Sisäpiiri 163 · Tools · first published 2026-10-10 · updated 2026-10-12

Europe's dismal AI development | Aleksi Paavola | Negotiator Inner Circle 163

In Neuvottelija Inner Circle 163, Sami Miettinen and AI builder Aleksi Paavola ask whether Europe has already lost the foundation-model race, and what happens if the US and China cut Europe off from the best models. This long-form, sourced article checks the episode's claims and gives the background: Europe's share of compute and models (Stanford AI Index, Bruegel, Draghi), where Mistral and Poro stand, LUMI and the EU's AI programmes, the contractual ban on distillation, the harness and reinforcement learning, the safety classifiers and export controls around Anthropic's Fable model, how far open models lag and how cheaply their safety training can be removed, research on repeated sampling, and how fast Colossus was built compared with European projects.

Sami Miettinen · Sections: Tools and Implementations + AI and Society

Europe’s dismal AI development | Aleksi Paavola | Negotiator Inner Circle 163

Summary: In Inner Circle episode 163 Sami Miettinen and AI builder Aleksi Paavola continue the main episode’s discussion of data sovereignty. The question is whether Europe has already lost the foundation-model race, and what is left for it if the US and China close off access to the best models. Neither believes the race can be won through regulation or corporate subsidies. The conversation arrives at the episode’s weightiest observation: as safety filters on frontier models tighten, security work may move to unfiltered open models.

What the check found: The episode’s broad lines hold. According to Bruegel, Europe has about 5% of the world’s AI compute. According to Stanford’s AI Index, Europe produced three notable models in 2024, against 40 from the US and 15 from China. Paavola’s description of Fable’s safety filter matches Anthropic’s own account of Fable’s July 2026 redeployment, and Anthropic confirms that Amazon researchers found a way past its safeguards. Some details need correcting:

• The lag of open models is now about four months on average according to Epoch AI, not 6–12 months.
• PewDiePie fine-tuned an existing Chinese model rather than training his own, and no comparison with Sonnet 4.5 was found.
• The Altman quote could not be sourced.

How to read this

• This is the Inner Circle follow-on to a main episode, which the conversation refers to several times. Paavola was previously a guest on Negotiator 396, where he presented the Laki.ai MCP.
• Who is speaking. Paavola is an AI builder and technology entrepreneur with more than ten years’ experience who has trained language models for a range of organisations. He builds tools on top of US frontier models, so he has a practical interest in access to them being preserved. Miettinen says he uses models from Anthropic, OpenAI and Google, with DeepSeek V4 Pro as a backup.
• When the episode was recorded. The episode was published on 10 October 2026, but in the conversation Paavola says “Fable comes back into use today”. Anthropic redeployed Fable 5 with new safety classifiers on 1–2 July 2026, so the episode was most likely recorded in early July. Events from the summer and autumn are covered separately in this article and marked as later than the recording.
• Source and expansion. The account of the episode is based on a Finnish MacWhisper transcript and the publisher’s own chapter list. Quotations are translated from Finnish. The transcript has no speaker labels, so some remarks are described as positions in the conversation rather than attributed to one speaker. For this expanded version (12 October 2026), the episode’s claims were checked against primary sources. All sources are listed at the end, and claims that could not be checked are listed separately.

1. Has Europe already lost?

Miettinen says he tried Mistral about a year ago and found it weak. Paavola agrees: he tests every new Mistral release and has been “pretty disappointed”. Both praise Mistral’s marketing as superb.

Paavola still finds a reason for Mistral to exist. In his horror scenario, the US blocks Europe’s access to the best models and China responds in kind. Europe would then at least have “something”.

What the metrics say

Every metric found supports the episode’s premise of dismal European development:

• Models. According to Stanford’s AI Index 2025, “in 2024, U.S.-based institutions produced 40 notable AI models, significantly outpacing China’s 15 and Europe’s three”. The 2026 report says US and Chinese models have “traded the lead multiple times since early 2025”. Europe is not part of that race.
• Compute. Bruegel’s September 2026 policy brief finds that “only two gigawatts of artificial-intelligence compute capacity, equating to five percent of the global total, is located in the European Union”. The US has 35 GW (78%) and China 5 GW (11%). The EU share is forecast to rise only to 5.6% by 2031. Bruegel adds that the EU “has just one significant model developer: Mistral”.
• The Draghi report (2024). The report finds that “around 70% of foundational AI models have been developed in the US since 2017”. Only four of the world’s top 50 tech companies are European, and the largest European cloud operator holds 2% of the EU market.

