Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 1 of 24
`
`
`
`
`
`v.
`
`NEURALMAGIC, INC.
`
`Plaintiff,
`
`FACEBOOK, INC. AND ALEKSANDAR
`ZLATESKI
`
`Defendants.
`
`
`
`
`
`UNITED STATES DISTRICT COURT
`FOR THE DISTRICT OF MASSACHUSETTS
`
`)
`)
`)
`)
`)
`)
`)
`)
`)
`)
`)
`)
`)
`)
`)
`)
`
`
`Civil Action No. 20-10444
`
`Jury Trial Demanded
`
`
`
`
`COMPLAINT
`
`Plaintiff NeuralMagic, Inc. (“Neural Magic” or the “Company”), by and through its
`
`undersigned attorneys, Quinn Emanuel Urquhart & Sullivan LLP, for its claims against
`
`Defendants, Facebook, Inc. (“Facebook”) and Aleksandar Zlateski (“Zlateski,” and, collectively,
`
`“Defendants”) hereby states as follows.
`
`NATURE OF THE ACTION
`
`1.
`
`Neural Magic is a small start-up co-founded by MIT professor Nir Shavit and
`
`MIT research scientist Alex Matveev in 2017 and based in Somerville, MA. One of Neural
`
`Magic’s technologies—a set of computer algorithms encompassed within a machine compiler—
`
`is the result of decades of research on neural networks and artificial intelligence. These
`
`algorithms have the potential to revolutionize the field of artificial intelligence (“AI”), in part by
`
`allowing complicated mathematical functions
`
`to run efficiently on commodity-based
`
`
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 2 of 24
`
`
`
`computers—using no specialized hardware. These algorithms will also allow research scientists
`
`to use much larger data sets, heretofore a severe limitation on the advancement of machine
`
`learning. Indeed, Neural Magic’s technology is a key to unlocking the next wave of
`
`advancements in many fields: from healthcare and cancer screening, to how customers shop
`
`online and identify items of interest. This technology can also help bring AI to the masses by
`
`reducing barriers of hardware scarcity and cost.
`
`2.
`
`Defendant Aleksandar Zlateski was a trusted part of the Neural Magic team
`
`almost from its inception; he was Neural Magic’s first employee. As Technology Director at
`
`Neural Magic, Zlateski had access to all of Neural Magic’s trade secrets, confidential,
`
`proprietary information, and business plans for the future. And, critically, he had access to and
`
`helped author the software for one of Neural Magic’s prized jewels: the source code of its
`
`compiler that encapsulates Neural Magic’s above-described algorithms.
`
`3.
`
`In July 2019, Zlateski left Neural Magic to work in a capacity that he represented,
`
`and Neural Magic understood, was unrelated to the particular compiler work he had done for the
`
`Company. While Facebook is one of the largest users of artificial intelligence in the world—and
`
`a potential customer of Neural Magic’s—Neural Magic trusted Zlateski to honor non-disclosure
`
`and non-compete agreements he had entered in connection with his work at Neural Magic, and
`
`Neural Magic believed it had nothing to fear from his anticipated work in an unrelated area of
`
`artificial intelligence.
`
`4.
`
`That trust was misplaced. Less than six months later, in December 2019,
`
`Facebook announced that it had published to the world—as open source—a compiler that,
`
`investigation would later reveal, includes the same proprietary algorithms that form the heart of
`
`Neural Magic’s technology and intellectual property. At the time, Facebook even publicly
`
`
`
`
`- 2 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 3 of 24
`
`
`
`thanked Zlateski for his role in cracking this key problem for Facebook’s continued advancement
`
`in the world of artificial intelligence, writing in their release notes for the misappropriated
`
`compiler algorithms that the “team would like to acknowledge and greatly appreciates the
`
`contributions of @zlateski to sparse kernels and unified code cache.”
`
`5.
