Bielik Co-Creator: Poland Could Become the EU's Second AI Power — If It Stops Losing Talent

A Polish AI Ambition Tested Against Reality

Remigiusz Kinas, co-creator of the Polish large language model Bielik, believes Poland has the raw ingredients to become the European Union's second AI power behind France — but only if it breaks a cycle of underinvestment and talent loss. In an interview with Forbes Poland, Kinas said that with serious funding, computing capacity and access to data, a Polish team could build a model on par with France's Mistral within six months and a model competitive with Chinese systems within a year.

Kinas's confidence is tempered by what he saw at this year's ICLR conference in Rio de Janeiro, where Polish research was barely visible. He argues that Poland's reputation for elite computer scientists is partly mythology: the country produces few breakthrough papers or research centres, and its most visible AI successes — ElevenLabs, DeepL and Pathway — were built by Poles abroad. Warsaw's only significant foreign AI outpost is a small Mistral engineering office.

Behind the public debate, Kinas describes an AI value chain with several layers. Polish companies are well represented among integrators who plug foreign models into products, and among developers of their own models, but least present in fundamental research — the layer that invents new architectures. He fears the gap with the most advanced countries will widen unless governments and companies treat AI as infrastructure rather than a talking point.

Kinas also discussed the practical bottlenecks at Bielik and in biology-focused AI. High-quality Polish-language data and access to compute, particularly AGH Cyfronet's Helios supercomputer, constrained Bielik. At Ingenix, the drug-discovery company he joined, noisy and heterogeneous biological data makes AI far harder than in text. He argues that large language models are unlikely to lead to artificial general intelligence, and that the next breakthroughs will come from multimodal world models that represent physical reality.

Where Poland's AI Ecosystem Actually Falls Short

Poland's AI Brain Drain Is Structural

Kinas points to a one-way migration: Poles helped build ChatGPT through Wojciech Zaremba and work at OpenAI, Anthropic and DeepMind, while foreign companies keep only small European outposts in Poland. This is a fact pattern, not a judgment. The interpretation Kinas offers is that until Poland hosts significant research centres, its global visibility will keep shrinking — and the talent that leaves will have fewer reasons to return.

Funding and Data: What Bielik Actually Needed

The Bielik model would not exist without AGH Cyfronet's Helios supercomputer, Kinas says. His more striking claim is that access to books from Polish publishers would let teams train high-quality Polish models without breaking copyright, because models distil knowledge rather than memorise text. That claim is an assessment, not a verified fact, but it identifies a concrete bottleneck: compute is partly solved, while curated Polish-language data remains scarce.

Fundamental Research Is the Missing Layer

At ICLR, Kinas saw almost no Polish university presence. He places Polish AI activity in the integration and model-building layers, but says the deepest layer — inventing new architectures — is the thinnest. That explains why Poland's AI successes tend to be products rather than research breakthroughs, and why the country's position in the global AI hierarchy may not improve just by spending more on compute.

Beyond LLMs: The Niches Kinas Thinks Are Still Open

Kinas argues that language models are a dead end for AGI because the world cannot be described by language alone. He points to physical AI and drug discovery as areas where adoption is low and barriers are high, citing Ingenix's multimodal approach, AlphaFold's Nobel recognition, and Peter Steinberger's OpenClaw being acquired by OpenAI weeks after launch. The implication is that Polish teams should compete where integration is hard and responsibility is high, rather than trying to beat US and Chinese labs at scale.

What Polish AI Leaders Can Do With the Advantages They Have

The audience that can act on Kinas's diagnosis is narrow: Polish AI founders, research leaders and the public institutions that fund compute and data. The following steps are tied directly to the bottlenecks he names.

• Negotiate access to Polish-language book and text corpora with publishers. Kinas says this single step would unlock Bielik-grade training data without copyright disputes, because models distil rather than reproduce text.
• Build public compute capacity beyond AGH Cyfronet's Helios. Kinas credits the supercomputer as the reason Bielik exists, and identifies compute as the second binding constraint after data.
• Fund model-building and fundamental-research teams, not only integration projects. Kinas says the shallowest layer of Polish AI is the people who invent architectures and improve models.
• Target high-barrier niches such as physical AI and drug discovery, where adoption is low and mistakes carry serious consequences. Ingenix's multimodal drug-discovery work and OpenClaw's rapid acquisition both point to room for smaller teams.