In late 2015, a computer scientist named Steven Flinn and a co-inventor filed a patent application describing a way for a neural network to decide which parts of an input deserved attention. The idea was not new to researchers. It was new to the patent office, which granted claims on it. On Thursday, a company Flinn founded used those claims to sue two of the most valuable artificial-intelligence companies in the world.
Flinn AI filed two complaints in the U.S. District Court for the District of Delaware, accusing Anthropic and OpenAI of infringing five patents covering attention-based neural network techniques. The applications behind them date to late 2015 and early 2016, before the papers that made attention the default architecture of the field were widely cited in industry products.
The complaint against Anthropic centers on the Claude models. It argues that each successive release widened the footprint of the infringement, reaching into attention-based inference, retrieval, memory, agent behavior and multimodal processing. The complaint against OpenAI uses similar language. Neither company responded to requests for comment.
The timing is the part patent lawyers noticed first. Filings made in 2015 and 2016 sit on the far side of a doctrinal line. Courts have spent a decade narrowing what counts as patentable subject matter in software, and a claim that reads as an abstract idea implemented on a generic computer is invalid under the Alice test. A claim that describes a specific technical improvement to how a machine processes data has a better chance. Attention mechanisms, which change the arithmetic a network performs rather than the business purpose it serves, fall into the second category more comfortably than most software patents do.
“Whether these claims survive is a question about the specification, not the headline,” said one patent litigator who reviewed the complaints. “If the patents describe a concrete implementation, the case has legs. If they describe a goal, it does not.”
Damages are the other half of the argument. A reasonable royalty is calculated against a royalty base, and the royalty base here would be enormous: revenue from the models and from the cloud services that resell them. Defendants typically argue for apportionment, meaning the royalty should attach only to the portion of the product that practices the patent. Flinn AI’s complaints describe the technology as constituting core functionality, which is a way of arguing that the apportionment math should not shrink the base much.
Delaware is a deliberate choice. The district has a deep bench of judges who handle patent cases routinely, no local bias toward either coast’s technology industry, and a docket that moves. It is also where companies incorporate, which means a judgment is easier to enforce against assets that live in the same state.
Prior art will be the defense’s first move. Attention, as a concept, appeared in academic literature before 2015, and the defendants will argue that the patents claim something the field already knew, dressed in the vocabulary of a patent application. Flinn’s answer, according to a person familiar with the company’s strategy, is that the claims cover a specific combination of mechanisms rather than the general idea, and that the combination was not obvious at the time.
The litigation history around Flinn is tangled. He and his co-inventor also run ManyWorlds, a separate company. Earlier this year, an entity with a similar name, Many Worlds 2T Innovations, sued OpenAI and Spotify over a different set of patents covering content recommendation. Patent lawyers said the overlap in names will invite defense arguments that the plaintiffs are coordinating, and that the plaintiffs will argue the opposite: that separate portfolios reflect separate inventions.
The economics of the plaintiff’s side deserve a note. Cases like this are expensive to run, and a small company suing two large ones usually does it with money it does not have. Contingency arrangements with litigation funders are common, and they shift part of the recovery to a third party in exchange for carrying the cost of expert reports, claim construction and the depositions that follow. A person familiar with the arrangements said the company has outside backing, without describing terms.
For the defendants, the cost of the case is mostly the distraction. Both companies are in the middle of capital-intensive expansion, and both face shareholder scrutiny of their legal exposure in filings that describe litigation as a risk factor. A patent suit filed in Delaware does not stop a product launch. It does put a number in the footnotes.
Settlement is the likeliest ending, as it is in most patent cases that survive a motion to dismiss. The question that decides the price is whether the claims reach the training process or only the inference step. A royalty on inference is a tax on usage. A royalty on training is a tax on existence, and defendants fight those much harder.
Flinn, for his part, has been arguing for years that the industry built its products on ideas that were written down and patented by people outside its celebrity circle. The complaints are his attempt to collect. Whether a judge agrees will take two to three years, and by then the models named in the filings will be obsolete while the patents will not.
