STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 07

WEEK OF FEB 8 – 14, 2025

A judge ruled against fair use, and two governments refused to sign the AI declaration

The first US decision squarely on whether training on copyrighted work is fair use went against the defendant — in the same week the Paris summit split the room and an open-data reasoning recipe became fully reproducible.

Judge Stephanos Bibas held that Ross Intelligence's copying of Westlaw headnotes to train a legal-research AI was not fair use — the first US decision addressing training-as-such, and it landed while the larger cases against OpenAI and Anthropic are still running.

At the Paris AI Action Summit, more than 60 countries signed a declaration on inclusive AI; the US and the UK declined. France pledged €109B and the EU launched a €200B InvestAI mobilization. OpenAI also rejected a $97.4B takeover bid from Elon Musk.

FRI · Feb 14, 2025industryopenaimusknonprofit governance

OpenAI's board unanimously rejects Elon Musk's $97.4B takeover bid

OpenAI's board rejected a $97.4 billion unsolicited bid from a consortium led by Elon Musk, saying the company is not for sale and calling any future bid disingenuous.

Board chairman Bret Taylor said on behalf of the board that the offer was Musk's latest attempt to disrupt his competition, and that any reorganization would strengthen the nonprofit. Musk submitted the bid on 10 February as an offer for the nonprofit that controls the for-profit arm; Sam Altman publicly refused it the same day. The bid is part of Musk's broader effort to block OpenAI's conversion to a for-profit structure, which is also the subject of his separate lawsuit.

OpenAI's board unanimously rejects Elon Musk's $97.4B takeover bid
www.gov.uk

WHY IT MATTERS

The for-profit conversion and its capital raise continue on OpenAI's own terms, with the nonprofit's board — not outside bidders — setting the price of control. It is also the first time the governance fight moved from a courtroom to a hostile bid, and the board's refusal frames the conversion as a defensive act rather than a commercial one.

Reuters — OpenAI board rejects Musk's $97.4 billion offerThe New York Times — Elon Musk leads $97.4 billion bid to control OpenAI
FRI · Feb 14, 2025safetypolicyuklab leadershipanthropic

UK's AI Safety Institute becomes the AI Security Institute — and signs an Anthropic deal

At the Munich Security Conference, UK Technology Secretary Peter Kyle recast the AI Safety Institute as the AI Security Institute, days after the Paris summit.

The remit narrows to national security and crime, including chemical and biological weapons, cyber-attacks, fraud and child sexual abuse, with a new criminal-misuse team run jointly with the Home Office and a partnership with DSTL to assess frontier-AI risk. The same announcement carried a new agreement with Anthropic, published as a Memorandum of Understanding, to apply Anthropic's models to public services.

UK's AI Safety Institute becomes the AI Security Institute — and signs an Anthropic deal
www.gov.uk

WHY IT MATTERS

Read next to the week's open-weight releases: the UK's evaluation capacity is being pointed at security harms rather than societal and bias harms, so the capability bar for entering the Institute's work has moved. A national evaluator signing an MOU with one lab at the same time is also worth watching as a structural question about who assesses whom.

GOV.UK — Tackling AI security risksGOV.UK — MoU between the UK and AnthropicAnthropic — Anthropic signs MoU with UK Government
WED · Feb 12, 2025researchsafetymodel specopenairoadmap

OpenAI rewrites its Model Spec into 63 pages and releases it under CC0

OpenAI published a major Model Spec update — 63 pages, up from roughly 10 previously — and released it into the public domain under CC0 so anyone can reuse or fork it.

The rewrite foregrounds customizability, transparency and intellectual freedom, meaning fewer arbitrary content restrictions while keeping guardrails against real harm, and it explicitly loosens how the model handles controversial topics and adult content. OpenAI also published early adherence results against the spec's principles. On the same day, Sam Altman laid out a product roadmap: GPT-4.5 in weeks, GPT-5 in months as a unified system collapsing the o-series into the mainline model.

