Open source AI outperforms private models like Chat-GPT.
Generative artificial intelligence (AI) models based on centralized cloud infrastructure like ChatGPT currently lead overall performance, but new research shows that open source competitors are catching up.
Current generative AI market leaders, such as Google and OpenAI, have taken a centralized approach to building their infrastructure, effectively limiting public access to a variety of information, including the data sources used for training models.
Cathy Wood’s ARK Invest research team argues that this could change, suggesting that open source AI models will likely surpass centralized AI models by 2024.
The graph above shows the progress of open source AI models since 2022, some of which have performed better than private models. OpenAI, Google, and parent company Alphabet dominate the centralized AI space with popular models such as ChatGPT-4 and Gemini Ultra. On the other hand, Meta (formerly Facebook), Mistral, and some Chinese AI models have chosen an open source approach.
Please confirm @ARKInvestOpen source and closed performance updates #AI Model. The open source LLM model is still gaining ground, despite increasing dispersion. thank you @downingARK and @JozefARK Thank you for keeping us updated on the most important technology race in history! https://t.co/hiikHFJYwr
— Cathie Wood (@CathieDWood) December 14, 2023
In 2023, Yi 34B, Falcon 180B, and Mixtral 8x7B emerged as some of the best open source AIs with similar performance to market leaders. ARK Invest researcher Jozef Soja noted that Mixtral surpassed GPT 3.5 in the absolute log error of the Massive Multitask Language Understanding (MMLU) benchmarking performance, highlighting “how far ahead GPT-4 is today.”
Additionally, Meta’s base model, LLaMA, also recorded significant improvements in its 2023 iteration. AI engineer Brian Roemmele said the need for performance in open source AI models stems from the importance of democratizing access to generative AI.
Related: ChatGPT passed the neurology exam for the first time.
Current AI market leaders, Google’s Gemini and OpenAI’s ChatGPT-4, were recently tested. Cointelegraph asked the free versions of Gemini (via Bard) and ChatGPT-4 some questions about cryptocurrency and compared their answers.
In both cases, the AI took a similar approach, alerting users when asked about investment advice and encouraging them to seek out a professional financial advisor for more “personalized advice.” Additionally, both AI models provided answers that highlighted similar considerations, despite providing different details in some cases.
magazine: Weaponized Terrorism and Israel-Gaza War to Destroy Cryptocurrencies