OpenAI is addressing concerns about ChatGPT’s perceived “laziness” by introducing an enhanced model, aiming to rejuvenate its AI’s performance and responsiveness after complains that it was refusing to do its job.
“Today, we are releasing an updated GPT-4 Turbo preview model,” OpenAI announced. “This model completes tasks like code generation more thoroughly than the previous preview model and is intended to reduce cases of ‘laziness’ where the model doesn’t complete a task.”
People had been complaining across social media that ChatGPT was giving predetermined answers, refusing to do its job, telling people to do the task themselves, or simply halted its thinking process when asked to do certain things, mainly coding.
Copilot pro is completely useless today, I’m not sure if there was some update behind the scenes but it has adopted the recent behaviour of ChatGPT and it’s getting lazy or just straight up refusing to do the simplest tasks that did fine yesterday.@JordiRib1 @MParakhin 👀 pic.twitter.com/5B5d5t7oXN
— Angel (@Angel_pap85) January 29, 2024
OpenAI’s latest update—named gpt-4-0125-preview —focuses on code generation capabilities and counteracting the AI’s previous inclination to underperform. The core of this upgrade lies in the incorporation of “embeddings,” a sophisticated AI feature designed to enhance the model’s understanding and output.
Embeddings are essentially sequences of numbers representing the concepts within content, such as natural language or code. By parsing the relationships between content, embeddings facilitate complex tasks like clustering or retrieval, thereby powering applications like knowledge retrieval in ChatGPT and the Assistants API.
Embeddings guide the generative process of the LLMs, conditioning it to follow a certain path during its reasoning as they capture specific data over a relevant object, as explained by Cloudflare. The LLMs now have little pieces of code “embedded” into it to make it better.
OpenAI introduced 3 embeddings that helped GPT increase its performance. And of all the embeddings, one named text-embedding-3-large is the company’s new best-performing creation.
Average benchmark scores rose from 31% to 54% to 61 to 64%, the company claims.
Some users are already reporting good results from the update. Dwayne Charrington, a core team member of Aurelia.io, said he was pretty satisfied with the update. “I don’t know what you did, but you fixed it. You still need to prompt properly, but you fixed the issue.” he said on a Twitter post.
Tech, APIs, and education
OpenAI also announced an update to its GPT-3.5 Turbo model and a significant price reduction today.
The third-generation chatbot received enhancements relating to the API and its responsiveness and an overall issue regarding non-English language function calls.
By cutting its input prices by 50% and output prices cut by 25%, OpenAI hopes its model more attractive to users that find themselves deciding between ChatGPT and Claude. Users of the earlier gpt-3.5-turbo model will experience a seamless transition to the improved model without having to do anything else.
OpenAI also announced a collaboration with Common Sense Media, which advocates for the safety and empowerment of children and families in the digital age. This collaboration focuses on crafting guidelines and educational materials tailored for parents, educators, and young peopl. Finally, OpenAI says it will also join forces to curate a collection of family-friendly GPTs in the GPT Store, adhering to Common Sense’s ratings and standards.
“AI offers incredible benefits for families and teens, and our partnership with Common Sense will further strengthen our safety work, ensuring that families and teens can use our tools with confidence,” Sam Altman, CEO of OpenAI, said in an official press release.