WHY IT MATTERS

Both halves matter for anyone building on OpenAI APIs. The behavior contract is now public and forkable — which cuts both ways, since it also tells you exactly what the model will and won't refuse — and the model lineup is about to consolidate, which changes how you route between reasoning and standard models.

OpenAI — Sharing the latest Model SpecThe Verge — OpenAI is rethinking how AI models handle controversial topicsArs Technica — Altman lays out roadmap for GPT-5
WED · Feb 12, 2025open weightsreasoningpost-trainingdistillationmultilingual

Fully open reasoning: OpenThinker-32B matches closed-data distills, and BSC ships an Apache-2.0 family

OpenThoughts released OpenThinker-32B under Apache 2.0, built on a Qwen2.5-32B base and trained on OpenThoughts-114k — a dataset whose problems, ground-truth solutions, test cases and DeepSeek-R1 reasoning traces are all published, along with the data-curation code.

The team's point is methodological: scale the data, verify reasoning traces, scale the model, and an open-data 32B model outperforms existing open-data reasoning models on math, code and science — while the comparable R1-Distill-32B is a closed-data model. The same day, the Barcelona Supercomputing Center published Salamandra: 2B, 7B and 40B decoder-only models trained from scratch on 35 European languages plus code, released under Apache 2.0 with training and evaluation scripts public.

WHY IT MATTERS

Nobody here is at frontier scale — no 32B open model yet closes on the flagship reasoning systems — but the reasoning-distillation recipe is now fully reproducible on a single 32B-class training run rather than a trade secret, and that is where the open-versus-closed cost gap actually gets set. Note that the differentiator is published data and verification, not raw compute.

OpenThoughts — Scaling up open reasoning with OpenThinker-32BHugging Face — open-thoughts/OpenThinker-32B (Apache 2.0)arXiv — Salamandra Technical Report
TUE · Feb 11, 2025regulationpolicycomputedata centersparis summit

Paris AI Action Summit: US and UK refuse to sign, France pledges €109B, EU launches €200B InvestAI

The AI Action Summit in Paris ended with more than 60 countries, the African Union Commission and the EU signing a declaration on inclusive and sustainable AI — and the United States and the United Kingdom declining to sign it, on grounds of over-regulation and national-security framing respectively.

The investment numbers were the concrete output: France announced €109 billion in private AI commitments, and the European Commission launched InvestAI, a €200 billion mobilization including a €20 billion European fund for AI gigafactories of roughly 100,000 chips each — about four times the size of the AI factories already being built.

WHY IT MATTERS

Hold the contrast: the same week two governments declined a soft declaration, an openly licensed 32B reasoning model with a fully published data pipeline was released. The coordination question and the capability question are moving at different speeds, and the money is going into compute rather than into agreement.

The Guardian — US and UK refuse to sign Paris summit declarationBBC — UK and US refuse to sign international AI declarationEuropean Commission — EU launches InvestAI to mobilize €200 billion
TUE · Feb 11, 2025litigationcopyrightfair usetraining data

Delaware judge rejects fair use for AI training in Thomson Reuters v. Ross Intelligence

Judge Stephanos Bibas of the US District Court for the District of Delaware granted Thomson Reuters partial summary judgment, holding that Ross Intelligence's copying of Westlaw headnotes to train a legal-research AI was not fair use.

The court rejected the transformative-use argument because the material was used to build a direct market substitute, and split the fair-use factors against Ross. It is the first US decision squarely addressing whether training an AI model on copyrighted material qualifies as fair use, and it landed while the larger training-data cases against OpenAI, Anthropic and others are still running.

WHY IT MATTERS

The useful takeaway is narrower than the headlines suggested: the ruling turns on substituting for the original work in the same market, not on training as such. But it is now the precedent judges and defendants will argue from in every other training-data case — and it is the first data point suggesting the answer will be fact-specific rather than categorical.

Reuters — Thomson Reuters wins AI copyright fair use rulingCourt opinion (PDF)