI funding went to deep learning. Constraint programming research continued in universities and specialized companies. Formal verification remained essential in safety-critical industries. Knowledge engineering powered semantic web applications that worked reliably.
This group exists. They're substantial. Europe has 10+ million ICT specialists. Germany has 2.3 million. Central and Eastern Europe have 3.5 million. Employment among developers is 84%. This isn't a tiny niche. This is a massive workforce with underutilized expertise.
Here's the crucial insight: the EU AI Act's requirements for explainability, transparency, and auditability create regulatory demand for precisely what this group knows.
Formal verification proves system correctness. Constraint programming provides interpretable reasoning paths. Logic programming enables traceable decision-making. Knowledge engineering builds hybrid neuro-symbolic systems that combine learning with reasoning.
The second group's expertise maps directly to next-generation AI requirements. They don't need to learn AI from scratch. They need to recognize that their existing skills ARE the next generation of AI.
The unusable AI coming soon
What happens when current AI approaches hit their mathematical limits? We're already seeing it.
Google AI Overview suggested putting glue on pizza. CNET's AI wrote articles with a 53% error rate. Medical diagnosis systems failed on 80% of pediatric cases. Legal AI models hallucinate in 1 out of 6 queries. These aren't edge cases. They're symptoms of fundamental architectural problems.
Zhang ir kt. (2023) nustatytas haliucinacijų sniego gniūžtės efektas rodo, kad LLM pernelyg įsipareigoja ankstyvoms klaidoms ir generuoja papildomus melagingus teiginius, kad jas pateisintų. Klaidos ne tik plinta. Jos auga.
98 % problema reiškia, kad po 34 samprotavimo žingsnių sistemos dažniau klysta, nei būna teisios. Po 100 žingsnių jos klysta 86,7 % atvejų. Štai kodėl 95 % AI agentų žlunga gamyboje. Štai kodėl 95 % generatyvaus AI bandomųjų projektų niekada nepasiekia gamybos.
Greitai artėjame prie netinkamo naudoti AI. Sistemų, kuriomis negalima pasitikėti atliekant daugiapakopį samprotavimą. Sistemų, kurių kiekvieną rezultatą žmogus turi tikrinti. Sistemų, kurių taisymas kainuoja daugiau, nei jos sukuria vertės.
Pirmoji grupė atsakys tuo pačiu. Didesniais modeliais. Daugiau skaičiavimų. Geresniais raginimais. Jie įstrigę paradigmoje, kuri, kaip rodo matematika, negali veikti.
Antroji grupė į tai žiūri kitaip. Jie visą karjerą kūrė sistemas, kurios NĖRA linkusios į eksponentinį klaidų kaupimąsi. Formalūs metodai įrodo savybes. Apribojimų sprendimas randa optimalius sprendimus. Loginis programavimas užtikrina patikimą samprotavimą. Žinių inžinerija kuria paaiškinamas sistemas.
Poliarizacija tampa aiški. Viena grupė dvigubina pastangas taikydama žlungantį metodą. Kita grupė turi sprendimus, kurių rinka ilgainiui pareikalaus.
Europos skaitmeninis suverenumas per turimą kompetenciją
ES Komisija skaitmeninį suverenumą apibrėžia kaip „skaitmeninę infrastruktūrą, produktus ir paslaugas, kurios saugo Europos saugumą, strateginį turtą ir interesus". Berlyno deklaracija priduria „gebėjimą veikti savarankiškai ir laisvai pasirinkti savo skaitmeninį kelią".
Dabartiniai dirbtinio intelekto suverenumo metodai orientuoti į europinių Amerikos DI versijų kūrimą. Didelių kalbos modelių mokymas. Europinių pamatinių modelių kūrimas. Konkuravimas Amerikos sąlygomis.
Tai praleidžia esmę. Europos kelias į DI suverenumą nėra daryti tą patį, ką daro Amerika, tik vietoje. Tai daryti tai, ką Europa daro geriausiai, kitaip.
ES DI akto akcentuojamas paaiškinamumas, skaidrumas ir į žmogų orientuotas dizainas nėra atsitiktinis. Jis atspindi Europos vertybes. Jis taip pat atspindi Europos techninius pranašumus. Reglamentas sukuria paklausą sistemoms, kurias galima paaiškinti, patikrinti ir audituoti.
Kas gali sukurti šias sistemas? Ne gilaus mokymosi specialistai, sukūrę „juodosios dėžės" modelius. Formaliosios metodikos ekspertai, kurie 25 metus įrodinėja sistemų teisingumą. Apribojimų programavimo tyrėjai, kurie du dešimtmečius kuria interpretuojamas optimizavimo sistemas. Žinių inžinieriai, kurie nuo 2000-ųjų pradžios kuria paaiškinamas semantines sistemas.
Tai Europos pranašumas. Europai nereikia importuoti DI kompetencijos. Europai reikia pripažinti, kad jos turima techninė kompetencija YRA naujos kartos DI kompetencija.
Persikvalifikavimo galimybė, kurią matome, bet nepastebime
Europos Komisijos įgūdžių paktas per perkvalifikavimo ir kvalifikacijos kėlimo iniciatyvas pasiekė 2,6 mln. žmonių. CEDEFOP praneša, kad daugiau nei 25 % Europos dirbančių suaugusiųjų jau eksperimentuoja su DI darbe.
Tradicinis pasakojimas teigia, kad šie darbuotojai turi išmokti gilųjį mokymąsi, neuroninius tinklus, transformerių architektūras. Jie turi tapti panašūs į pirmąją grupę.
Tai atvirkščiai. Darbuotojai, kuriuos Europa jau turi, turi būtent tuos įgūdžius, kurių reikia naujos kartos DI. Jiems tereikia suprasti, kad jų kompetencija yra vertinga DI kontekste.