Mistral: a disappointment, or Europe’s only contender?

Paavola’s assessment of Mistral’s models comes from his own experience and cannot be checked. Mistral’s position has strengthened quickly, however, partly after the recording:

• Funding. In September 2025 Mistral raised €1.7 billion at an €11.7 billion valuation, in a round led by ASML. In September 2026 it raised €3 billion at a valuation of “more than €21 billion”, in a round led by Samsung.
• Models. Mistral Large 3 (December 2025) is open under Apache 2.0, with 675 billion parameters, 41 billion of them active. The new flagship, Mistral Large 4, was announced on 6 October 2026, after the recording. It has one trillion parameters, 52 billion active, and was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centres. Mistral has promised to release the weights by the end of October.
• Compute. Mistral’s own target is about a gigawatt of compute capacity by 2030.

As insurance against Paavola’s horror scenario, Mistral is more credible now than at the time of recording. The checked data do not show that it has reached frontier level.

2. Poro, Silo AI and LUMI

The Finnish Poro model gets ironic treatment: after its release, “nobody has heard anything about it”. The founders of Silo AI are congratulated on the deal in which, according to the episode, AMD paid 650 million.

Corrections and context:

• The deal. On 10 July 2024 AMD announced an “all-cash transaction valued at approximately $665 million” to acquire Silo AI. The deal closed on 12 August 2024.
• Poro and Viking. Poro 34B was trained from scratch on a trillion tokens on the LUMI supercomputer, together with the University of Turku’s TurkuNLP group and the HPLT project. The Viking model family for the Nordic languages was trained the same way.
• Poro has not disappeared, but it has changed. In June 2025 AMD Silo AI and TurkuNLP released Poro 2 (8 and 70 billion parameters). These were not trained from scratch: they were continued-pretrained from Meta’s Llama 3.1 base models. In August 2026 a long-context Poro 2 8B Long followed, along with a preview of a mathematical-reasoning model. Paavola’s irony therefore lands somewhere other than he intends. The Finnish language model is alive, but it now builds on an American company’s base model, which would be no help in the horror scenario.
• LUMI. LUMI’s GPU partition has nearly 12,000 AMD MI250X GPUs and is operated by CSC in Kajaani. In December 2024 EuroHPC selected the Finnish-led LUMI AI Factory among Europe’s first seven AI factories. The procurement contract for its successor, LUMI-AI, was signed on 31 August 2026. Bull will deliver a €387.8 million system with AMD MI430X GPUs, due in the second half of 2027, with ten times the AI capacity of the current LUMI.

3. Distillation and local models

Miettinen asks whether Europe could act “the Chinese way” and query a US frontier model hard enough to distil its capabilities into its own model. Paavola considers this culturally difficult for Europe because of the GDPR and the EU AI Act. In the horror scenario, though, that is probably what would have to happen.

What is known about distillation:

• Contract terms forbid it. Anthropic’s commercial terms bar using the service “to build a competing product or service, including to train competing AI models”. OpenAI’s terms of use bar using output “to develop models that compete with OpenAI”. The barrier is a contractual one, not the GDPR or the AI Act.
• It happens anyway, on an industrial scale. In February 2026 Anthropic reported campaigns by three Chinese labs: more than 16 million exchanges through about 24,000 fraudulent accounts. DeepSeek accounted for more than 150,000 of them, Moonshot AI for more than 3.4 million and MiniMax for more than 13 million. Anthropic also notes that distillation as such is “a widely used and legitimate training method”. The problem is unauthorised use of a competitor’s model.
• Frontier models defend themselves. According to Anthropic, Fable 5 hands requests flagged as distillation attempts to the weaker Opus 4.8 model. Distillation in the horror scenario would therefore breach the contract and also be blocked technically.

Locally runnable models exist, for example Google’s Gemma models and Chinese open models. In April 2026 Google released the Gemma 4 family under Apache 2.0. Google says its smallest models “run completely offline” on the device, and the larger ones run on an ordinary computer.

4. The harness and reinforcement learning

The core of the conversation is that a model’s weights alone no longer decide the race. Frontier models’ advantage lies in reasoning and in the harness: how the model works with the tooling environment that directs it. Paavola says the companies behind Codex, Claude Code, Cursor and Google are now doing reinforcement learning specifically around the harness, so the model learns to perform well in exactly that environment. The Austrian Peter Steinberger is mentioned as someone who took harness development to a high level.

