Tom Goldstein, a computer science professor at the University of Maryland, discusses the advantages and disadvantages of open-source generative AI models and why open approaches are less likely to become widespread. A new language model called LLaMa 2, developed by Meta (Facebook's parent company), is poised to challenge OpenAI's ChatGPT. LLaMa 2 is capable of competing with other well-known models but sets itself apart by being open source, meaning its source code is publicly available for study and further development at no cost. While companies like OpenAI and Google prefer to keep their latest models proprietary, an open approach offers benefits for research and enterprise and can also be advantageous for these companies themselves. Professor Goldstein explains that open source helps companies enhance their public image and reputation by showcasing their research and contributing to the academic community. This perception of innovation and technological advancement attracts talented individuals who are more likely to join these companies. On the other hand, there are risks associated with open source, particularly concerning the potential for misuse of language models. Malicious uses range from generating spam and fake reviews to engaging in sophisticated cybercrimes like spear-phishing and spreading conspiracy theories. While proprietary models can also be used for these purposes, open-source tools make it easier for individuals to run these models on their own computers, making moderation more challenging. However, there are advantages to open-source approaches as well. The academic community relies on open-source language models for research, and there is a demand for creative applications that arise when people have access to these tools.

The transparency of open-source models allows for greater scrutiny and scientific applications. It is worth noting that the transparency of models may vary, with some models being more closed about their training process, possibly due to concerns about liability. Regarding the future of AI models, Professor Goldstein believes that the number of model releases will decrease as the bar for quality continues to rise. Models like LLaMa 2 represent significant efforts that will be challenging to surpass, even with industrial resources or within the open-source community. He predicts that closed-source models will emerge more frequently, not only from large American companies but also from data providers in Asia. While open-source models will still have a role, their rate of development may slow down. It is important to understand that the AI model landscape is not binary in terms of openness. Popular models like ChatGPT and Midjourney fall somewhere in the middle, offering accessibility to the public while limiting outsiders' ability to study their workings. LLaMa 2, while further down the openness axis, still maintains some level of opacity by not disclosing its training data. The debate around model openness extends to legal considerations, like fair use and copyright. The European Union is considering regulations that may require more transparency in training data disclosure, but companies like OpenAI have successfully lobbied to soften such requirements. "Marketplace Tech" is a platform that examines the digital economy and explores topics beyond just Big Tech, including technology's intersection with climate change, inequality, and disinformation.