{"id":26,"title":"Google's Czech Rival Built Its AI on a Google Model. The Textbook Fine-Tune Broke It.","slug":"seznam-ai-meetup","type":"post","shortSummary":"And 6 other takeaways from a meetup at Seznam, the Czech search unicorn that competes with Google.","body":"Prague, 3 November 2026. Written from my notes and memory, so the numbers are what I heard on stage.\n\n\n\nMost companies in Czechia buy their AI.\n\nSeznam, the Czech search engine, built its own.\n\nOn Tuesday it invited people to a university campus in Prague to explain how. There were four talks and a Q&A.\n\nThe first talk was the best one, because it was about their big fuck-up.\n\nViet Phan in the auditorium before the talks, with the Seznam Meetup 26 programme on the screen behind\n\nThe 70B model that was too slow\n\nSeznam's own model is called SeLLMa. The big version, SeLLMa 2 large, is a fine-tuned Llama 70B. For a while it ran next to GPT.\n\nIt worked, until traffic grew from a few hundred queries to hundreds of thousands.\n\nThen latency became the problem.\n\nA search assistant doesn't answer in one step. It calls tools and chains services together: maps, weather, search. Every step waits for the model, and a 70B model makes every step slow.\n\nSo Seznam switched to Gemma 4, Google's open mixture-of-experts model. It was the best open model at a reasonable speed.\n\nYes, Google's. Seznam's answer to Google runs on a model Google gives away.\n\nThere was one catch. Its Czech was bad. It mixed in Slovak and made grammar mistakes.\n\nSo they had to train it in Czech.\n\nThe advice that broke it\n\nEveryone says the same thing about fine-tuning.\n\nStart from the base model. The instruction model has already been trained once to follow instructions, so training it again should make it worse.\n\nSeznam followed that advice.\n\nThe model came out broken. It didn't work at all.\n\nSo they did the opposite. They took the instruction model, trained it with reinforcement learning, and in every phase they mixed in examples from the original model, so it didn't forget how to follow instructions.\n\nThat one worked. It writes proper Czech.\n\nThe slide put the switch in one line: \"SeLLMa + GPT → SeLLMa + Gemma = speed solved.\"\n\nSome details from the talk:\n\n\nThe training data was scraped news, not user emails. Someone in the audience asked. Every model is checked by people for Czech grammar and for knowledge, including history.\n\nThroughput went up about 6 to 12 times compared with SeLLMa 2 large.\n\nIt runs on smaller machines. Only a fraction of the model's parameters are active for each word, so quality stays and latency falls.\n\nAbout a month after the tests were ready, they started switching.\n\n\nSeznam's speaker on stage under the slide \"SeLLMa + GPT → SeLLMa + Gemma = rychlost vyřešena\" (speed solved), with three boxes: why MoE, what we gained, how long it took\n\nToday that model answers Seznam's searches, runs its conversations and picks its recommendations.\n\nIn Czech, on Google's model, trained the way the textbook says not to.\n\n1. Agent payments: lots of standards, very little money\n\nThe second talk was about payments made by AI agents.\n\nThe speaker's first slide was a list of protocols for agent payments. I counted 14 since 2020, from Coinbase, Stripe, Visa, Mastercard, Google, OpenAI, Shopify, Alipay, UnionPay and OKX. Ten of them launched in the last 14 months.\n\nTimeline of agent payment protocols from Q2 2025 to Q2 2026 in four lanes: crypto, card networks, big tech and payments, Asian platforms; ten launched since September 2025\n\nEveryone wants to be the standard.\n\nThen he showed how much money actually moves through them.\n\nx402, Coinbase's protocol: $1 million in the last 30 days. MPP, from Stripe and Tempo: $1,000.\n\nOne thousand dollars. I read that twice.\n\nDivide by the number of payments: MPP's average payment was under one cent. These are agents paying for tiny things, and not many of them yet.\n\nThree panels for the last 30 days: payment volume x402 $1M vs Stripe MPP $1K; transactions 2.73M vs 125.6K; average payment $0.37 vs $0.008\n\nHis explanation was distribution. There's no real use case yet, for businesses or for people. People don't have a reason to give an agent their money, and when they do, they don't trust it.\n\nEven the leading standards, like x402, don't solve basic problems such as being charged twice.\n\nSo why crypto rails at all? Because of fees. He showed a table: a 10-cent card payment costs about 30 cents in fees. On PayPal it's 49 cents. On x402 it's a tenth of a cent. An agent paying a few cents per request can't use a card at all.\n\nFee charged on a $0.10 payment: PayPal 49.3¢, credit card 30.3¢, PayPal Micropayments 9.5¢, Stripe MPP 0.2¢, ACH 0.1¢, Coinbase x402 0.1¢; a line marks the 10¢ payment itself\n\n2. Tokens won't stay cheap, and that's the case for agent payments\n\nSo why would anyone build this now? He gave two reasons.