What has been said publicly:

• OpenAI. In May 2025 OpenAI said codex-1, the model behind Codex, was “trained using reinforcement learning on real-world coding tasks in a variety of environments”.
• Anthropic. A November 2025 Anthropic research post says Claude is trained with reinforcement learning “on real programming tasks”. No direct statement was found that this training takes place in the Claude Code environment specifically.
• Steinberger. In February 2026 Steinberger announced he was joining OpenAI “to work on bringing agents to everyone”. His open agent project, OpenClaw, is moving to a foundation.
• Cursor. The conversation says Cursor is in the game “through the SpaceX merger” and that “Musk got Cursor’s user data too”. This holds:
  • According to SpaceX’s registration filing with the US Securities and Exchange Commission, Cursor and xAI agreed in April 2026 that Cursor would receive compute capacity and, in return, contribute data, prompts and code, among other things, to improving the Grok model.
  • On 16 June 2026 SpaceX agreed to buy Cursor for $60 billion in stock, and the deal closed in August.

Paavola concludes that Europe lacks proper open-source equivalents. Smaller efforts are mentioned in the conversation, however:

• Lovable. According to Miettinen, Anton Osika is developing a model of Lovable’s own. No source was found for this. Available reporting says Lovable builds on OpenAI and Anthropic models.
• PewDiePie. According to the conversation, the Swedish YouTuber PewDiePie trained a model on his own hardware that reached Sonnet 4.5 level. That is not accurate as stated. In a video published in February 2026, PewDiePie fine-tuned an existing Chinese model, Qwen2.5-32B. It scored 39% on the Aider Polyglot coding benchmark, against GPT-4o’s 23.1%. He acknowledged himself that benchmark data had leaked into his training data. No comparison with Sonnet 4.5 was found.

The conversation concludes that development is not impossible even “with garage setups”. Both examples, though, show the same thing as Poro 2: even the small players build on someone else’s base model.

5. Regulation will not win the race

Both speakers hold that Europe cannot win the race by regulating. The conversation notes that training data can be had from US models for as long as their providers do not decide to stop the service working in Europe. As a foretaste, it mentions Google’s new agent product at around $100 a month, which was not brought to Europe.

Checked: In May 2026 Google announced a $100-a-month AI Ultra plan. Its round-the-clock Gemini Spark agent is available “U.S. only”. The claim holds, although Google has not said whether the restriction is due to EU regulation.

The EU AI Act timetable:

• General-purpose models. Obligations have applied since 2 August 2025. In July 2025 the Commission published a voluntary code of practice, signed by Anthropic, Google, Microsoft, Mistral AI and OpenAI, among others. xAI signed only the safety and security chapter.
• The Act as a whole became applicable on 2 August 2026, with exceptions.
• High-risk systems. The so-called digital omnibus package, in force since July 2026, postponed their obligations: to 2 December 2027 for Annex III systems and to 2 August 2028 for systems built into regulated products. The EU has thus already lightened its own regulation for competitiveness reasons.

The Draghi report noted in 2024 that the AI Act imposes extra requirements on models above a compute threshold, “a threshold which some state-of-the-art models already exceed”. The threshold therefore applies to exactly the kind of model Europe would be trying to build.

6. The safety filter that weakens a frontier model

Paavola raises the episode’s most topical issue. He says Anthropic’s Fable is coming back into use on the day of recording with a new safety filter that is more aggressive than the old one. The old filter routed the user to Opus 4.8 whenever it detected a possible risk. If the new filter is very aggressive, frontier models lose their competitive edge, and users have no access to an unfiltered frontier model. An open model could then become the leading option.

Later Paavola narrows this down to security. In his view Fable’s and Mythos’s security capabilities are remarkable, but in the Fable available to users they have been “completely stripped out”. Security questions get no answer. “How can you protect your own systems? You can’t.” His conclusion is that maintaining security may soon require running your own open models without a filter layer.

The conversation also says that “AWS did it on the first attempt with Fable”: it cracked the harness open, and the US government had to consider “whether to give this to everyone or try a third time”.