`
`Zlateski breached the non-disclosure and non-competition agreement he signed
`
`with Neural Magic as Technology Director. Moreover, Zlateski and his new employer Facebook
`
`engaged in acts and conduct in the Commonwealth of Massachusetts that violate Massachusetts
`
`trade secrets laws, Chapter 93A, and the Defend Trade Secrets Act of 2016. Facebook has
`
`refused repeated requests to cease these acts and remove misappropriated material from its own
`
`code. Neural Magic has therefore been forced to bring these claims to protect its intellectual
`
`property.
`
`PARTIES
`
`6.
`
`Plaintiff Neural Magic is a Delaware corporation, with its corporate headquarters
`
`in Somerville, Massachusetts.
`
`7.
`
`Defendant Facebook is a Delaware corporation with a principal place of business
`
`at 1 Hacker Way, Menlo Park, California. Facebook maintains offices in Cambridge,
`
`Massachusetts.
`
`8.
`
`Defendant Aleksandar Zlateski is a natural person who, on information and belief,
`
`resides at 211 N End Ave Apt 3K New York, NY 10282. Zlateski worked as a Neural Magic
`
`employee in Massachusetts for over a year pursuant to the employment contract he signed with
`
`Neural Magic, governed by Massachusetts law. During that time, he helped develop the
`
`proprietary algorithms and trade secrets that are the subject of this action.
`
`
`
`
`- 3 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 4 of 24
`
`
`
`JURISDICTION & VENUE
`
`9.
`
`This is an action for trade secret misappropriation under state trade secrets law
`
`and the Defend Trade Secrets Act of 2016 (18 U.S.C. § 1836 et seq.), and state and common law
`
`claims for breach of contract, tortious interference with contractual relations and prospective
`
`contractual relations, unjust enrichment, and unfair trade practices. This Court has jurisdiction
`
`pursuant to 28 U.S.C. §§ 1331, 1338(a), 1338(b), and 1367(a).
`
`10.
`
`This Court has personal jurisdiction over Facebook, Inc. Facebook maintains
`
`offices in Cambridge, MA. Furthermore, Facebook transacts business in Massachusetts, has
`
`caused tortious injury to Neural Magic in Massachusetts, has an interest in using or possessing
`
`real property in Massachusetts, and contracts to supply services or things in Massachusetts.
`
`11.
`
`This Court has personal jurisdiction over Aleksandar Zlateski. Aleksandar
`
`Zlateski resided in Massachusetts, performed services for a Massachusetts company, Neural
`
`Magic, and entered into a contract in Massachusetts governed by Massachusetts law, forming the
`
`foundation for many of the claims at issue here. Zlateski worked on the algorithms and trade
`
`secrets that form the basis for this action during his time at Neural Magic in Massachusetts.
`
`Furthermore, Zlateski has transacted business with Neural Magic in Massachusetts, has served as
`
`Technology Director of Neural Magic in Massachusetts, maintains a stake in Massachusetts-
`
`based Neural Magic, and has caused tortious injury to Neural Magic in Massachusetts.
`
`12.
`
`Venue is proper in this Court pursuant to 28 U.S.C. §§ 1391(b) and 1391(c).
`
`Aleksandar Zlateski is a former resident of Massachusetts, a substantial part of the events giving
`
`rise to the claims alleged herein occurred in Massachusetts, and Facebook, Inc. has a regular and
`
`established place of business in Massachusetts.
`
`
`
`
`- 4 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 5 of 24
`
`
`
`FACTUAL ALLEGATIONS
`
`Neural Magic’s Founding
`
`Nir Shavit is the CEO & Co-Founder of Neural Magic. He is a Professor of
`
`A.
`
`13.
`
`Computer Science at the Massachusetts Institute of Technology. Over the course of his more
`
`than three-decade career, Professor Shavit has worked tirelessly to advance computer processing.
`
`In 2016, after a long and successful career in the field of multicore processing, Professor Shavit
`
`embarked on a new challenge—the development of artificial intelligence systems to reconstruct
`
`the connectivity of neural tissue in brains.
