Wow, what a week! Numerous AI companies released major products. Google announced that it would allow developers to access its AI language models. Anthropic, an AI startup, also revealed its AI assistant Claude. OpenAI's multimodal large language model GPT-4, is the best announcement. William Douglas Heaven, my colleague, got an exclusive preview. Check out his first impressions.
GPT-4, unlike OpenAI's viral ChatGPT, is only currently available to developers. The tech is still in its early stages and will take some time to integrate into new products or services. People are testing its capabilities in public. Here are some of my favorite ways they do it.
Hustling
Jackson Greathouse Fall, a branding designer, asked GPT-4 for as much money possible using a $100 budget. Fall claimed he was a "human liaison" who bought whatever the computer program instructed him to.
GPT-4 suggested that he create an affiliate marketing website to promote links to other products, in this case eco-friendly. Fall asked GPT-4 to generate prompts to allow him to create an OpenAI logo-generating AI system DALLE 2. GPT-4 was also requested by Fall to create content and allocate money for social media advertising.
People on social media were interested in Fall's GPT-4-inspired marketing company, so the stunt attracted a lot of attention. Fall ended up with $13,78.84 cash. Although this is a publicity stunt it also shows how AI can be used to help people think of new ideas.
Productivity
Big tech companies want you to use AI in your work environment. This is likely how most people will use the new technology. Microsoft wants you use GPT-4 within its Office suite to help with PowerPoint presentations and summarize documents. This is exactly what we predicted back in January.
Google also announced similar AI tech within its office products including Gmail and Google Docs. This will allow people to create emails, proofread text, and generate images for presentations.
Health care
Nikhil Buduma, Mike Ng and Mike Ng were my cofounders at Ambience Health. OpenAI funds Ambience Health. GPT-4 is used by the startup to generate medical documentation that is based on conversations between provider and patient. The startup claims that it will reduce doctors' workloads by taking out the tedious parts of the job like data entry.
Buduma claims that GPT-4 does a better job following instructions than its predecessors. It's not clear how it will perform in areas like health care where accuracy is crucial. OpenAI claims it has fixed some flaws in AI language models, but GPT-4 still has many of these. It fabricates things and presents falsehoods confidently in its place as facts. It is still biased. Ng says that this is why it's important to ensure these models are safe and soundly deployed.
Princeton University computer science professor Arvind Narayanan claims that it took him less than ten minutes to get GPT-4 code to convert URLs into citations.
Narayanan claims that he has been using AI tools to generate text, images, and code, and that code generation is his favorite. He tweeted, "I believe the benefit of LLM (large language model) code generation is both psychological and time-saving."
Greg Brockman, OpenAI founder and CEO, used GPT-4 in a demonstration to create a website using a simple image of a design that he had drawn on a napkin. Narayanan explains that this is precisely where the power lies in these AI systems: automating repetitive, low-stakes and time-consuming tasks.
[TAG0]
Writing books
Reid Hoffman, the cofounder and executive chair of LinkedIn, and an early investor at OpenAI, said that he used GPT-4 for help in writing a book called Impromptu. Amplifying our Humanity through AI. Hoffman claims it is the first book to be co-written by GPT-4. Its predecessor ChatGPT was used to create tons more books.
Hoffman was granted access to the system in the summer of last year and has been writing about his thoughts on how the AI model could be used for education, journalism, and the justice system. He also includes extracts from conversations with the system in the book. The book contains his vision of the future, GPT-4 is used to write new ideas and he analyzes the answers.
Last word… The coolest new shiny toy for the AI community is GPT-4. It is an assistive technology that can be used to generate ideas, simplify text, clarify concepts and automate tedious tasks. This is a great development for white-collar workers in the knowledge sector.
OpenAI warns against using the model and urges caution. There are other potentially dangerous uses that it could be used for, which we have not yet encountered. Let's not get too excited. There is no reason to stop people using these new powerful models to do harm, and there are no grounds to hold them responsible if they do.
Deeper Learning
Chinese tech giant Baidu has just released its answer for ChatGPT
So. Many. Chatbots. Baidu, a Chinese tech giant, is the latest to join the AI chatbot market. Baidu announced a new large-language model called ErnieBot late last week. It can solve math problems, write copy and answer questions about Chinese literature.
An alternative in Chinese: Ernie Bot. The name stands for "EnhancedRepresentation from kNowledgeIntEgration" and its Chinese name is Wen Xin Yi Yan. It excels at tasks that are specific to Chinese culture such as explaining a historical fact, or writing a traditional Chinese poem. Learn more from Zeyi Yang, my colleague.
Even Deeper Learning
Language models might be able "self-correct" biases if you ask them
Due to the large amount of human-produced content they are exposed to, large language models can be notorious for expressing toxic biases. If the models are sufficiently large, some biases may be corrected by them. Amazingly, we only need to ask.
Researchers at AI lab Anthropic discovered this fascinating finding after testing a variety of language models with different sizes and training levels. This work raises the obvious question of whether or not this "self-correction” could and should be built into language models right from the beginning. To learn more, read the entire story by Niall Firth.
Bits and Bytes
Google makes its generative AI tools accessible to developersAnother Google announcement was overshadowed in the OpenAI hype train. The company made some of its AI technology available to developers via an API that allows them to build products using its large language model PaLMs. (Google)
Midjourney's AI text-to-image has finally been mastered
This year, image-generating AI systems will get incredible results. Exhibit A. The most recent iteration Midjourney's text-to-image AI software can create images of people using five fingers. Mangled fingers were previously a sign that an image had been generated by a computer program. All this means that it will become increasingly difficult to discern what is real and what is not. (Ars Technica)
Artists could use a new tool to protect their images against being scraped by AI
University of Chicago researchers have created a tool that allows artists and designers to create a protective layer on their work. This prevents them from using it to train AI models that generate images. (University of Chicago).
Runway has launched an even more powerful text-to video AI system
The advancements in generative AI continue to make Runway, the video editing startup that co-created Stable Diffusion's text-to-image model, a major update to its generative-video-making software a month after it launched the previous version. Gen-2 is a new model that improves upon the Gen-1 model (which Will Douglas Heaven has written about). It increases the quality of the generated video and allows you to create videos from scratch using only a text prompt.
Thank you for reading!
Melissa
————————————————————————————————————————————————————————————
By: Melissa Heikkilä
Title: How AI experts are using GPT-4
Sourced From: www.technologyreview.com/2023/03/21/1070102/how-ai-experts-are-using-gpt-4/
Published Date: Tue, 21 Mar 2023 11:06:19 +0000
Leave a Reply