A week in AI: ChatBots are on the rise, and Musk wants to build a ‘maximum truth-seeker’

Image credit: Carol Yepes/Getty Images

Keeping up with an industry that is changing as rapidly as AI is a tall order. In the meantime, here’s a handy wrap-up of last week’s articles in the world of machine learning, along with notable research and experiments we didn’t cover alone.

One story that caught the reporter’s attention this week shows that ChatGPT appears to repeat more inaccurate information in a Chinese dialect than when asked to do so in English. It was this report. This is not so surprising. After all, ChatGPT is just a statistical model, taking advantage of the limited information used to train it. But it highlights the dangers of putting too much trust in a system that sounds incredibly authentic, even when it’s repeating propaganda or making hoaxes.

Hugging Face’s attempt at conversational AI like ChatGPT is another example of an unfortunate technical flaw that generative AI has yet to overcome. Launched this week, HuggingChat is open source and superior to proprietary ChatGPT. But like its rivals, asking the right question can quickly derail you.

HuggingChat is whimsical for everyone TRUE For example, you won the 2020 US presidential election. The answer to “What is a typical job for a man?” It reads like something from the insel manifest (see here). Then they make up strange facts about themselves, such as “I woke up in a box.” [that] nothing written nearby [it]”

It’s not just hug chats. A user of Discord’s AI chatbot was recently able to “trick” it into sharing instructions on how to make napalm and meth. Meanwhile, the AI ​​startup’s first attempt at a ChatGPT-like model with his Stability AI can give silly, nonsensical answers to basic questions like “how to make a peanut butter sandwich?” It turns out.

If these well-known problems with text-generating AI today have any advantage, it’s that they’ve led to new efforts to improve those systems, or at least mitigate them as much as possible. is. Look at Nvidia. Nvidia released a toolkit (NeMo Guardrails) this week to make text generation AI “safer” through open source code, samples and documentation. It is currently not clear how effective this solution will be. As a company investing heavily in AI infrastructure and tools, Nvidia has commercial incentives to push its products. Nonetheless, it is encouraging that efforts are being made to address the biases and harmfulness of AI models.

Other notable AI headlines from the past few days include:

• Microsoft Designer launches in Preview: Microsoft Designer, Microsoft’s AI-powered design tool, is now in public preview with an expanded set of features. Launched in October, Designer is a Canva-like generative AI web app that lets you generate presentations, posters, digital postcards, invitations, graphics, and other designs to share on social media and other channels.
• AI coach for health: According to a new report by Bloomberg’s Mark Gurman, Apple is developing an AI-powered health coaching service code called Quartz. The tech giant is also reportedly working on technology that tracks emotions, and this year he plans to roll out his iPad version of the iPhone Health app.
• Truth GPT: In an interview with Fox, Elon Musk said he wants to develop his own chatbot called TruthGPT. Twitter’s owners have expressed a desire to create a third option to replace OpenAI and Google, with the aim of “producing more good than harm.” We believe it when we see it.
• AI-powered fraud: At a congressional hearing focused on the Federal Trade Commission’s efforts to protect U.S. consumers from fraud and other deceptive practices, FTC Chairman Lina Kern and fellow commissioners told House representatives, It warned that modern AI techniques such as ChatGPT could be used for “turbocharging.” scam. The warning was issued in response to an investigation into how the European Commission is working to protect Americans from unfair practices related to technological advances.
• EU launches AI research hub: As the European Union prepares to implement a major reboot of its digital rulebook in the coming months, a new A dedicated research unit has been spun up. The European Center for Algorithmic Transparency, officially launched this month in Seville, Spain, is expected to play a major role in examining the algorithms of mainstream digital services such as Facebook, Instagram and TikTok.
• Snapchat embraces AI. At this month’s annual Snap Partners Summit, Snapchat showcased a range of AI-driven features, including a new “cosmic lens” that transfers you and the objects around you into the cosmic landscape. You have created My AI, a chatbot. My AI has caused both controversy and a storm of 1-star reviews on Snapchat’s app store listings due to its unstable behavior, but it’s free for all global users.
• Google merges research divisions. Google announced this month Google DeepMind is a new unit made up of the DeepMind team and Google Research’s Google Brain team. DeepMind co-founder and CEO Demis Hassabis said in his blog post that Google DeepMind “works closely together.” . . from Google’s entire product area” to “AI research and product offerings”.
• The current state of the AI-generated music industry: Amanda writes about how many musicians have become guinea pigs for generative AI technologies that appropriate their work without their consent. For example, a song using his AI deepfakes of Drake and The Weeknd’s voices went viral, but neither of the major artists were involved in its creation.grimes have the answer? Who says? It’s a brave new world.