`
`14. While Professor Shavit and Matveev were working in their MIT Lab running tests
`
`on a large set of neurobiology data, they discovered that, with the right algorithms, they could
`
`run neural networks on these large neurobiological datasets using only standard computers
`
`without specialized hardware. In particular, Professor Shavit and Matveev realized that they
`
`could achieve these exceptional speeds on standard computers that have much higher memory
`
`capacities. That was the birth of Neural Magic, and together Professor Shavit and Matveev
`
`founded the company in 2018 to bring their vision to fruition and to the marketplace.
`
`15.
`
`After extensive multicore computing research, testing, analysis, and refining of
`
`the algorithms that Professor Shavit initially developed in his neurobiology lab, the novel
`
`algorithms, techniques, design patterns, optimization strategies, and formulas (collectively, the
`
`“Neural Magic Algorithms”) at the heart of Neural Magic were created. The Neural Magic
`
`Algorithms enable certain types of neural networks to run in a highly-efficient manner on
`
`commodity Central Processing Units (“CPUs”)—something found on every laptop and
`
`desktop—instead of specialized chips like GPUs. Through these novel algorithms, and the
`
`unique way in which they work together to overcome obstacles experts in the field viewed as
`
`
`
`
`- 5 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 6 of 24
`
`
`
`impassable, Neural Magic is shattering the hardware barriers that were impeding the field of
`
`machine learning and artificial intelligence. This breakthrough represents an incredible value
`
`proposition with implications throughout medicine, the online economy, and artificial
`
`intelligence, and is poised to make machine learning more accessible and affordable, and to
`
`unlock the next great wave of machine learning research.
`
`16.
`
`Indeed,
`
`the
`
`investment community recognized
`
`the significance of
`
`this
`
`breakthrough. Neural Magic has raised $20 million from respected investors including Comcast
`
`Ventures, NEA, Andreessen Horowitz, Pillar VC, and Amdocs Ventures.
`
`17.
`
`Defendant Aleksandar Zlateski was an MIT post-doctoral researcher in a
`
`computer vision group when Professor Shavit and Matveev founded Neural Magic. Excited
`
`about the possibilities of turning CPU-based neural network computation into a commercial
`
`product, Zlateski joined Neural Magic as its first employee, becoming the full-time Technology
`
`Director of Neural Magic on March 19, 2018. As Technology Director, Zlateski was offered a
`
`base salary of $165,000 per year, and was also granted an option to purchase shares of the
`
`company’s stock.
`
`B.
`
`18.
`
`Neural Magic Protects Its Trade Secrets
`
`One of Neural Magic’s key
`
`technologies,
`
`its secret sauce, are
`
`its
`
`algorithms⎯proprietary and heretofore confidential techniques that enable commodity multicore
`
`machines, that is, servers with multiple computing units, to execute complex mathematical
`
`operations in a computationally and storage-efficient manner. Those trade secrets bring the
`
`company significant value, and as such, Neural Magic guards those algorithms closely. When
`
`onboarding any employee—including Zlateski, Neural Magic’s first employee—Neural Magic
`
`highlights the need for confidentiality with a confidentiality and non-disclosure agreement.
`
`
`
`
`- 6 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 7 of 24
`
`
`
`19.
`
`As part of the Neural Magic offer, Zlateski signed a non-disclosure agreement.
`
`Zlateski agreed that he would “not at any time, whether during or after the termination of [his]
`
`engagement by the Company, reveal to any person or entity any of the trade secrets or
`
`confidential, proprietary or other non-public information concerning the organization, business,
`
`finances or assets of the Company . . . including but not limited to information related to
`
`Company
`
`inventions, research,
`
`testing, manufacturing, production, marketing, supplies,
`
`suppliers, consultants, strategic partners, products, designs, methods, know-how, techniques,
`
`systems, processes, software programs and/or code, works of authorship, customer and
`
`collaborator lists, projects, plans, proposals, any Developments, and the notes, memoranda,
`
`reports, lists, records, drawings sketches, specifications, data, documentation or other materials
`
`of any nature containing such trade secret or confidential information (the “Confidential
`
`Information”).
