Keeping up with an industry as rapidly changing as AI is a challenge. So until AI can do it for you, here’s a quick recap of recent stories in the world of machine learning, as well as notable research and experiments that we couldn’t cover on our own.
This week in the AI space, OpenAI hosted the first of what will likely be many developer conferences to come. During the keynote, the company unveiled a number of new products, including an improved version of GPT-4, a new text-to-speech model, and an API for image generation DALL-E 3.
But arguably the most important announcement was GPT.
OpenAI’s GPT provides a way for developers to build their own conversational AI systems that leverage OpenAI’s models and publish them on an OpenAI-hosted marketplace called the GPT Store. OpenAI CEO Sam Altman said on stage at the conference that developers will soon be able to monetize GPT based on the number of people using it.
“We believe that if you give people better tools, they will do amazing things,” Altman said. “Once you build a GPT, you can publish it for others to use. It also has a combination of instructions, expanded knowledge, and actions that make it more useful.”
OpenAI’s transition from AI model provider to platform was certainly interesting, but not entirely unexpected. The startup announced its ambitions in March by releasing a plugin for its AI-powered chatbot, ChatGPT, which will bring third parties into OpenAI’s model ecosystem for the first time. Did.
But what caught me off guard was the breadth and depth of OpenAI’s GPT construction (and commercialization) tools.
According to my colleague Devin Coldewey, who attended OpenAI’s conference in person, the GPT experience works more or less as advertised, although it was “a little glitchy” in the demo. GPT requires no coding experience and can be as simple or complex as the developer desires. For example, you can train GPT on a cookbook collection to answer questions about ingredients in a particular recipe. Alternatively, GPT can bring in a company’s own codebase, allowing developers to check their own style and generate code that aligns with best practices.
GPT effectively democratizes the creation of generative AI apps, at least for apps that use OpenAI’s family of models. And if I were his OpenAI rival, at least a rival without the backing of Big Tech, I would rush to the metaphorical war room to gather a response.
GPT has the potential to kill consulting firms whose business model is to build what is essentially GPT for their clients. Additionally, for customers with developer talent, it may be possible to create model providers such as: please do not Given the complexity of having to incorporate a provider’s API into existing apps and services, offering any form of app building tooling is less attractive.
Is that a good thing? I’m not necessarily, but I’m concerned about the possibility of monopoly. But OpenAI has a first-mover advantage and is taking advantage of it, for better or for worse.
Here are some other notable AI stories from the past few days.
- Samsung unveils generative AI: Just days after the OpenAI development event, Samsung announced its own generative AI family Samsung Gauss at Samsung AI Forum 2023. It consists of his three models: a large language model similar to ChatGPT, a code generation model, and an image generation model. Editorial Model — Samsung Gauss is currently being used internally by his Samsung staff and will be available to general users “in the near future,” the company said.
- Microsoft provides free AI computing to startups. Microsoft announced this week that it is updating its startup program, Microsoft for Startups Founders Hub, to include a free Azure AI infrastructure option for “high-end” Nvidia-based GPU virtual machine clusters to train and run generative models. did. Y Combinator and its community of startup founders will first access the cluster in a private his preview, followed by his Microsoft venture fund, his M12 and startups in the M12 portfolio, and then other startups. Investors and accelerators may also participate.
- YouTube tests generative AI features. YouTube will soon begin experimenting with new generative AI features, the company said. announced this week. As part of the Premium His package, which is available to YouTube’s paid subscribers, users can use his AI to answer questions and make recommendations about his YouTube content, as well as a conversation tool that allows you to change topics within the comments of your videos. You can try out the ability to summarize.
- Interview with DeepMind’s Head of Robotics: Brian spoke to Vincent Vanhoucke, head of robotics at Google DeepMind, about Google’s big robotics ambitions. The interview touched on a wide range of topics, including general-purpose robots, generative AI, and most importantly, office Wi-Fi.
- Kai-Fu Lee’s AI startup launches model: Kai-Fu Lee, the computer scientist known in the West for his bestseller AI Superpowers and in China for his bet on AI unicorns, has achieved remarkable status with his AI startup. 01.AI. Seven months after its founding, his $1 billion-valued 01.AI released its first model, open source. Yi-34B.
- GitHub introduces customizable Copilot plans. GitHub this week announced plans for an enterprise subscription tier that will allow companies to fine-tune Copilot pair programmers based on their in-house codebase. The news forms part of a slew of notable information the Microsoft-owned company revealed at Wednesday’s annual GitHub Universe developer conference, including a new partner program and Copilot Chat (recently It includes more clarity about when Copilot’s announced chatbot-like features will be used. Officially available.
- Hugging Face’s two-person model team: AI startup Hugging Face offers a wide range of data science hosting and development tools. But some of the company’s best and most capable tools these days are produced by his two-person team he just formed in January, called H4.
