AI Weekly #17 OpenAI once again attracts the attention of the world
This week’s cover: @Cristopher
Prompt: Portrait of a girl with short brown hair, glasses and a blue jacket and pink shirt
OpenAI palace fight turmoil
The nearly week-long OpenAI palace battle came to an end with Sam Altman returning to OpenAI as CEO ( click here to view the timeline of events ). This is the most sensational “tiddy news” in the technology industry in recent years. It is full of speculation and doubts. What is the cause of the incident?
As more and more information comes to light, the truth behind Ultraman's dismissal is gradually emerging. OpenAI's internal AI project Q* (pronounced Q-Star) was found to potentially threaten humanity, and a letter about an "AI breakthrough called Q*" prompted the board to fire Altman, media reported, citing people familiar with the matter.
At present, the truth behind it is still unknown. Elon Musk also called in the past few days that given that the potential dangers of AI are so great, OpenAI should disclose to the public the reasons for its dismissal of Ultraman.
Anthropic launches Claude 2.1 model, supports 200K contexts and reduces hallucinations
Artificial intelligence startup Anthropic has announced its latest conversation model Claude 2.1, which doubles Claude's context length limit to 200,000 tokens and reduces the misstatement rate by 50%.
After the model was released, X user Greg Kamradt conducted a long-context accuracy test called "needle in the haystack" on GPT-4 (128K) and Claude 2.1 (200K). We experimented with the memory differences between the two AIs in the content at different locations in the document when receiving contexts of different lengths. (GPT-4 is still far ahead)
Claude 2.1 Easier to remember (regardless of length): the second half of the text.
Claude 2.1 is even harder to remember (90K long text): the first half of the text.
Claude 2.1 Nearly 100% memorized (regardless of length): beginning of text & end of text.
Less context = higher accuracy.
The tested API call cost approximately $1016.
Stablility.ai launches video generation model Stable Video Diffusion
Stable Diffusion developer Stablility.ai released Stable Video Diffusion, a basic video generation model based on the image model Stable Diffusion, on November 21. Together with the previously released AI audio generation model Stable Audio, Stability.AI has achieved almost full coverage of media content generation.
Stable Video Diffusion comes in two model forms: SVD and SVD-XT. SVD converts still images into 14-frame 576×1024 video, while SVD-XT increases the number of frames to 24. It is reported that Stability.AI was initially trained on a data set containing millions of videos, and was subsequently fine-tuned on smaller sets containing hundreds of thousands to a million clips. However, the exact origin of the training videos remains unclear, raising some concerns about copyright issues.
ChatGPT opens voice capabilities to all users
On November 22, OpenAI opened the ChatGPT Voice function to all users for free . The feature, previously only available to ChatGPT Plus subscribers, is now available to all free users. Users only need to open the app settings on Android or iOS and toggle the "headphones" icon to activate the text-to-speech feature.
OpenAI launched a text-to-speech feature based on ChatGPT in September this year, using a neural network model to generate human-like speech from text. The system is able to create custom voices with different accents and styles by sampling a few seconds of audio recorded by a voice actor.
Microsoft will open Copilt to users in mainland China on December 1
According to Microsoft, the Web AI chat function Copilot, formerly known as Bing Chat Enterprise, will be launched for enterprises and educational institutions in mainland China on December 1.
Copilot is based on the latest OpenAI models, including GPT-4 and DALL-E 3, and provides text and image generation capabilities. Microsoft does not have direct access to user data, nor is it used to train the underlying model .
Bill Gates: AI will make three-day work week possible
Bill Gates, co-founder of Microsoft and former richest man in the world, recently appeared on the podcast of well-known talk show host Trevor Noah. Gates believes that AI may not be able to replace humans, but it can make a three-day work week possible.
When Noah asked about the threat of AI to jobs, Gates said that one day, humans "won't have to work so hard." "If you end up in a society where you only have to work three days a week, that might not be so bad," Gates said. He believes that in the future, "machines could make all the food and stuff" and people wouldn't have to work five days a week. To make a living, the remaining time can be used for useful things. “I have a lot of free time, so I have to think about how to use it.”