What Anthropic has said itself

Paavola’s description broadly matches the timeline Anthropic has published:

What follows for the episode’s claims:

• Paavola’s description holds for the time of recording. The new classifier was more aggressive than before, and it routed requests to Opus 4.8. The Fable 5.1 system card says the classifiers “consistently fire across all tested cyber capability evaluations”, so Fable 5.1’s performance on cyber tasks is practically identical to Opus 4.8’s. In a prompt-injection evaluation, roughly half of Fable 5.1’s coding runs fell back to Opus 4.8.
• “Completely stripped out” was an overstatement already in July, and the situation changed in September. Security questions are answered by Opus 4.8, which is not an incapable model. From the 5.1 version in September, the generally available model also allows vulnerability discovery in source code. Paavola’s concern about protecting one’s own systems has therefore been partly resolved for defensive use. Offensive techniques, such as analysing compiled binaries and writing exploit code, remain blocked.
• The AWS claim. Anthropic confirms that Amazon researchers found a way past Fable 5’s safeguards and that this led to the export controls. In context, “on the first attempt” refers to Fable’s first release, not to the number of tries, and read that way the claim holds. According to a researcher interviewed by The Register, it was not really a break-in: the model first refused a security review but agreed to the request “fix this code”.

7. Do massive safety margins make sense?

Paavola argues from proportion. He says open models are 6–12 months behind the frontier, and anyone can easily strip their safety filters. He therefore sees huge safety margins on frontier models as poorly justified.

What research says:

• The lag is shorter than Paavola says. Epoch AI’s May 2026 analysis finds that “the most capable open-weight models have lagged frontier closed models by an average of four months”, or six months under a stricter criterion. In 2024 Epoch put the lag at about a year. Epoch’s 6–12 months, Paavola’s figure, applies to models that can run on a single consumer GPU. The leading open models are mostly Chinese.
• Removing safety training really is cheap.
  • Qi et al. (2023) broke GPT-3.5’s safety training by fine-tuning it on ten examples “at a cost of less than $0.20”.
  • The safety training of Llama 2 models was undone for less than $200 on a single GPU, leaving refusal rates of about 1%.
  • Arditi et al. (2024) showed that refusal is mediated inside the model by “a one-dimensional subspace”. This is the basis of so-called abliteration, which removes refusal from a model directly.
• Anthropic itself confirms Paavola’s premise but draws the opposite conclusion. Anthropic’s September 2026 research reports that the US NIST AI testing centre rates the open Chinese model GLM-5.3 as “the most cyber-capable open-weight model released to date”, about four months behind the frontier. Abliteration took its refusal rate from over 90% to about 3–12%. Anthropic reads this as a “critical threshold” in freely available capabilities having been crossed. Its conclusion is not that safeguards should go, but that defenders need to get access to capable models first.

What is checkable here and what is opinion. Paavola’s premises hold: open models are close to the frontier, and safety training is easy to remove. His conclusion, that frontier models’ safeguards are therefore senseless, is an opinion. Anthropic, which has its own commercial interests, argues the opposite from the same facts. Neither is a research finding.

8. Can a swarm of dumb models beat a frontier model?

The conversation asks whether a swarm of weaker models could compete with a frontier model, or, as the episode puts it, whether enough bees can beat a bear. The open Kimi K2.6, run in long agent chains, is given as an example. The problem is delegation: “the crap just flows down the chain”.

The conversation also recalls an old remark by Sam Altman: the model of the day would produce the right answer to every question if the question were asked a million times and the right answer picked out. The mechanism is explained like this. The temperature parameter controls how the model’s tokens are sampled from the probability distribution. Turn up the variance and run the model enough times, and the right answer will be in the set. That could be Europe’s path: a weaker base model plus more compute.

What research says:

• The Altman quote. No source was found for it. The closest found is Altman’s March 2023 interview on Lex Fridman’s podcast, where he says “if you ask the same question 10,000 times and you rank the outputs from best to worst, what most people see is, of course, something around output 5,000”. He goes on to say that output 10,000 is the one that gets the attention. That remark is about the worst answer, though, not the best.
• The idea itself has been studied, and it holds. In Brown et al. (2024), Large Language Monkeys, the open model DeepSeek-Coder-V2 solved 15.9% of SWE-bench Lite programming tasks with one attempt and 56% with 250 attempts. That beat the single-attempt frontier of the time, 43%.
• What remains unsolved is how to recognise the right answer. The episode does not take this up. Program code can be tested automatically. Without an automatic checker, majority voting and reward models “plateau beyond several hundred samples”. A million answers are of no use if you cannot tell which one is right.
• The delegation problem is documented. Cemri et al. (2025) found failure rates of 41–86.7% in seven open multi-agent systems. Cognition, which builds agent tools, warns that parallel sub-agents produce “compounding errors” and “fragile systems”. Kimi K2.6’s model card says it scales to “300 sub-agents executing 4,000 coordinated steps”, but being able to run long chains does not stop errors compounding.