Formaliosios patikros specialistui nereikia mokytis neuroninių tinklų. Jam reikia pritaikyti savo patikros kompetenciją apribojimais grindžiamoms DI sistemoms. Apribojimų programavimo tyrėjui nereikia persiorientuoti į gilųjį mokymąsi. Jam reikia suprasti, kad jo apribojimų sprendimo metodai YRA naujos kartos DI samprotavimas. Žinių inžinieriui nereikia atsisakyti semantinio žiniatinklio dėl didelių kalbos modelių. Jam reikia kurti hibridines sistemas, kurios sujungia mokymąsi su samprotavimu.
Persikvalifikavimo iššūkis nėra mokyti Europos specialistus DI nuo nulio. Tai padėti jiems suprasti, kad jų dešimtmečių kompetencija YRA DI. Tas DI, kurio dabartiniai metodai nesugeba pateikti.
Pajamų modelis, kurį jie rizikuoja prarasti, nėra dingęs. Jis transformuojasi. Įmonės, kurios supranta formaliosios patikros, apribojimų tenkinimo ir žinių inžineriją, bus tos, kurios kurs DI sistemas, kurios iš tikrųjų veikia gamyboje. DI sistemas, atitinkančias ES reglamentus. DI sistemas, kuriomis Europos įmonės gali pasitikėti.
Niuanse išvada: poliarizacija kaip galimybė
Poliarizacija tarp „mums reikia DI" ir „mums nereikia DI" yra klaidinga. Abi pozicijos praleidžia niuansą.
The nuanced reality: Europe needs AI, but not current AI. Europe needs next-generation AI built on European strengths. Explainable, verifiable, efficient, human-centered. The AI that the EU AI Act mandates. The AI that European enterprises actually need. The AI that European specialists already know how to build.
The first group's narrative that we must adopt current AI or die is dangerous. It leads European companies to invest in systems that will fail. It perpetuates dependency on American technology. It ignores European expertise.
The second group's skepticism of current AI is justified, but their conclusion that we don't need AI is wrong. We need AI. We just need the RIGHT AI. AI built on foundations they've been developing for decades.
The opportunity lies in the synthesis: recognize that the next generation of AI expertise already exists in Europe. It's the 10+ million ICT specialists. It's the formal methods experts with 25+ years of experience. It's the constraint programming researchers with two decades of work. It's the knowledge engineers who built the semantic web.
These people don't need to be replaced by AI specialists from elsewhere. They need to be empowered to recognize that their expertise IS next-generation AI expertise. They need to see the bridge between their current skills and the AI systems Europe actually needs.
At Dweve, we're building that bridge. Core provides the formal verification framework. Loom implements constraint-based domain specialists. Nexus provides the orchestration layer. Spindle provides the knowledge governance. Every component leverages European technical strengths. Every component creates opportunities for European specialists to apply their expertise to next-generation AI.
The polarization isn't a problem. It's an opportunity. The clash between hype and reality will force European companies to recognize what actually works. When current AI fails, European enterprises will seek alternatives. Those alternatives exist. They're built on 25-50 years of European research. They're held by 10+ million European specialists. They just need to be recognized as the next generation of AI.
What you need to remember
- The polarization is false. "We must have AI" versus "we don't need AI" misses the nuance. Europe needs next-generation AI built on European strengths, not imported current AI.
- Current AI is failing. Exponential error accumulation makes multi-step systems unusable. 95% of AI pilots fail to production. The mathematics doesn't care about revenue models.
- European expertise exists. 25+ years of formal methods. 20+ years of constraint programming. 50+ years of logic programming. 20+ years of knowledge engineering. This isn't history. It's the foundation of next-generation AI.
- The EU AI Act demands what Europe has. Explainability, transparency, auditability. These aren't burdens for European specialists. These are what they've been building for decades.
- 10+ million ICT specialists. Europe has substantial workforce with relevant expertise. They don't need to learn AI from scratch. They need to recognize their existing skills ARE next-generation AI.
- The reskilling opportunity is massive. Pact for Skills reached 2.6 million people. 25% of European workers experiment with AI. The challenge is recognizing existing expertise applies to AI, not importing foreign expertise.
- European AI sovereignty comes from within. Not by copying American approaches, but by leveraging European strengths in formal methods, constraint programming, and knowledge engineering to build AI that actually works.
The bottom line
The debate about European AI sovereignty focuses on the wrong question. It's not whether Europe needs AI or doesn't need AI. It's what KIND of AI Europe needs.
Current AI, built on deep learning and neural networks, is hitting mathematical limits. Exponential error accumulation. Hallucination snowballs. 95% production failure rates. This isn't the future European enterprises can rely on.
Next-generation AI, built on formal verification, constraint satisfaction, and knowledge engineering, is already emerging in Europe. It leverages 25-50 years of European research. It employs 10+ million European specialists. It aligns with EU AI Act requirements. It actually works.
The first group will continue shouting that Europe must adopt current AI or die. Their revenue model depends on it. The second group will remain skeptical, missing that their expertise IS the next generation of AI.
The nuanced truth lies between. Europe needs AI sovereignty. But the path to sovereignty isn't importing American technology. It's empowering European specialists to build the AI they already know how to create. The AI built on foundations they've been developing for decades. The AI that the EU AI Act mandates. The AI that will actually work in production.
The next generation of AI experts isn't someone Europe needs to find or train from scratch. They're already here. They've been here for 25-50 years. They just need to recognize that their expertise is exactly what next-generation AI requires.
Ready to build European AI sovereignty on European foundations? Dweve's constraint-based platform leverages decades of European research in formal methods, constraint programming, and knowledge engineering. No importing American failures. Just building on European strengths. Join our waitlist.