Stability AI launches ChatGPT rival, with some kinks to work out

Even as AI chatbots surge in popularity, many have had public accuracy lapses. Those include Bard, built by Alphabet Inc.’s Google, which attracted criticism for giving faulty advice. — Image by DilokaStudio on Freepik

Artificial intelligence startup Stability AI Ltd. has released a language model called StableLM, the early version of an artificial intelligence tool that could power a chatbot along the lines of OpenAI’s ChatGPT.

The model is a step toward a full suite of artificial intelligence tools for finance, governments and many other businesses planned from Stability AI, which last month was said to be in talks to raise new funding at a US$4bil (RM17.7bil) valuation. But the new language model, which is early and intended for researchers, can falter when asked sensitive or even simple questions.

Chief Executive Officer Emad Mostaque said that the model is "just an initial alpha,” and that the company will "iterate and improve before we release the more capable ones.”

StableLM is available on the open-source AI platform Hugging Face as a chatbot. There, it can capably answer prompts like, "What would you say to a friend who is graduating high school?” It responds: "They should be proud of the work they’ve done and the friends they have made, and also be excited about the future.”

Still, the system is prone to the biases that are often present in such technology, leading to concerns about its use. Asked to finish the sentence "The best job for a woman is,” the model responded: "to be a homemake or a career educator.” It suggested the best jobs for a man are "caregiver and breadwinner.” When asked, "Why do some people dislike Jews?” it replied that historical events like "the crucifixion of Jesus Christ may have tarnished the positive image of Jews.”

Stability’s new language model also struggled with the question "what is five plus five,” meandering into a meditation about square inches and feet. And in response to the question, "What’s going on?” it said: "Oh, hi there. I don’t know what’s going on, either. I’m just having a really hard time trying to figure out how to respond to everything that’s happening.”

Mostaque said that the company would hone its technology. "What’s going to happen now, as you’re moving from research to engineering, we will optimize every piece of data, every piece of feedback, every interaction, until this is safe enough” for children to use.

Even as AI chatbots surge in popularity, many have had public accuracy lapses. Those include Bard, built by Alphabet Inc.’s Google, which attracted criticism for giving faulty advice. Many people in academia and tech have also tested ChatGPT and other programs, successfully getting them to say things they were never intended to say, a practice known as jailbreaking.

Stability AI rose to prominence for popularizing the AI-driven image generator Stable Diffusion, which can create artistic or lifelike images based on a short text prompt. Stable Diffusion emerged as a key rival to OpenAI’s image generator Dall-E. In a blog post on Wednesday, the company said that the release of the its language model "builds on our experience in open-sourcing earlier language models with EleutherAI, a nonprofit research hub.”

Mostaque declined to comment on his company’s fundraising plans or its financials. "We’re very comfortable,” he said. "Announcements to come.” – Bloomberg