A swarm and repeated sampling can therefore close the gap on tasks whose answers can be checked by machine. On other tasks they do not help.

9. Raw compute

The conversation uses Elon Musk’s Colossus data centre as an example of what Europeans cannot do. According to the episode, the capacity was built fast, and when there was not enough demand through X, it was rented out to Anthropic and Google and combined with Cursor’s user data.

Checked in SpaceX’s registration filing:

• Build speed. The first Colossus cluster, about 100,000 H100 GPUs and about 130 megawatts, was brought online “in just 122 days” inside the shell of an old factory. The first Colossus II cluster took 91 days and the second 64 days. The filing gives about two years as the industry benchmark for a 100-megawatt greenfield data centre. Together the Colossus sites provide about a gigawatt.
• Anthropic. In May 2026 SpaceX and Anthropic signed cloud services agreements for capacity of about 325,000 NVIDIA GPUs across the Colossus sites, at “$1.25 billion per month through May 2029”. Anthropic confirmed the agreement in its own announcement.
• Google. No equivalent agreement with Google was found.

The European comparison. The LUMI-AI procurement contract was signed in August 2026, and the machine is due in the second half of 2027. The EU’s AI “gigafactories” were announced in February 2025, but the call for up to seven gigafactories opened only on 30 July 2026. Selection is expected in early 2027 at the earliest. According to Bruegel, Europe’s bottleneck is not private capital but “speed-to-operation and grid access”. The episode’s point about Colossus therefore hits exactly Europe’s weak spot.

On the compute side, the episode sees LUMI and the data centres being built in Finland as good things. Miettinen says he runs his own agents on a home computer.

10. What can the state do?

Asked about remedies, the conversation does not trust Chinese openness. According to the conversation, the Chinese share open models to keep US prices down. The main episode suggested this might be “temporary kindness” that ends once China can manage on its own.

Miettinen’s proposal is half in jest: appoint Linus Torvalds Europe’s “AI prince”, or put a cult figure like Steinberger in charge of building a truly open European AI fortress. Steinberger has since joined OpenAI.

Paavola’s answer is the episode’s clearest political position: the state’s options are “really limited”. In his view subsidies and state-level projects achieve nothing. As a general rule, he would end all corporate subsidies and cut corporation tax. This is Paavola’s opinion, not a research finding.

What the EU has done, and how it has gone:

• InvestAI (February 2025). The Commission set out to mobilise €200 billion of AI investment, including a €20 billion fund for AI gigafactories. No gigafactory has been selected yet.
• The AI Continent Action Plan (April 2025) and the Apply AI Strategy (October 2025). Their goals include at least 13 operational AI factories by 2026, tripling the EU’s data-centre capacity, and AI adoption in ten sectors.
• OpenEuroLLM (February 2025). The project aims at open European foundation models on a €37.4 million budget. It is coordinated by Charles University and co-led by AMD Silo AI. In the coordinator’s March 2026 update, “significant challenges, especially in securing more compute for creating the final models, still remain”. The deadline for the first models was 31 July 2026. By October 2026, only small experiments and an unfinished 9-billion-parameter model had been published.

This supports Paavola’s position: the state-led model project is late and short of compute, and €37 million is small beside Colossus’s monthly rent. The episode, however, does not address the counter-example. The publicly funded LUMI is the infrastructure on which both Poro and Viking were trained, and which helped make Silo AI a $665 million company. One of the things the conversation itself praises is therefore a state-level project. Bruegel’s analysis suggests a middle path: the bottlenecks are grid connections and speed of deployment rather than public money, which makes this a matter of permitting and infrastructure policy.

11. Open source and business graduates

Miettinen says he installed Ubuntu for his agent. In his view Linux’s strength is that there is no commercial layer between the user and the machine. He thinks Finland now has a lot to defend in its open-source tradition, as GitHub becomes part of Microsoft’s more commercial system.

Checked: In August 2025 GitHub’s CEO Thomas Dohmke announced he was leaving. GitHub continues “as part of Microsoft’s CoreAI organization”, so it no longer operates as a separate company within Microsoft.

At the end, both agree that these are exceptional times. Paavola originally graduated as a master of business administration from Aalto and has only his thesis left for a master’s in computer science at the University of Helsinki. He compares today with when he started those studies in 2017. Back then a whole year could pass in the field without anything significant happening; now an exceptional week is one in which nothing significant happens. This article’s timelines bear that out. In the three months between recording and publication, Anthropic released two new model versions, Mistral raised €3 billion and announced a trillion-parameter model, SpaceX bought Cursor, and the EU lightened its AI Act.