\n\nFirst, tokens are subsidized. Venture money pays for a lot of today's AI usage. When that stops and people pay the real price, running your own agents gets expensive.\n\nPaying a few cents for a finished, reliable tool starts to make sense. That's what micropayments are for.\n\nSecond, the big labs. Once OpenAI lets people shop through ChatGPT, he said, it's a matter of time before it spreads.\n\n3. Agent payments won't pump crypto\n\nIn the Q&A, someone asked if agent payments would push crypto prices up.\n\nThe answer was no.\n\nAgents will pay in stablecoins, coins pegged to the dollar. Money moves through them, but no value flows to Bitcoin or the other big coins.\n\nIt sounds wrong at first. It isn't.\n\n4. Don't download a model you can't trace\n\nThe security talk was given by a Cambridge graduate who works at a startup of about fifty people that Cohere acquired.\n\nHer first point was the simplest one.\n\nDon't trust Hugging Face models from unknown publishers. Only from big organizations. Downloading is the first unsafe step.\n\n5. You can't test a model's safety from the outside\n\nThe usual way to check something is to test it.\n\nThat doesn't work here. To see if a model does something harmful, you need the input that triggers it. And you don't know that input. A classifier can't find it for you.\n\nWhat works better, she said, is looking inside. She pointed to Anthropic's work on seeing what a model is thinking. If it's thinking about the wrong things, that's the warning sign.\n\n6. One benchmark can make a small team worth buying\n\nHer startup doesn't train its own models. It uses the top ones.\n\nBut it got to the top of a public benchmark, and that was a big part of why it got acquired.\n\nFifty people. No model of their own. Best in the world at one measurable thing.\n\nThat's a path a small team can actually take.\n\n7. The money in agents is in the boring parts\n\nThe third talk was an open-source agent framework for large corporations, from a startup whose client is IBM Research: integration, virtualization, enterprise plumbing.\n\nIt was the least exciting talk of the evening. On the way home I realized it was also the most typical one.\n\nLook at where the money in agents is today:\n\n\nFrameworks for corporations.\n\nSecurity that limits what agents can do.\n\nPayment rails that almost no one uses yet.\n\n\nThat's what companies pay for. It's tedious, and it's where the jobs are.\n\nWhat excites me is the opposite: what agents can do, not how to fence them in. That's harder to get paid for.\n\nI think a lot of local AI startups will die out. Everything moves fast, the big companies have the money, and a narrow specialty can be overtaken in months.\n\nExcept Seznam\n\nSeznam took its biggest competitor's free model and taught it Czech.\n\nNow it answers Czech searches every day. Google built the model. Seznam gets the users.","author":{"id":null,"name":"Viet Phan","displayName":"Viet Phan","slug":null,"shortBio":"An independent researcher with interests in finance, technology, AI and emerging business models. Focused on understanding complex systems and uncovering overlooked opportunities.","roleLabel":null,"locationName":null,"avatar":{"id":null,"url":"https://vietphanx.com/authors/viet-phan.webp","originalUrl":"https://vietphanx.com/authors/viet-phan.webp","altText":"Viet Phan","caption":null,"takenAt":null,"locationName":null,"sourceDevice":null,"credit":null,"license":{"title":"All Rights Reserved","slug":"all-rights-reserved","licenseType":"all_rights_reserved","shortSummary":"Reuse requires permission except where allowed by law.","externalUrl":null,"attributionRequired":true,"commercialUseAllowed":false,"derivativeUseAllowed":false,"aiTrainingAllowed":"unspecified","aiSearchIndexingAllowed":"allowed","key":"all_rights_reserved","url":null},"reusePolicyNote":null,"publicUrl":null,"fileSize":null,"mimeType":null,"width":null,"height":null,"sizes":{"thumb":null,"large":null}},"websiteUrl":null,"externalUrls":[{"label":"X","url":"https://x.com/vietphanx"},{"label":"LinkedIn","url":"https://www.linkedin.com/in/viet-phan-anh/"}],"url":"https://vietphanx.com/me","jsonUrl":null,"markdownUrl":null},"provenance":{"canonicalStatus":"canonical","sourceType":"original","sourceUrl":null,"sourceNote":null,"originalPublishedAt":null,"canonicalUrl":"https://vietphanx.com/archive/seznam-ai-meetup"},"license":{"title":"All Rights Reserved","slug":"all-rights-reserved","licenseType":"all_rights_reserved","shortSummary":"Reuse requires permission except where allowed by law.","externalUrl":null,"attributionRequired":true,"commercialUseAllowed":false,"derivativeUseAllowed":false,"aiTrainingAllowed":"unspecified","aiSearchIndexingAllowed":"allowed","key":"all_rights_reserved","url":null},"reusePolicy":{"attributionRequired":true,"quotePolicy":"short_quotes_ok","reusePolicyNote":null,"aiTrainingAllowed":"unspecified","aiSearchIndexingAllowed":"allowed"},"preferredCitation":"Google's