The field of artificial intelligence (AI) is rapidly advancing, and there have been some exciting breakthroughs in recent months that are set to revolutionize the field. From new training techniques to innovative applications of AI in everyday life, these advancements promise to change the way we think about AI.
One of the most impressive breakthroughs is the development of a new training technique called Parameter Efficient Fine Tuning (PEFT). Using this technique, researchers at LLaMA-Adapter were able to introduce instruction tuning and multimodality into a model in just one hour of training. Even more impressive is the fact that the model achieved a new state-of-the-art (SOTA) result on multimodal ScienceQA using just 1.2M learnable parameters.
Another exciting development came from the University of California, Berkeley, which launched a new dialogue model called Koala. The model was trained entirely using freely available data, and the researchers took the crucial step of measuring real human preferences between their model and the widely used ChatGPT. The results were surprising: while ChatGPT still held a slight edge, more than 50% of the time users either preferred Koala or had no preference. What’s more, the training cost for Koala was just $100.
Open Assistant also made a significant contribution to the field with their launch of an open-source model and dataset for Alignment via Reinforcement Learning from Human Feedback (RLHF). Their model achieved a human preference score of 48.3%, which was very close to ChatGPT’s score of 51.7%. This dataset can also be applied to Pythia-12B, giving researchers the option to use a fully open stack to run the model. Because the dataset is publicly available, RLHF has become cheap and easy for small experimenters.
But as exciting as these developments are, there are still some concerns about the societal impact of AI. Sam Altman, CEO of OpenAI, asked the question of whether GPT-4 should be let off the leash. OpenAI is taking steps to limit the potential societal destruction and danger that could arise if AI falls into the wrong hands. This comes as some experts claim that open letters on AI risks lack clarity and sensationalize potential dangers.
In other news, Microsoft has launched an AI-powered Security Copilot for cybersecurity professionals. This tool enhances the quality of detection, speed of response, and ability to strengthen security postures. At the same time, leaked documents from Google AI reveal that the company believes they have no moat and neither does OpenAI, suggesting that competition in the field is fierce.
Finally, there has been a lot of buzz about a proposed ‘pause’ on AI development, which has been rejected by Bill Gates. While there are legitimate concerns about the impact of AI on society, experts argue that a pause would hinder progress and could ultimately do more harm than good.
In summary, the world of AI is rapidly advancing with groundbreaking developments, but we must approach it with both caution and enthusiasm. While acknowledging the legitimate concerns about its societal impact, we should work towards ensuring that AI is developed in a responsible and beneficial way for all.
vikrambhardwaj 