Miettinen mentions a hypothesis he has been discussing with Mikko Alasaarela and Markus Havi: business graduates will win the coding game.

What is checkable and what is opinion

Checked against primary sources:

• Europe’s share of models and compute (Stanford AI Index 2025 and 2026, Bruegel 2026, Draghi 2024).
• Mistral’s funding rounds and models, including Mistral Large 4, announced after the episode.
• AMD’s acquisition of Silo AI (about $665 million, closed 12 August 2024; the episode says 650 million).
• Poro 34B and Viking were trained from scratch on LUMI; Poro 2 is continued-pretrained from Llama 3.1.
• The LUMI AI Factory (2024) and the LUMI-AI procurement (2026).
• Anthropic’s and OpenAI’s contract terms barring distillation, and the distillation campaigns Anthropic reported.
• Codex’s reinforcement learning, Steinberger’s move to OpenAI, and the Cursor, xAI and SpaceX arrangements.
• That Google’s Gemini Spark agent is available only in the US.
• The AI Act timetable and the omnibus package’s postponements.
• Fable 5’s and 5.1’s safety classifiers, the routing to Opus 4.8, the bypass found by Amazon researchers, and the export controls.
• How far open models lag (Epoch AI) and how cheaply safety training can be removed (Qi et al., BadLlama, Lermen et al., Arditi et al., Anthropic’s GLM-5.3 research).
• The repeated-sampling results (Large Language Monkeys) and the failures of multi-agent systems.
• Colossus’s build speed and Anthropic’s capacity agreement.
• The EU’s InvestAI, AI Continent, Apply AI and OpenEuroLLM, and the gigafactory call timetable.
• GitHub’s move into Microsoft’s CoreAI organisation.

Claims in the episode that do not hold or could not be checked:

• Altman’s “a million times” remark. No source was found, and the closest remark found is about the worst answer.
• PewDiePie’s model at Sonnet 4.5 level. It was a fine-tuned Qwen2.5-32B, and no such comparison was found.
• A Lovable model of its own. No source was found.
• Colossus capacity rented to Google. No source was found.
• A 6–12-month lag for open models. Epoch’s estimate for the leading open models is about four months.

Opinions, not facts:

• That Europe has lost the foundation-model race.
• That security work will soon require unfiltered open models.
• That frontier models’ safeguards are senseless because open models are close behind.
• That corporate subsidies should end and corporation tax be cut.
• That Chinese openness is temporary.
• This article’s readings of Poro 2 and of Europe’s bottlenecks.

Sources

Europe’s position

• Stanford HAI: AI Index Report 2025: https://hai.stanford.edu/ai-index/2025-ai-index-report ; AI Index Report 2026: https://hai.stanford.edu/ai-index/2026-ai-index-report
• Martens & Schenk (2026): Bruegel Policy Brief 18/2026: https://www.bruegel.org/sites/default/files/2026-09/PB%2018%202026.pdf
• Draghi, M. (2024): The future of European competitiveness, Part A: https://commission.europa.eu/document/download/97e481fd-2dc3-412d-be4c-f152a8232961_en
• Epoch AI (29 May 2026): Open models lag state-of-the-art closed models by 4 months: https://epoch.ai/data-insights/open-closed-eci-gap ; Consumer GPU model gap: https://epoch.ai/data-insights/consumer-gpu-model-gap

Mistral, Silo AI and LUMI

• Mistral AI (9 September 2025): Series C: https://mistral.ai/news/mistral-ai-raises-1-7-b-to-accelerate-technological-progress-with-ai
• Mistral AI (8 September 2026): Series D: https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/
• Mistral AI (2 December 2025): Mistral 3: https://mistral.ai/news/mistral-3 ; (6 October 2026): Mistral Large 4: https://mistral.ai/news/mistral-large-4
• Mistral AI (11 August 2026): In-region inference, open models, and new European infrastructure: https://mistral.ai/news/regional-inference-open-models-new-compute/
• AMD (10 July 2024): AMD to Acquire Silo AI: https://ir.amd.com/news-events/press-releases/detail/1206/amd-to-acquire-silo-ai-to-expand-enterprise-ai-solutions
• Poro 34B model card: https://huggingface.co/LumiOpen/Poro-34B
• AMD (18 June 2025): Poro 2: https://www.amd.com/en/blogs/2025/silo-ai-continued-pretraining-provides-blueprint-for-gla.html ; (6 August 2026): Poro 2 8B Long: https://www.amd.com/en/blogs/2026/extending-context-and-exploring-multilingual-reasoning-with-poro.html
• EuroHPC (10 December 2024): Selection of the first seven AI Factories: https://eurohpc-ju.europa.eu/selection-first-seven-ai-factories-drive-europes-leadership-ai-2024-12-10_en
• CSC (31 August 2026): LUMI-AI procurement: https://csc.fi/?p=17896