`
`20.
`
`Zlateski further agreed that any inventions he created—either alone or with
`
`others—“during [his] engagement by the Company” that “relates to the business of the Company
`
`. . . results from tasks assigned to [him] by the Company or results from the Confidential
`
`Information or use of facilities, premises or personal property . . . [of] the Company” are the
`
`“sole and absolute property of the Company.”
`
`21.
`
`Neural Magic protects its trade secrets and confidential, proprietary information
`
`through a variety of mechanisms. Neural Magic uses its secure computer network with a VPN
`
`protocol and SSH keys used to login required for any access, physical servers in a secure
`
`location protected by firewall software, and secure office space with electronic keycards required
`
`for entry. It further requires execution of non-disclosure agreements by all employees.
`
`
`
`
`- 7 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 8 of 24
`
`
`
`C.
`
`
`22.
`
`Aleksandar Zlateski Enjoyed Full and Unfettered Access to Neural Magic’s Trade
`Secrets, Research, and Development as Technology Director
`
`Once Zlateski signed
`
`the non-disclosure agreement and expressed his
`
`understanding of the need to keep all information about the company confidential, Zlateski
`
`became a part of the small and trusted Neural Magic team. As a member of that team, Zlateski
`
`enjoyed unfettered access to the research and algorithms that Neural Magic was working so
`
`tirelessly to refine.
`
`23.
`
`Zlateski benefitted greatly from the mentoring and tutelage provided by Professor
`
`Shavit, Dr. Matveev, and other members of the Neural Magic team. These scientists worked
`
`together and collaborated to address many of the most challenging problems in the field of
`
`artificial intelligence, and to practically implement solutions to these problems, one such solution
`
`being the Neural Magic Algorithms at issue here.
`
`24.
`
`As part of his work at Neural Magic, Zlateski co-authored software code that
`
`implemented the Neural Magic Algorithms. That source code is one expression of those
`
`Algorithms and, as such, that code is a physical embodiment of Neural Magic’s trade secrets.
`
`25.
`
`In marketing its products and services, Neural Magic does not reveal this source
`
`code—or the trade secrets contained therein—to its customers or partners. None of this
`
`information was publicly available, and it was only made available to Zlateski through his role as
`
`Technology Director of Neural Magic.
`
`D.
`
`26.
`
`Zlateski Leaves Neural Magic to Work for Facebook
`
` After almost a year and a half at Neural Magic, Zlateski informed Neural Magic
`
`that he had received a new position at Facebook to work on technology that Zlateski assured
`
`Neural Magic would not be related to his work at Neural Magic.
`
`
`
`
`- 8 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 9 of 24
`
`
`
`27.
`
`Given the supposedly different nature of Zlateski’s role at Facebook, Neural
`
`Magic presumed Zlateski did not plan to violate his non-compete obligation. Neural Magic
`
`nonetheless reminded Zlateski of the need to keep all Neural Magic proprietary information and
`
`trade secrets strictly confidential.
`
`E.
`
`28.
`
`Zlateski and Facebook Publish Neural Magic’s Secret Code to the World
`
`In November 2019, Facebook⎯with no notice to Neural Magic⎯published code
`
`(on GitHub, an open-source software development platform that anyone can access online for
`
`free) disclosing the Neural Magic Algorithms.
`
`29.
`
`A month later, and once again unknown to Neural Magic at the time, at an AI
`
`Developers Conference (the TVM conference in Seattle), Facebook announced it had disclosed
`
`to the whole world—and made open source—an algorithm that would enable neural networks to
`
`run efficiently on commodity Central Processing Units (CPUs); later investigation by Neural
`
`Magic would reveal, however, that its open source implements the very same algorithms as the
`
`Neural Magic Algorithms.
