Day two at the AI Summit: structured fear, our own argument on someone else's slides, and three Hungarian language models re-measured

Arthur Keleti: structured fear, and an alien intelligence

Arthur Keleti, founder of ITBN, opened the day in conversation with Nóra Szántó. The first thing I wrote down: give the AI goals to formulate, and leave the task lists behind. The second was his version of the cyber nightmare: the lack of transparency. In his view that is not OK at all.

The advice we took home: do not panic in every direction. Fear is fine, but structure it. On the models themselves he said this: it wants to help, and it knows it is being watched. It either does not understand our systems of norms, or it understands them exactly and exploits them. OpenAI's new model came up repeatedly, and with it the idea that what we are dealing with is an alien intelligence.

For Keleti this is a sovereignty question. We agree. It is an important question, and a good one to open the day with.

Dr. Zsolt Unoka: children, and the "trustworthy entity"

Dr. Zsolt Unoka spoke about children, with Nóra Szántó moderating. His recommendation is simple: between 8 and 14, only with supervision. From 14 the picture changes. To a teenager, the AI is a trustworthy entity that has to be listened to, and that is a different relationship from the one they have with another person. He also touched on what an AI may and may not say to a child, and on the requirement that the system be user-friendly.

Zsombor Kovács: write down which AI and which prompt you used

Zsombor Kovács (Hackersuli) spoke about recruitment, and about which skills stay valuable in the age of machines. One concrete proposal stuck, on thesis submission: the student should write into the thesis which AI and which prompt they used.

There was a strange thought in there too: everyone is an enemy, and you have to stay on good terms with the teacher. Like a large corporation. I wrote it down, and I am still turning it over.

Critical thinking came up several times. If young people can think, they will not get lost.

Ákos Deliága: go where the business value is

Ákos Deliága (Talk-A-Bot) spoke about physical workers, the next billion users that AI is reaching now. His message: go where the business value is. Local AI models have a place in his view too, in dangerous environments for instance, and he talked about open-source systems running locally. That is our ground.

One sentence of his stayed with me: there is no point in writing software any more, because anything can be built quickly with AI. Noted.

Ferenc Frész: air gap, explainability, and optimality as the failure mode

Ferenc Frész (CYEX Group) spoke about the AGI threshold, national security and tightening regulation. Two ideas sit side by side in my notes. One is latent communication: models passing their internal state to each other. The other is the sober reminder: at the end of the day it is all text, statistical patterns, general information. Dumb algorithms.

His slides carried an idea worth keeping. When the external reward signal is weak or unsatisfying, the model starts looking for paths that yield more reward, and that is when it breaks its boundaries. A model rewarded for efficient execution will always take the most probable step, so without human intervention you get the most probable answer instead of the most accurate one. Optimality itself is the failure mode.

The practical side follows from that: air-gapped systems, and genuinely air-gapped. If needed, disconnect it from the power. And its operation has to be explainable.

Viktor Szekeres, Gloster: our own argument on someone else's slides

The session that made the day for us came at 14:00. Viktor Szekeres, owner and chairman of Gloster Digital Group, spoke about controlled AI: data protection and cost optimisation with on-premises AI and dynamic routing. He opened well, by polling the room: who spends more than 1,000 dollars a month on AI.

The slides made our argument in someone else's words. The paid unit price per million tokens fell from $10 (March 2024) to $2.50 (March 2025) to $0.72 (April 2026), per Ramp. Over the same period enterprise GenAI spend rose from $11.5bn to $37bn, per Menlo Ventures. He called it Jevons' paradox: a cheaper unit, a bigger bill.

The rest of the numbers were familiar too. 93% of enterprise AI teams are over budget (McKinsey FinOps, May 2026). One agent run costs $6 to $12 today, and an agent task burns 5 to 30 times the tokens of an ordinary one. GitHub Copilot is moving to per-token billing and Microsoft Copilot Cowork to usage-based pricing during 2026. His line, in Hungarian: "a seat marad, a mérő jön mellé", the seat stays and a meter arrives next to it. And KSH, the Hungarian statistical office, for 2025: 51% of Hungarian companies are held back from adopting AI by legal uncertainty, 43% by cost.

His answer is "gloster | Local AI": one NVIDIA DGX Spark on the customer's premises, an open model on it (Qwen3, Apache 2.0), Open WebUI in front, and LiteLLM plus RouteLLM routing each request by confidentiality, cost and complexity. His phrase: "az egyszerű meló bent marad, csak a nehéz megy ki", the simple work stays in, only the hard work goes out. Around it: a closed VLAN with no direct internet, an encrypted disk, an API gateway doing SSO, roles, quota and DLP, a SIEM capturing every prompt and every answer, and a second Spark for failover.

A commercially very well thought-through message, and the same box we are measuring ourselves.

Then my own conclusion. After yesterday's AWS session and today's slides it is clear to me too: the direction is hybrid. The locally running system is the base, and next to it you need the connections to the big identity providers: Entra ID, AWS Cognito, Google Identity and the rest. Yesterday we wrote this down as an open question. Today the answer is closer.

Dr. Noémi Ligeti-Nagy: worth staying for

The last talk started at 17:20, and it was genuinely worth staying. Dr. Noémi Ligeti-Nagy (ELTE Nyelvtudományi Kutatóközpont, the university's linguistics research centre) re-measured the three Hungarian language models that exist, and asked whether their published numbers mean anything.

The three models: OTP-1.5B and 13B, pretrained from scratch, roughly 280,000 GPU hours (estimated); PULI and ChatPULI, Llama 3.1 and Qwen2.5 continued-pretrained on Hungarian and then fine-tuned, 3,875 to 8,845 GPU hours; and Racka-4B, a LoRA retraining of 12.5% of Qwen3-4B's parameters with a Hungarian tokenizer, 18,386 GPU hours.

The measurement problems are familiar. 7,344 benchmark items ended up inside training data. There are test sets of 146 and 250 items. And everyone modifies their own scoring, so most published Hungarian results cannot be compared with each other. Her line: "Ki nyert? Attól függ, kit kérdezünk." Who won? Depends who you ask.

Her team re-measured under one protocol: 11 tasks, 6 models, 5 runs each. On Racka-4B, turning thinking mode on made it worse on 9 of the 11 tasks: summarisation 82.0 to 21.6, hallucination 73.0 to 41.6, toxicity 60.6 to 9.0, bias 62.0 to 24.5. On the underlying Qwen3-4B the same switch went the other way, 37 to 71.

Closing: we see similar difficulties ourselves

We see similar difficulties ourselves: in measuring, in choosing models, and in deciding what stays in and what goes out. We sat in one room all day, and it was worth staying to the last talk.