IP Law Daily Disputed facts precluded finding that start-up’s former employee shared secret algorithms with Facebook
Friday, June 5, 2020

Disputed facts precluded finding that start-up’s former employee shared secret algorithms with Facebook

By Robert Margolis, J.D.

Neural Magic failed to prove as a matter of law that its former technology director published secret code on open source platform after joining Facebook.

A federal district court in Boston has denied Neural Magic, Inc.’s motion for preliminary injunctive relief against its former Technology Director, Dr. Aleksander Zlateski, and his new employer Facebook, Inc., finding too many disputed facts about whether the algorithms at issue actually were trade secrets and whether the defendants used improper means to acquire them. Neural Magic sued Facebook and Zlateski for federal and state law trade secret misappropriation, and Zlateski for breach of his non-disclosure and non-competition agreement, based on his alleged disclosure of algorithms to Facebook, Facebook’s alleged internal use of them, and its alleged posting of them on the open source forum GitHub (Neural Magic, Inc. v. Facebook, Inc., May 29, 2020, Casper, D.).

Neural Magic is a recent start-up, the goal of which is to create technology that can "perform sparse matrix multiplication on a conventional CPU [computer processing unit] at an unprecedented rate," which would speed up CPUs and allow them to analyze larger data sets. The technology would be useful to researchers, universities, direct-to-customer services and even Facebook itself (and Google), which use "hundreds of thousands of CPUs and GPUs to operate their business and research."

Neural Magic hired Aleksander Zlateski as Technology Director in March 2018 to help develop that technology. He signed a nondisclosure and non-competition agreement. Though Zlateski worked together with Neural Magic’s founders to create programming and algorithms, they did not have a product to sell by Fall 2018. Neural Magic contends that it planned to launch its product in April 2020.

In March 2019, Facebook offered Zlateski a position; he told Neural Magic this immediately, and stayed at Neural Magic until July 2019, when he joined Facebook. Neural Magic allowed Zlateski to keep his company laptop, as there was hope they could still collaborate even after he went to Facebook.

At Facebook, Zlateski was asked to write optimization code for Facebook General Matrix Multiplication ("FBGEMM"), what the court describes as "a code library that [Facebook] develops and uses to optimize performance across numerous machine learning ("ML") and AI applications." FBGEMM development began before Zlateski was hired, and is not a product that Facebook sells. On November 1, 2019, Facebook posted the FBGEMM that Zlateski had optimized on GitHub, and one of Zlateski’s supervisors on LinkedIn credited Zlateski for using a tool that Neural Magic claims was developed by Zlateski when he worked there. Zlateski contends that he wrote the Facebook code "from scratch" and did not use any Neural Magic code.

Neural Magic demanded that Facebook remove the FBGEMM from GitHub, which Facebook eventually did, though it did not agree to all of the relief Neural Magic asked for, leading Neural Magic to sue and seek injunctive relief on its misappropriation and breach of contract claims.

Widely known. The court found that Neural Magic did not show a likelihood of success on any element of its trade secret claim. First, there were sufficient disputed facts to call into question whether the algorithms that Neural Magic claimed are trade secrets even qualify. Facebook and Zlateski argued that the code is based on widely known concepts and approaches, and that Zlateski knew and used them before joining Neural Magic. The defendants presented enough evidence to support these contentions, such that Neural Magic could not meet the high "likelihood of success" standard for injunctive relief, the court held.

No economic value. Similarly, the court held that Neural Magic could not make the requisite "likelihood of success" showing that it derived economic value from the alleged trade secrets. Evidence of potential benefits to potential customers notwithstanding, it does not have a product yet on the market and has not yet earned any revenue, the court pointed out.

Reasonable steps. The court found it a "close call" whether Neural Magic took reasonable steps to protect its algorithms. Much of its argument rested on its non-disclosure agreement with Zlateski, which it requires of employees, potential employees, and "alpha customers and beta testers." In addition, its product is designed as a "closed-source software product," meaning no customer receives the underlying source code and any algorithms are hidden from the user. While these are important protective steps, that still left open the question of whether Neural Magic took sufficient steps with regard to Zlateski’s departure to Facebook, which Neural Magic knew to be a major user of AI and uses many CPUs and GPUs. The court agreed with Neural Magic that it should be able to rely on Zlateski to abide by his agreement, but also found the defendants’ argument reasonable that more should have been done in this specific case. Zlateski was allowed to remain with Neural Magic for several months after he announced his intent to leave for Facebook, he was not denied access to their systems, required to turn in his laptop, or scrub his computer before his departure. Moreover, while Neural Magic took the reasonable step of contacting Facebook to remove the publication of the alleged trade secret from GitHub, when some remnants of it remained on GitHub, Neural Magic did not contact GitHub directly, and there was some evidence that GitHub itself can remove sensitive data.

No improper means. Nor did Neural Magic establish that the defendants acquired the alleged trade secret by improper means. While Neural Magic argued that Facebook shifted to a use of the technology embodied by its trade secret after Zlateski became employed, evidence in the record supported Facebook’s contention that it had used some of that technology since 2015, and other aspects before Zlateski’s arrival. Thus, Neural Magic did not show sufficient evidence of a shift in its product to establish that trade secrets were wrongfully acquired through Zlateski. In addition, Neural Magic acknowledged differences exist between its assembly code and Facebook’s, and Zlateski testified that he wrote the code for Facebook from scratch. Given all that, Neural Magic did not show a substantial likelihood of success on this element.

Contract claim. Neural Magic alleged that Zlateski breached the non-disclosure agreement’s obligation not to disclose confidential information. As noted above, the court found that Neural Magic failed to establish a reasonable likelihood of success in making such a showing, which doomed injunctive relief for the contract claim as well as the trade secrets claim.

This case is No. 1:20-cv-10444-DJC.

Attorneys: Kaitlyn M. O'Connor (Quinn Emanuel Urquhart & Sullivan, LLP) for Neural Magic, Inc. Christopher W. Henry (Latham & Watkins LLP) and Russell Beck (Beck Reed Riden LLP) for Facebook, Inc. and Aleksandar Zlateski.

Companies: Neural Magic, Inc.; Facebook, Inc.

MainStory: TopStory TechnologyInternet TradeSecrets GCNNews MassachusettsNews

Back to Top

Interested in submitting an article?

Submit your information to us today!

Learn More

IP Law Daily: Breaking legal news at your fingertips

Sign up today for your free trial to this daily reporting service created by attorneys, for attorneys. Stay up to date on intellectual property legal matters with same-day coverage of breaking news, court decisions, legislation, and regulatory activity with easy access through email or mobile app.