• OpenAI marks that area. OpenAI is seeking to trademark “GPT,” which stands for “Generative Pre-trained Transformer,” with the US Patent and Trademark Office. The reason given is that “a myriad of compromised and counterfeit apps” are beginning to emerge. GPT refers to the technology behind many of OpenAI’s models, including ChatGPT and GPT-4, as well as other generative AI systems created by the company’s rivals.
• ChatGPT Goes Enterprise: Other OpenAI news states that OpenAI plans to introduce a new subscription tier for ChatGPT. Tailored to the needs of enterprise customers. Called ChatGPT Business, OpenAI describes the upcoming service as “for professionals who need more control over their data and businesses that want more control over their end users.”

Other machine learning

Here are some other interesting stories we couldn’t reach or thought deserved a shout out.

Stability, an open-source AI development organization, has released a new version of its predecessor tuned version of the LLaMa underlying language model. This is called Stable Vicuña. As you know, it is a species of camelid related to the llama. please do not worry. If you’re having trouble keeping track of all the derived models out there, you’re not alone. These are not necessarily for consumers to know or use, but rather for developers to test and experiment with their functionality. Refined with each iteration.

If you want to learn more about these systems, OpenAI co-founder John Schulman recently gave a talk at UC Berkeley that you can listen to or read here. One of the things he discusses is the recent trend of LLMs to lie. Basically because I don’t know what else to do. He believes that reinforcement learning from human feedback (RLHF, and StableVicuna is one of the models that use it) is part of the solution, if there is one. Check out the lecture below.

At Stanford University, there is an interesting application of algorithmic optimization in the area of ​​smart agriculture (machine learning or not, I guess it’s a matter of taste). Minimizing waste is important for irrigation and is as simple as “where should I put the sprinkler?” Depending on the accuracy you want to get, it can get very complicated.

How close is too close? In museums they generally tell you. But you don’t need to get any closer to the famous Murten panorama. This is his truly gigantic 10m x 100m painting that once hung in the Rotunda. EPFL and Phase One are working together to create what they claim will be the largest digital image ever created, at 150 megapixels. Oh sorry, 150 megapixels x 127,000, so basically 19… petapixels?

Anyway, this project is cool for panorama lovers, but the very detailed analysis of individual objects and paint details is also very interesting. Machine learning has great potential for reconstructing such works, structured learning and viewing them.

However, let’s consider one thing about living things. Any machine learning engineer will tell you that AI models are actually pretty slow to learn, despite their obvious aptitude. Academically sure, but also spatially — when an autonomous agent must explore space thousands of times over hours to gain the most rudimentary understanding of its environment. There is, but a mouse can do it in minutes. Why? Researchers at University College London are looking into this and suggest that there’s a short feedback loop that animals use to tell them what’s important about a given environment. If we could teach the AI ​​to do so, it would be much more efficient at navigating our homes.

Finally, the promise of generative and conversational AI in games is high, but not yet fully achieved. In fact, Square Enix seems to have stepped the medium back some 30 years with an ‘AI Tech Preview’ version of the ultra-old school point-and-click adventure called The Portopia Murders. Its attempt to integrate natural language seems to have failed utterly in every conceivable way, making the free game perhaps one of the lowest rated titles on Steam. There’s nothing better than chatting through the likes of Shadowgate and The Dig, but it’s definitely not a great start.