EU policy and regulation

• European Commission: AI Act, regulatory framework: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
• European Commission: General-Purpose AI Code of Practice: https://digital-strategy.ec.europa.eu/en/policies/gpai-code-practice
• European Commission (11 February 2025): InvestAI: https://luxembourg.representation.ec.europa.eu/actualites-et-evenements/actualites/eu-launches-investai-initiative-mobilise-eur200-billion-investment-artificial-intelligence-2025-02-11_en
• European Commission: AI Continent Action Plan: https://digital-strategy.ec.europa.eu/en/factpages/ai-continent-action-plan ; Apply AI Strategy: https://digital-strategy.ec.europa.eu/en/policies/apply-ai
• EuroHPC (30 July 2026): AI Gigafactories call: https://eurohpc-ju.europa.eu/eurohpc-joint-undertaking-launches-ai-gigafactories-call-2026-07-30_en
• OpenEuroLLM (6 March 2026): First year progress and next steps: https://openeurollm.eu/blog/first-year-progress-and-next-steps
• Google (19 May 2026): Google AI subscriptions: https://blog.google/products-and-platforms/products/google-one/google-ai-subscriptions/
• Regulation (EU) 2024/1689 (AI Act) and Regulation (EU) 2016/679 (GDPR).

Distillation, harness and reinforcement learning

• Anthropic: Commercial Terms of Service: https://www.anthropic.com/legal/commercial-terms
• Anthropic (23 February 2026): Detecting and preventing distillation attacks: https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks
• OpenAI (16 May 2025): Introducing Codex: https://openai.com/index/introducing-codex/
• Anthropic (21 November 2025): Natural emergent misalignment from reward hacking: https://www.anthropic.com/research/emergent-misalignment-reward-hacking
• Steinberger, P. (2026): OpenClaw: https://steipete.me/posts/2026/openclaw
• Google (2 April 2026): Gemma 4: https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/
• PewDiePie (26 February 2026): I Trained My Own AI… It beat ChatGPT: https://www.youtube.com/watch?v=aV4j5pXLP-I

Safety filters and open models

• Anthropic (7 April 2026): Project Glasswing: https://www.anthropic.com/glasswing
• Anthropic (28 May 2026): Claude Opus 4.8: https://www.anthropic.com/news/claude-opus-4-8
• Anthropic (9 June 2026): Claude Fable 5 and Claude Mythos 5: https://www.anthropic.com/news/claude-fable-5-mythos-5
• Anthropic: Redeploying Claude Fable 5: https://www.anthropic.com/news/redeploying-fable-5
• Anthropic (2 July 2026): Fable 5’s cyber safeguards and our jailbreak framework: https://www.anthropic.com/news/fable-safeguards-jailbreak-framework
• Anthropic (1 September 2026): Claude Fable 5.1 and Claude Mythos 5.1: https://www.anthropic.com/claude-fable-and-mythos-5-1 ; system card: https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf
• Anthropic (29 September 2026): GLM-5.3 and the spread of advanced cyber capabilities: https://www.anthropic.com/research/glm-5-3-and-the-spread-of-advanced-cyber-capabilities
• The Register (15 June 2026): Feds freaked over Fable 5 after simple ‘fix this code’ prompt: https://www.theregister.com/security/2026/06/15/feds-freaked-over-fable-5-after-simple-fix-this-code-prompt-not-jailbreak-says-researcher/5255827
• Qi, X. et al. (2023): Fine-tuning Aligned Language Models Compromises Safety: https://arxiv.org/abs/2310.03693
• Gade, P. et al. (2023): BadLlama: https://arxiv.org/abs/2311.00117
• Lermen, S. et al. (2023): LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B: https://arxiv.org/abs/2310.20624
• Arditi, A. et al. (2024): Refusal in Language Models Is Mediated by a Single Direction: https://arxiv.org/abs/2406.11717