- Mozilla releases AI chatbot: Earlier this year, Mozilla acquired Fakespot, a startup that leverages AI and machine learning to identify fake and deceptive product reviews. Currently, Mozilla is fake spot chat, An AI agent that helps consumers shop online by answering questions about products and suggesting questions that may help with product research.
More machine learning
After perusing many previous examples, we have seen in many fields how machine learning models can make very good short-term predictions for complex data structures. . For example, the warning period for an upcoming earthquake could be extended, giving people a critical 20 to 30 seconds to evacuate. And Google has shown that it’s also good at predicting weather patterns.
Some of this came from a post showing how MetNet integrates data into ML-based predictions. Image credits: Google
MetNet-3 is a suite of physically-based weather models that looks at a variety of variables, including precipitation, temperature, wind, and cloud cover, and produces incredibly high-resolution (temporally and spatially) predictions of what’s likely to happen. This is the latest version. Please come next time. Many of these types of predictions are based on fairly old models, which may or may not be accurate, and can be made more accurate by combining data with other sources. This is what MetNet-3 does. I won’t go into too much detail, but They posted a very interesting post on this topic This from last week gives us a good idea of how modern weather prediction engines work.
In other very special science news, researchers at the University of Kansas say: AI-generated text detector…for magazine articles about chemistry. Of course, it won’t help most people, but after OpenAI and others put the brakes on their detector models, it will at least help show that something more specific is possible. “Most of the field of text analysis is looking for a truly general-purpose detector that will work with anything,” says co-author Heather Desaire. “We were really looking for accuracy.”
Their model was trained on articles from the Journal of the American Chemical Society and learned to write the introductory section from just the title and abstract. We were then able to identify intros written in ChatGPT-3.5 with near-perfect accuracy. Obviously, this is a very narrow use case, but the team notes that it was fairly quick and easy to set up. This means you can set up detectors for different sciences, journals, and languages.
There isn’t one for college admissions essays yet, but AI will soon be on the other side of the process, helping admissions officers identify diamonds in the rough rather than deciding who gets admitted. You may be able to do it. Researchers at the University of Colorado and his UPenn have shown that ML models can: Successfully identify sentences in student essays that demonstrate interests and qualitiesleadership and “prosocial purpose.”
Although students are not graded this way (again, not yet), it is a much-needed tool in an administrator’s toolbox, and administrators have to process thousands of applications. Yes, I sometimes use my hands. You can also use an analysis layer like this to group essays and randomize them better to avoid having everyone talking about camp in a row. And the study found that the language used by students was surprisingly predictive of certain academic factors, such as graduation rates. Of course, they’ll be looking into it more deeply, but it’s clear that ML-based styrometry will continue to be important.
But we must not lose sight of AI’s limitations, as highlighted by a group of researchers at the University of Washington who tested the compatibility of AI tools with our unique accessibility needs. Their experiences have been decidedly mixed, with summarization systems adding bias and illusions (making them inappropriate for people who can’t read the source material), and inconsistent application of accessibility content rules. .
But at the same time, people on the autism spectrum have found that by using language models to generate messages on Slack, they can overcome their lack of confidence in their normal communication skills. Although her colleagues felt the message was a bit “robotic”, it was a net benefit to the users and this was a start. More information about this study can be found here.
However, both of the above items raise thorny issues of bias in sensitive areas and the general weirdness of AI, leading some states and local governments to revise their rules regarding the use of AI in public service. It is not surprising that they are considering enacting it. For example, Seattle We just announced a set of “Governing Principles” A toolkit that should be referenced or applied before using AI models for official purposes. There is no doubt that different and perhaps contradictory sets of rules will be introduced at all levels of governance.
Inside VR, a machine learning model acts as a flexible gesture detector. A very interesting set of ways to interact with virtual objects. “What’s the point of using VR if using it is the same as using a keyboard and mouse?” asked lead author Per Ola Christenson. “We need to give them almost superhuman powers that they can’t get anywhere else.” Good perspective!
You can see exactly how it works in the video above, and it makes perfect intuitive sense when you think about it. I don’t want to select “Copy” from the menu and then select “Paste”. mouse finger. I want to hold the object in one hand, open the palm of the other hand, and suddenly duplicate it. So if you want to cut it, should you use your hands as scissors? This is amazing!
Image credits: EPFL
Finally, speaking of cut/paste, New exhibition at the Swiss university EPFL, students and professors explored the history of comics since the 1950s and how AI can enhance or interpret comics. Although it is clear that his art is not yet completely popular, some artists are clearly experimenting with new technologies and interpreting historical materials, despite ethical and copyright issues. Eager to explore. If you’re lucky enough to be in Lausanne, check out Couper/Coller, a catchy local version of the ubiquitous digital action.