Open-source LLMs have taken the world by storm in just a little over 2 months, ever since LLaMA’s weights were made available for anyone to tinker and play with. Just less than 2 weeks after the untrained LLaMA model was released by Meta.
💡 A model’s weights are the values set to each parameter after training the model on a dataset, with the parameters being various factors (such as token size, number of layers) that allow the model to give more complex answers to what’s input by the user.
This led to a flurry of advancements from dedicated open-source community members. Through just the use of their personal hardware, they were able to make leaps and bounds in their quest to place the most powerful AI in the hands of everyday people.
👨💻 Recommended: A Quick and Dirty Dip Into Cutting-Edge Open-Source LLM Research
OpenAI’s leadership seems to have taken quite a notice of these events, because they seem to be planning to release an open-source LLM, according to a report by Reuters. It’s virtually unanimous that OpenAI’s GPT-4 is the best-performing LLM model out there. So an open-source model from them would be no small event, even if it is weaker than GPT-4.
Finding out exactly how an OpenAI foundation model is built would give the open-source community a wealth of knowledge that they can apply to their other projects.
It would also go to show how seriously OpenAI views open source and the community surrounding it. It would show that they’re fully aware that the only chance for them to maintain their LLM dominance is if they allowed the world to improve and iterate on their designs.
Open-source showing such swift and definite progress toward taking the crown away from OpenAI can hardly be a surprise. The law of the wisdom of the crowd was foretelling of that. The insight and understanding of the relative few can never match the capability of the collective knowledge and experience of the tens of millions.
The Ultimate Open-Source LLM Battle – Who Wins?
In a chatbot arena site managed by LYMSYS, visitors are asked to enter a prompt, and two randomly-selected models will each provide a response.
The model that the user chooses as having given the best response is then raised up on the leaderboard while the other gets lowered.
The following models are the top three highest-performing models in that arena, just behind GPT-4 (ELO rating of 1274), Anthropic’s Claude (rating of 1224), and GPT-3.5-Turbo (rating of 1155).
Trained by LYMSYS, an open research organization based in UC Berkeley, it is the most promising model from the LLaMA leak.
💡 Recommended: 11 Best ChatGPT Alternatives
It reportedly achieves 90% response quality compared to ChatGPT and Google’s Bard, using a casual evaluation method done through GPT-4. They were able to accomplish this with just a training cost of $300. It has a rating of 1083.
Coming from BAIR, another group within UC Berkeley, this is a dialogue model meant for academic research. It aims to answer the question of whether open-source models can overcome the massive scale advantage of closed models through better curation of training data. It comes in with a rating of 1022.
Impressively, this model was developed by a single person known by the username BlinkDL.
Even more impressively, it’s an RNN LLM (Recurrent Neural Network) rather than the ubiquitous Transformer LLM. The advent of Transformers is what led to the power of GPT-4 being achieved.
People like BlinkDL figuring out ways to optimize more archaic architectures could soon lead to a hybrid architecture that overtakes Transformers in both performance and speed. This model’s rating is a respectable 989.
Civilization-Defining Power Through Artificial General Intelligence
Open-source is a term that can bring out patronizing feelings in people because, after all, a lot of the best programs we know today are closed-source and are chosen by billions of people each year. But that is only due to there being no real reason for the wider community to develop superior open-source alternatives preferred by the wider public.
It’s a much different case with AI.
A few companies holding such immense and civilization-defining power for themselves is not a future that anyone who truly understands the capabilities of AI would want.
Artificial general intelligence is just around the corner, and with it, a complete reimagining of society as we know it. It is a tool that every single person should have equal access to. That reality would bring about a golden age that humanity has never before experienced in all its history.
No matter what anyone says, hoarding any AI knowledge for oneself is a complete disservice to the good of humanity.
Rather than being reserved for the privileged few, a world where AI can be developed and iterated upon by any and all is the only way any sort of utopia can be achieved. Through open-source AI, the dreams and optimism of some of our favorite sci-fi stories will finally be brought to life.