`
`30.
`
`At that time, and once again unknown to Neural Magic at the time, Facebook
`
`employee Bram Wasti described the sudden increase in speed and performance in Facebook’s
`
`software, and said it was thanks to “some work done by Aleks Zi,” using shorthand for Zlateski’s
`
`last name. He explained that this code significantly sped up Facebook’s most popular neural
`
`network executions.
`
`31.
`
`On or around January 18, 2020, a LinkedIn post from Facebook employee
`
`Jongsoo Park alerted Dr. Matveev at Neural Magic to the fact that Facebook and Zlateski had
`
`misappropriated the Neural Magic Algorithms. Notes accompanying Mr. Park’s LinkedIn post
`
`credited Zlateski for his contributions “to sparse kernels”—a high-level reference to work
`
`
`
`
`- 9 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 10 of 24
`
`
`
`Zlateski had performed in developing Neural Magic’s Algorithms. This area was completely
`
`outside the work that Zlateski said he would be pursuing at Facebook.
`
`32.
`
`Upon subsequently reviewing the open source code that Facebook published on
`
`GitHub, Neural Magic realized⎯for the first time⎯that Zlateski had misappropriated the Neural
`
`Magic Algorithms and given them to Facebook. And Facebook had published these algorithms
`
`to the world in the form of its open source compiler that implements (and embodies) the Neural
`
`Magic Algorithms. Facebook referred to its compiler as the “Sparse GEMM JIT”—but this code
`
`in pertinent part implements the Neural Magic Algorithms.
`
`33.
`
`GitHub maintains the entire modification history of every file posted on its
`
`website. The GitHub modification history for the Sparse GEMM JIT shows that the initial
`
`submitter is the lead of the FBGEMM software package at Facebook, Jongsoo Park, and Zlateski
`
`is listed as the first “reviewer” of the code.
`
`34.
`
`The code and compiler Facebook posted to GitHub implement the same Neural
`
`Magic Algorithms used in Neural Magic’s compiler code, to achieve the same computational and
`
`storage efficiencies running on commodity hardware (CPUs). Indeed, Neural Magic has tested
`
`the Facebook compiler side-by-side against its compiler, and the results from this direct
`
`comparison establish that the algorithms implemented in the Facebook compiler are the Neural
`
`Magic Algorithms. As explained below, Neural Magic subsequently informed Facebook of this
`
`testing and these test results, but Facebook has not changed course, forcing Neural Magic to take
`
`the present action.
`
`
`
`
`- 10 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 11 of 24
`
`
`
`F.
`
`Facebook Refuses to Remove Neural Magic’s Compiler from GitHub
`
`35.
`
`Despite direct and conclusive evidence that Facebook’s newly published compiler
`
`uses Neural Magic’s Algorithms, Facebook has refused to remove from GitHub the compiler
`
`containing Neural Magic’s trade secrets.
`
`36.
`
`On January 22, 2020, counsel for Neural Magic sent letters to both Facebook and
`
`Zlateski identifying the offending GitHub publication and asking for its removal pending
`
`investigation and further discussion between Neural Magic, Facebook, and Zlateski.
`
`37.
`
`In a series of letters, counsel for Facebook and Zlateski flatly refused to take
`
`down the code or agree to cease further use of Neural Magic’s proprietary and confidential
`
`information that Zlateski misappropriated as a Facebook employee.
`
`38.
`
`Neural Magic can only see what Facebook has already published to GitHub,
`
`which includes further Neural Magic technology in addition to the said compiler. Zlateski could
`
`disclose (and on information and belief has disclosed) even more to Facebook, for use internally
`
`or in future open source releases. Those continued disclosures would further damage Neural
`
`Magic’s business and market opportunities. Trade secrets at risk of further disclosure (perhaps
`
`already disclosed) include trade secret techniques used by Neural Magic to run fully connected
`
`neural networks and convolutional neural networks at speed and to speed up the training of such
`
`networks.
