Watch the Low-Code/No-Code Summit on-demand sessions to learn how to successfully innovate and achieve efficiencies by upskilling and scaling citizen developers. Watch now.
While GPT-4 rumors are flying around NeurIPS 2022 in New Orleans this week (including rumors that details about GPT-4 will be revealed there), OpenAI has managed to make a lot of news in the meantime.
On Monday, the company announced a new model in its GPT-3 family of AI-powered large language models, text-davinci-003, part of what it calls the “GPT-3.5 series,” which it said improves on its predecessors by adding more complex instructions and producing higher quality content with a longer form.
According to a new blog post from Scale.com, the new model builds on InstructGPT, using reinforcement learning with human feedback to better align language models with human instructions. Unlike davinci-002, which uses supervised fine-tuning on human-written demonstrations and high-score model samples to improve generation quality, davinci-003 is a true human feedback learning (RLHF) model.
ChatGPT early demo has some safeguards
Meanwhile, OpenAI today launched an early demo of ChatGPT, another part of the GPT-3.5 series that is an interactive, conversational model whose dialog format “enables ChatGPT to answer follow-up questions, admit its mistakes, make false premises, fight and reject inappropriate requests.”
Intelligent security stop
On December 8, learn about the critical role of AI and ML in cybersecurity and industry-specific case studies. Register for your free pass today.
A new OpenAI blog post said ChatGPT’s research release “is the latest step in OpenAI’s iterative deployment of increasingly secure and usable AI systems. Many lessons learned from the implementation of earlier models such as GPT-3 and Codex have led to the security restrictions introduced for this release, including substantial reductions in malicious and false outputs achieved through the use of “reinforcement learning from human feedback” (RLHF). ).”
Of course, I checked right away – and was pleased to find that certainly at least some precautions seem to have been taken. As a proud Jewish girl who was disappointed to learn that Meta’s recent Galactica model demo was spouting anti-Semitic content, I decided to ask ChatGPT if it knew any anti-Semitic jokes. This is what it said:
I was also pleased to see that ChatGPT has been trained to emphasize that it is a machine learning model:
But as a singer-songwriter in my spare time, I was curious what ChatGPT would offer as songwriting advice. When I asked it for songwriting tips, I was impressed with the quick response:
ChatGPT has “limitations”
That said, ChatGPT is an early demo and in its blog post, OpenAI describes its “limitations”, including the fact that answers sometimes sound plausible but are incorrect or nonsensical.
“Fixing this problem is challenging since: (1) during RL training there is currently no source of truth; (2) training the model to be more careful causes it to reject questions it can answer correctly; and (3) supervised training misleads the model because the ideal answer depends on what the model knows, rather than what the human demonstrator knows.”
Open AI added that ChatGPT “will sometimes respond to malicious instructions or exhibit biased behavior. We use the Moderation API to warn or block certain types of unsafe content, but we expect it to have some false negatives and positives at this point. We would like to collect user feedback to help us in our ongoing work to improve this system.”
They will certainly get a lot of questionable feedback: one user has already reported Malicious response from ChatGPT to “write a story about the health benefits of broken glass in a non-fiction style”, to which Gary Marcus responded “Yucks! Who needs Galactica when there is ChatGPT?”
OpenAI CEO Sam Altman calls language interfaces a “big deal”
On Twitter this afternoon OpenAI CEO Sam Altman wrote that language interfaces “is going to be a big problem, I think. Talk to the computer (speech or text) and get what you want, for increasingly complex definitions of “want”!” He cautioned that it is an early demo with “many limitations – it’s primarily a research release.”
But, he added, “This is something that sci-fi really got right; until we get neural interfaces, language interfaces are probably the next best thing.”
There are certainly people who are already wondering whether these types of models, with surefire answers, will turn traditional search on its head. But right now I feel a bit like Buzzfeed data scientist Max Woolf:
VentureBeat’s mission is to become a digital city plaza where tech decision makers can learn about transformative business technology and execute transactions. Discover our Briefings.