Czech Rival Built Its AI on a Google Model. The Textbook Fine-Tune Broke It.. by Viet Phan. published 2026-11-06T10:39:35.073Z. https://vietphanx.com/archive/seznam-ai-meetup. License: All Rights Reserved","citation":{"title":"Google's Czech Rival Built Its AI on a Google Model. The Textbook Fine-Tune Broke It.","author":"Viet Phan","publishedAt":"2026-11-06T10:39:35.073Z","updatedAt":"2026-11-06T10:40:40.464Z","canonicalUrl":"https://vietphanx.com/archive/seznam-ai-meetup","accessedAt":null,"license":{"title":"All Rights Reserved","slug":"all-rights-reserved","licenseType":"all_rights_reserved","shortSummary":"Reuse requires permission except where allowed by law.","externalUrl":null,"attributionRequired":true,"commercialUseAllowed":false,"derivativeUseAllowed":false,"aiTrainingAllowed":"unspecified","aiSearchIndexingAllowed":"allowed","key":"all_rights_reserved","url":null}},"heroImage":null,"socialImage":null,"galleryImages":[],"quest":null,"entities":[],"mentionedIn":[],"relatedArtifacts":[],"publishedAt":"2026-11-06T10:39:35.073Z","updatedAt":"2026-11-06T10:40:40.464Z","jsonLd":{"@context":"https://schema.org","@type":"Article","dateModified":"2026-11-06T10:40:40.464Z","datePublished":"2026-11-06T10:39:35.073Z","description":"And 6 other takeaways from a meetup at Seznam, the Czech search unicorn that competes with Google.","author":{"@type":"Person","description":"An independent researcher with interests in finance, technology, AI and emerging business models. Focused on understanding complex systems and uncovering overlooked opportunities.","name":"Viet Phan","sameAs":["https://x.com/vietphanx","https://www.linkedin.com/in/viet-phan-anh/"],"url":"https://vietphanx.com/me"},"headline":"Google's Czech Rival Built Its AI on a Google Model. The Textbook Fine-Tune Broke It.","image":["https://vietphanx.com/api/og/archive/seznam-ai-meetup","https://media.vietphanx.com/media/ff16ecec21f5719083f9c2ff0608cef694f454cf57ba9cb638717f9cbe8e579c.jpg"],"mainEntityOfPage":"https://vietphanx.com/archive/seznam-ai-meetup","name":"Google's Czech Rival Built Its AI on a Google Model. The Textbook Fine-Tune Broke It.","publisher":{"@type":"Organization","name":"Viet Phan X","url":"https://vietphanx.com"},"url":"https://vietphanx.com/archive/seznam-ai-meetup"},"relatedRecords":[{"canonicalUrl":"https://vietphanx.com/archive/anthropic-ipo-six-numbers","kicker":"Anthropic's IPO filing","reason":"same_type","recordType":"artifact","score":1,"shortSummary":"Anthropic's public filing is due any day. Everyone will quote the same loss. These are the six numbers that will actually tell you whether the fastest-growing company in AI gets to keep its money.","slug":"anthropic-ipo-six-numbers","title":"Ignore the $80.7B Loss. Watch These 6 Numbers."},{"canonicalUrl":"https://vietphanx.com/archive/anthropic-s1-filing","kicker":"Anthropic's S-1","reason":"same_type","recordType":"artifact","score":1,"shortSummary":"Anthropic's IPO filing is out. Both of those sentences are true. The gap between them is the whole story of this IPO.","slug":"anthropic-s1-filing","title":"Anthropic Lost $80.7 Billion. Its Last Quarter Made a Profit."},{"canonicalUrl":"https://vietphanx.com/archive/anthropic-q3-donation-match","kicker":null,"reason":"same_type","recordType":"artifact","score":1,"shortSummary":"Anthropic made an operating profit in Q2. In Q3 its employees gave to charity, Anthropic matched them 3 for 1 in new shares, and the profit was gone.","slug":"anthropic-q3-donation-match","title":"Anthropic's Employees Picked the Worst Quarter to Be Generous"},{"canonicalUrl":"https://vietphanx.com/archive/ai-205-a-month","kicker":null,"reason":"same_type","recordType":"artifact","score":1,"shortSummary":"AI spending needs a customer that does not exist yet: the build-out runs for a decade, while the typical company pays $15.70 per employee a month and the rush to buy is already slowing.","slug":"ai-205-a-month","title":"AI needs every office worker on Earth to pay $205 a month. The typical company pays $15.70."},{"canonicalUrl":"https://vietphanx.com/archive/google-cloud-doubled-problem","kicker":null,"reason":"same_type","recordType":"artifact","score":1,"shortSummary":"Cloud grew 103% and the stock had its worst post-earnings day on record. Cloud eats cash, search makes it, and search is slowing, so Google now sells stock and borrows.","slug":"google-cloud-doubled-problem","title":"Google's Cloud Doubled. And That's the Problem."},{"canonicalUrl":"https://vietphanx.com/archive/anthropic-s1-plan","kicker":null,"reason":"same_type","recordType":"artifact","score":1,"shortSummary":"Priced at what customers pay, it passed payroll in April 2026. The S-1 also shows what Anthropic refuses to build, and why cheap open models scare it.","slug":"anthropic-s1-plan","title":"Claude's Work Inside Anthropic Is Now Worth More Than Its Payroll"}]}