`
`
`
`
`
`
`
`
`- 11 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 12 of 24
`
`
`
`COUNT I
`
`(Misappropriation of Trade Secrets and Confidential Information (M.G.L. c. 93 § 42))
`(Facebook, Inc., Aleksandar Zlateski)
`
`The allegations contained in the above paragraphs are hereby incorporated by
`
`39.
`
`reference as if fully set forth herein.
`
`40.
`
`Aleksandar Zlateski was Technology Director at Neural Magic and is a former
`
`employee of Neural Magic. He was and is under a contractual obligation not to use or disclose
`
`Neural Magic’s confidential or proprietary information, including all trade secrets. As
`
`Technology Director, Zlateski had access to and did access Neural Magic’s trade secrets and
`
`confidential proprietary information, including, most critically for present purposes, Neural
`
`Magic’s Algorithms. As noted above, the Neural Magic Algorithms comprise performance
`
`engineering techniques reflected in Neural Magic’s unpublished compiler to improve efficiency;
`
`as well as algorithms, design patterns, optimization strategies, and formulas developed by Neural
`
`Magic in the course of developing compiler code and related code for various machine
`
`architectures, including code for the AVX2 and AVX512 instruction sets. This information was
`
`held in strict confidence and confidentiality by Neural Magic and was only made available to
`
`Zlateski in his capacity as Neural Magic’s Technology Director. Prior to the acts complained of
`
`herein, Neural Magic’s Algorithms⎯its trade secrets⎯were not generally known to others in the
`
`industry. Further, Neural Magic does not publish or disclose these algorithms, and Neural Magic
`
`products or services that implement the Neural Magic Algorithms do so without the algorithms
`
`themselves being made public or otherwise generally available.
`
`41.
`
`Consistent with standard industry practices, Neural Magic took reasonable and
`
`appropriate measures to protect and maintain its trade secrets, including, but not limited to,
`
`having its employees and any entities or individuals that performed work on its behalf sign
`
`
`
`
`- 12 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 13 of 24
`
`
`
`confidentiality and non-disclosure agreements prohibiting use and/or removal of any materials
`
`containing confidential, proprietary information from its premises except in the pursuit of Neural
`
`Magic’s business, limiting access to its offices, limiting access to its computer systems, and
`
`configuring computers to use high levels of encryption.
`
`42.
`
`Neural Magic has expended significant resources to develop its trade secrets and
`
`other confidential and proprietary information, to offer a breakthrough compiler that⎯by
`
`implementing the Neural Magic Algorithms⎯allows users to build and exploit their neural
`
`networks on commodity CPUs as opposed to specialized, hard-to-use, and expensive GPU
`
`hardware. Neural Magic’s trade secrets and confidential and proprietary information derive
`
`independent economic value, actual or potential, from not being generally known to, and not
`
`being readily ascertainable through proper means by, another person who can obtain economic
`
`value from the disclosure or use of the information. These trade secrets and confidential,
`
`proprietary information are highly valuable to Neural Magic and to others who do not know or
`
`have access to them.
`
`43.
`
`As just one example of the value of Neural Magic’s trade secrets, Intel recently
`
`purchased the Israel artificial intelligence company Habana Labs for $2 billion on the promise of
`
`performance speed upgrades for machine learning that provide 2-3x over Nvidia’s GPUs. The
`
`Neural Magic Algorithms at
`
`issue herein⎯when
`
`implemented
`
`in
`
`the Neural Magic
`
`compiler⎯offer comparable performance enhancements on CPUs. Unlike these accelerators,
`
`Neural Magic’s CPU solution will deliver speedups while eliminating the severe memory
`
`constraints associated with these devices and enable neural networks to run anywhere, from
`
`laptops to servers, not just in large data centers.
`
`
`
`
`- 13 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 14 of 24
`
`
`
`44.
