This week, Google released a family of AI Open, GMMA 3 models, which quickly acquired its impressive efficiency. But as number to Developers Gemma 3 license, which unfortunately dates back to X, commercial use of models is a risky proposal.
It is not a unique problem for GEMMA 3. Companies such as Meta also apply the non -standard licensing conditions to their models openly, and the conditions for companies ’legal challenges. Some companies, especially smaller operations, are concerned that Google and others can “withdraw the carpet” on their business by confirming the most connected sentences.
“The restricted and inconsistent license of the” Open “models of artificial intelligence creates great uncertainty, especially for commercial adoption,” Nick Vidal, the head of society in the open source initiative, Long -term institution It aims to determine and “host” all open source things, tell Techcrunch. “While these models are marketed as open, actual terms impose many legal and practical obstacles that companies prevent them from integrating them into their products or services.”
Open models developers have the reasons for launching models under ownership licenses instead of industry greeting options Apache and MIT. Start AI, for example, Be About her intention to support scientific work – but not commercial – at the top of its models.
But the Llama and Meta licenses in particular have restrictions that limit the methods that companies can use for models without fear of legal revenge.
Definition, for example, The developers are prohibited From the use of “directing or results” Llama 3 models to improve any model besides Llama 3 or “derived works”. It also prevents companies that exceed more than 700 million active users per month from spreading Llama models without obtaining an additional private license first.
Gima license Generally less tense. But it gives Google the right to “restrict use (remote or otherwise)” GEMMA), which Google believes is violating the company Obligated use policy Or “applicable laws and regulations.”
These conditions are not only applied to the original Llama and Gemma models. Llama or Gemma models should also abide by Llama and Gemma licenses, respectively. In the case of GEMMA, this includes models trained on artificial data created by GEMMA.
Florian Brand, a researcher at the German AI Research Center, believes that – though – though What will you believe in giants in technology? Licenses such as GEMMA and Llama “cannot be reasontly called” open source “.
“Most companies have a set of approved licenses, such as APache 2.0, so any dedicated license is a lot of trouble and money,” Brand told Techcrunch. “Small companies, without legal difference or money for lawyers, will adhere to standard licensing models.”
Brand noted that the developers of artificial intelligence models who have custom licenses, such as Google, have not imposed strongly their conditions yet. However, the threat is often sufficient to deter adoption.
“These restrictions have an impact on the ecosystems of Amnesty International – even on active intelligence researchers,” said Brand.
Han-Cung Lee, Moody’s Automated Manager, agrees that allocated licenses such as GEMMA and Llama make models “unused” in many commercial scenarios. The same applies to Eric Tramil, the world of the employee dish at Ai Startup Gretel.
“Licenses for models make specific points for typical derivatives and distillation, causing anxiety about the klawak,” said Tameel. “Imagine a work that specifically produces the model ranges of its customers. What is the license that GEMMA-Data should have the title of Llama? What will the effect of all of their customers in the direction of the river course?”
Tramil said that the scenario, which is more afraid of publishing men, said that the models are a Trojan horse of some kind.
He said: “The Mabek Mabka group can take out models (open), and wait for the work of business cases that develop using these models, and then transfer them to successful spirits through extortion or lawfare.” “For example, GEMMA 3, in all manifestations, looks a strong version – which can have a wide impact. But the market cannot adopt because of the license structure. So, companies are likely to adhere to Apache 2.0 models weaker and less reliable.”
To be clear, some models have made a large -scale distribution despite their restricted licenses. Lama, for example, was Hundreds of millions of times were downloaded Anti -products from major companies, including Spotify.
But they could be more successful if they are constantly licensed, according to Yacine Jernite, the head of automatic learning and community at AI Startup Luging Face. Jernite has called on service providers like Google to move to open licensing frameworks and “directly cooperate” with users on widely acceptable conditions.
“Given the lack of consensus on these conditions and the fact that many basic assumptions have not yet been tested in the courts, they work primarily as a declaration of intention from these actors.” “(But if some sentences are interpreted) on a very large scale, a lot of good work will find itself on unconfirmed legal land, which is especially frightening for organizations to build successful commercial products.”
Vidal said that there is an urgent need for artificial intelligence models that can integrate, modify and participate without fear of sudden licensing changes or legal ambiguity.
Vidal said: “The current scene of licensing the artificial intelligence model is full of confusion, restrictions, and misleading demands for openness.” “Instead of redefining” open “to suit the interests of companies, the artificial intelligence industry must be consistent with open source principles in force to create a truly open ecosystem.”
https://techcrunch.com/wp-content/uploads/2024/01/GettyImages-1548038240.jpg?resize=1200,814
Source link