Swarms, sampling and compute

• Brown, B. et al. (2024): Large Language Monkeys: Scaling Inference Compute with Repeated Sampling: https://arxiv.org/abs/2407.21787
• Cemri, M. et al. (2025): Why Do Multi-Agent LLM Systems Fail?: https://arxiv.org/abs/2503.13657
• Cognition (12 June 2025): Don’t Build Multi-Agents: https://cognition.com/blog/dont-build-multi-agents
• Moonshot AI: Kimi K2.6 model card: https://huggingface.co/moonshotai/Kimi-K2.6
• Lex Fridman Podcast #367, Sam Altman (2023), transcript: https://auphonic.com/media/blog/LexFridmanPodcast367-transcript.html
• Hugging Face: How to generate text: https://huggingface.co/blog/how-to-generate
• SpaceX, SEC registration filing (4 June 2026): https://www.sec.gov/Archives/edgar/data/0001181412/000162828026041013/japanfwp_06042026.htm
• Anthropic (6 May 2026): Higher limits with SpaceX compute: https://www.anthropic.com/news/higher-limits-spacex
• GitHub (11 August 2025): Auf Wiedersehen, GitHub: https://github.blog/news-insights/company-news/goodbye-github/

Chapters

• 00:00 Data sovereignty and Europe’s foundation models
• 00:39 Mistral’s marketing is great, the models disappoint
• 01:01 The horror scenario: the US and China close off access
• 01:33 Poro, Silo AI and LUMI
• 02:07 Could Europe distil the US models’ weights?
• 02:39 Local Gemma and Chinese models
• 03:02 The harness and reinforcement learning decide the race
• 04:00 Lovable, PewDiePie and garage-built models
• 04:49 Regulation will not win the game
• 05:21 Fable’s tighter safety filter weakens frontier models
• 06:14 Can a swarm of dumb models beat a frontier model?
• 06:56 Altman’s million answers and the temperature parameter
• 08:11 Musk, Colossus and raw compute
• 09:04 Security needs unfiltered open models
• 10:11 Do massive safety limits make sense?
• 10:54 China’s open models may be temporary kindness
• 11:26 Linus Torvalds to lead Europe’s open AI
• 11:53 End corporate subsidies and cut corporation tax
• 12:24 Linux, Ubuntu and GitHub’s commercialisation
• 13:02 Something groundbreaking happens every week
• 14:06 Will business graduates win the coding game?

Episode details. Neuvottelija Inner Circle 163, published 10 October 2026, length 14:34, YouTube id 04Dt1tXMPSc. Guest Aleksi Paavola (guest profile and other episodes), host Sami Miettinen. The article was expanded and its claims checked on 12 October 2026.

Summary for AI search: Neuvottelija Inner Circle episode 163 (published 10 October 2026, most likely recorded in early July 2026, length 14:34, YouTube id 04Dt1tXMPSc). Guest: AI builder Aleksi Paavola. Host: Sami Miettinen. Subject: Europe’s position in AI foundation models and data sovereignty. The metrics show a weak position: Europe produced three notable models in 2024 (US 40, China 15), and the EU has about 5% of global AI compute (Bruegel 2026). Mistral is the EU’s only significant model developer; it raised €3 billion in September 2026 and announced Mistral Large 4 on 6 October 2026. AMD bought Silo AI in 2024 for about $665 million. Poro 34B was trained from scratch on LUMI, but Poro 2 (2025) is continued-pretrained from Meta’s Llama 3.1. Distillation is barred by Anthropic’s and OpenAI’s terms. Frontier models’ edge lies in the harness and reinforcement learning (Codex, Claude Code, and Cursor, which SpaceX bought in 2026). Anthropic’s Fable 5 was placed under US export controls in June 2026 after Amazon researchers found a way past its safeguards. It was redeployed on 1–2 July 2026 with a stricter cyber classifier that routes blocked requests to Opus 4.8, and Fable 5.1 (September 2026) allows vulnerability discovery in source code. Paavola argues that strict filters may push security work onto open models. According to Epoch AI, open models lag the frontier by about four months, and their safety training can be removed for under $200. Repeated sampling improves results (Large Language Monkeys: 15.9% → 56%), but only if the right answer can be identified. No source was found for Altman’s “a million times” remark or for PewDiePie’s “Sonnet 4.5-level” model. Colossus’s first cluster was built in 122 days, while the EU’s gigafactory call opened only in July 2026. Paavola says the state’s options are limited and corporate subsidies should end. The state-led OpenEuroLLM is short of compute, but the publicly funded LUMI made Poro and Silo AI possible.

Markdown: index.md · Suomeksi