`
`Aleksandar Zlateski misappropriated or unlawfully took, carried away, concealed,
`
`used and/or copied trade secrets and other confidential proprietary information from Neural
`
`Magic and disclosed Neural Magic’s trade secrets and other confidential proprietary information
`
`to Facebook, GitHub, and GitHub’s users in direct violation of Zlateski’s confidentiality
`
`agreement with Neural Magic.
`
`45.
`
`On information and belief, Facebook knew, or should have known, prior to the
`
`open source release of the Facebook Sparse GEMM JIT, that Zlateski’s work on that compiler
`
`(directly or indirectly) was a misappropriation of Neural Magic’s Algorithms. Indeed,
`
`Facebook’s steadfast refusal to take down the offending compiler code (which embodies and
`
`implements the Neural Magic Algorithms), even when faced with proof of this theft,
`
`demonstrates not only the value to Facebook of the Neural Magic Algorithms– it also
`
`demonstrates Facebook’s blatant disregard of Neural Magic’s work and valued intellectual
`
`property.
`
`46.
`
`On information and belief, Facebook, Inc. knowingly received the benefits from
`
`the disclosure and/or use of Neural Magic’s trade secrets. Aleksandar Zlateski and Facebook,
`
`alone or in concert, used improper means, in breach of Aleksandar Zlateski’s contractual
`
`obligations to Neural Magic, to acquire Neural Magic’s trade secrets and other confidential,
`
`proprietary information.
`
`47.
`
`As a direct and proximate result of Zlateski’s and Facebook’s misappropriation of
`
`trade secrets and other confidential and proprietary information, Neural Magic has suffered and
`
`will continue to suffer irreparable harm and other damages, including, but not limited to, loss of
`
`value of its trade secrets. Neural Magic is entitled to injunctive relief and monetary damages,
`
`including enhanced damages pursuant to the statute.
`
`
`
`
`- 14 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 15 of 24
`
`
`
`COUNT II
`
`(Violation of the Defend Trade Secrets Act of 2016 (18. U.S.C. § 1836))
`(Facebook, Inc., Aleksandar Zlateski.)
`
`The allegations contained in the above paragraphs are hereby incorporated by
`
`48.
`
`reference as if fully set forth herein.
`
`49.
`
`Neural Magic is the owner of trade secrets that Aleksandar Zlateski and Facebook
`
`misappropriated and are now using and disclosing. These trade secrets are used in, or intended
`
`for use in, interstate or foreign commerce.
`
`50.
`
`As Technology Director and a former employee of Neural Magic, Aleksandar
`
`Zlateski was under a contractual obligation not to use or disclose Neural Magic’s confidential or
`
`proprietary information, including all trade secrets.
`
`51.
`
`Neural Magic took reasonable measures to protect and maintain the secrecy of its
`
`trade secrets and confidential and proprietary information. These protections include, but are not
`
`limited to, having its employees and any entities or individuals that performed work on its behalf
`
`sign confidentiality and non-disclosure agreements prohibiting removal of any materials
`
`containing confidential, proprietary information from its premises except in the pursuit of Neural
`
`Magic’s business, limiting access to its offices, encrypting forms of electronic storage, and
`
`limiting access to its computer systems through a variety of mechanisms, including mechanisms
`
`discussed herein.
`
`52.
`
`Due to these security and confidentiality measures, Neural Magic’s confidential
`
`and proprietary trade secret information is not available for others in the field of neural networks
`
`and CPU optimization⎯or any other field⎯to use through any legitimate means.
`
`53.
`
`Neural Magic has expended significant resources to develop its trade secrets and
`
`other confidential and proprietary information, to offer a breakthrough compiler that allows users
`
`
`
`
`- 15 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 16 of 24
`
`
`
`to build neural networks that more efficiently utilize CPUs. Neural Magic’s trade secrets and
`
`confidential and proprietary information, derive independent economic value, actual or potential,
`
`from not being generally known to, and not being readily ascertainable through proper means by,
`
`another person who can obtain economic value from the disclosure or use of the information.
`
`These trade secrets and confidential, proprietary information are highly valuable to Neural Magic
`
`and to any other person or entity that wants to enter the field of artificial intelligence.
`
`54.
`
`Aleksandar Zlateski and Facebook knew, or had reason to know, that the trade
`
`secrets at issue herein had been acquired from Neural Magic through improper means including,
`
`without limitation, Zlateski’s breach of his non-disclosure obligations. Zlateski and Facebook
`
`disclosed Neural Magic’s trade secrets to GitHub and the public at large in direct violation of
`
`Zlateski’s contractual obligations to Neural Magic.
`
`55.
`
`On information and belief, Facebook, Inc. knowingly received the benefits from
`
`the disclosure and/or use of Neural Magic’s proprietary information. Aleksandar Zlateski and
`
`Neural Magic thus in concert used improper means, in breach of Aleksandar Zlateski’s
`
`contractual obligations to Neural Magic, to acquire Neural Magic’s trade secrets and other
`
`confidential, proprietary information.
`
`56.
`
`On information and belief, if Defendants are not enjoined, Defendants will
`
`continue to misappropriate and use Neural Magic’s trade secrets for their own benefit and to
`
`Neural Magic’s detriment.
`
`57.
`
`As a direct and proximate result of Zlateski’s and Facebook’s misappropriation of
`
`Neural Magic’s trade secrets and other confidential and proprietary information, Neural Magic
`
`has suffered and will continue to suffer irreparable harm and other damages, including but not
`
`limited to, loss of value of its trade secrets. Neural Magic is therefore entitled to civil seizure of
`
`
`
`
`- 16 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 17 of 24
`
`
`
`property, injunctive relief, monetary damages for its actual losses, and monetary damages for
`
`unjust enrichment where damages for its actual losses are not adequately addressed.
`
`COUNT III
`
`(Unfair Business Methods (M.G.L. c. 93A § 11))
`(Facebook, Inc., Aleksandar Zlateski)
`
`The allegations contained in the above paragraphs are hereby incorporated by
`
`58.
`
`reference as if fully set forth herein.
`
`59.
`
`At all times relevant to this action, Neural Magic has been engaged in trade or
`
`commerce within the meaning of M.G.L. c. 93A, § 11.
`
`60.
`
`Aleksandar Zlateski and Facebook, Inc. engaged in a course of conduct designed
`
`to unfairly harm Neural Magic, to Facebook’s advantage, through their business transactions
`
`with Neural Magic. Zlateski’s and Facebook’s unfair conduct included, but is not limited to
`
`stealing Neural Magic’s trade secrets—algorithms in Neural Magic’s GEMM JIT source code,
`
`including performance engineering techniques reflected in that source code to improve
`
`efficiency; as well as algorithms, design patterns, optimization strategies, and formulas
`
`developed by Neural Magic in the course of developing compiler code and related code for
`
`various machine architectures, including code for the AVX2 and AVX512 instruction sets—and
`
`publishing them publicly and for free. These acts and practices of Zlateski and Facebook
`
`constitute unfair methods of competition or unfair or deceptive acts and practices and business
`
`transactions that occurred primarily and substantially within Massachusetts within the meaning
`
`of M.G.L. c. 93A § 2.
`
`61.
`
`These unfair methods of competition or unfair or deceptive acts or practices were
`
`intentional, willful, and knowing.
`
`
`
`
`- 17 -
`
`

`

`Case 1:20-cv-10444 Document 1 Filed 03/04/20 Page 18 of 24
`
`
`
`62.
`
`As a direct and proximate result of Zlateski’s and Facebook’s unfair and
`
`deceptive acts and conduct as aforesaid, Neural Magic has suffered and will continue to suffer
`
`substantial and irreparable harm and other damages, including, but not limited to, loss of value of
`
`trade secrets, loss of custo