From ChatGPT to Gemini: how AI is rewriting the internet – The Verge

People have been trying to talk to computers for almost as long as they’ve been building computers. For decades, many in tech have been convinced that this was the trick: if we could figure out a way to talk to our computers the way we talk to other people, and for computers to talk back the same way, it would make those computers easier to understand and operate, more accessible to everyone, and just more fun to use.

ChatGPT and the current revolution in AI chatbots is really only the latest version of this trend, which extends all the way back to the 1960s. That’s when Joseph Weizenbaum, a professor at MIT, built a chatbot named Eliza. Weizenbaum wrote in an academic journal in 1966 that Eliza “makes certain kinds of natural language conversation between man and computer possible.” He set up the bot to act as a therapist, a vessel into which people could pour their problems and thoughts. 

The historically inaccurate images and text generated by Google’s Gemini AI have “offended our users and shown bias,” CEO Sundar Pichai told employees in an internal memo obtained by The Verge.

Last week, Google paused Gemini’s ability to generate images after it was widely discovered that the model generated racially diverse, Nazi-era German soldiers, US Founding Fathers who were non-white, and even inaccurately portrayed the races of Google’s own co-founders. While Google has since apologized for “missing the mark” and said it’s working to re-enable image generation in the coming weeks, Tuesday’s memo is the first time the CEO has widely addressed the controversy.

Imagine waiting in line at your local Wendy’s drive-thru during the lunch rush, only to pull up to the menu board and realize that the spicy chicken nuggets you’ve been craving all day will cost you a dollar extra. That nightmare may soon become a reality, because Wendy’s plans on testing surge pricing that will increase the price of its spicy nuggets, burgers, Frostys, and other favorites during its busiest times.

During an earnings call earlier this month, Wendy’s CEO Kirk Tanner said the fast food chain plans on investing $20 million to roll out digital menu boards to US-based restaurants by the end of 2025. As part of the change, Wendy’s will also introduce something called “dynamic prices” that will change the prices on the digital menu boards based on demand. It sounds similar to the surge pricing system implemented by Uber, which charges riders higher rates in busy areas.

Microsoft has announced a new multiyear partnership with Mistral, a French AI startup that’s valued at €2 billion (about $2.1 billion). The Financial Times reports that the partnership will include Microsoft taking a minor stake in the 10-month-old AI company, just a little over a year after Microsoft invested more than $10 billion into its OpenAI partnership.

The deal will see Mistral’s open and commercial language models available on Microsoft’s Azure AI platform, the second company to offer a commercial language model on Azure after OpenAI. Much like the OpenAI partnership, Microsoft’s partnership with Mistral will also be focused on the development and deployment of next-generation large language models.

Google has issued an explanation for the “embarrassing and wrong” images generated by its Gemini AI tool. In a blog post on Friday, Google says its model produced “inaccurate historical” images due to tuning issues. The Verge and others caught Gemini generating images of racially diverse Nazis and US Founding Fathers earlier this week.

“Our tuning to ensure that Gemini showed a range of people failed to account for cases that should clearly not show a range,” Prabhakar Raghavan, Google’s senior vice president, writes in the post. “And second, over time, the model became way more cautious than we intended and refused to answer certain prompts entirely — wrongly interpreting some very anodyne prompts as sensitive.”

Google and Samsung aren’t the only ones baking magical AI selective photo erasers into their devices — they’re about to become table stakes for Windows PCs too. Microsoft has just announced Generative erase, a feature that lets you do similar things in the Photos app that comes bundled with Windows.

Above and below, you can see how Microsoft disappears a dog’s leash and some unintentional photobombers using the power of generative AI:

Google is getting AI training data from Reddit as part of a new partnership between the two companies. In an update on Thursday, Reddit announced it will start providing Google “more efficient ways to train models.”

The collaboration will give Google access to Reddit’s data API, which delivers real-time content from Reddit’s platform. This will provide “Google with an efficient and structured way to access the vast corpus of existing content on Reddit,” while also allowing the company to display content from Reddit in new ways across its products.

Google has apologized for what it describes as “inaccuracies in some historical image generation depictions” with its Gemini AI tool, saying its attempts at creating a “wide range” of results missed the mark. The statement follows criticism that it depicted specific white figures (like the US Founding Fathers) or groups like Nazi-era German soldiers as people of color, possibly as an overcorrection to long-standing racial bias problems in AI.

“We’re aware that Gemini is offering inaccuracies in some historical image generation depictions,” says the Google statement, posted this afternoon on X. “We’re working to improve these kinds of depictions immediately. Gemini’s AI image generation does generate a wide range of people. And that’s generally a good thing because people around the world use it. But it’s missing the mark here.”

Microsoft’s Copilot Pro launched last month as a $20 monthly subscription that provides access to AI-powered features inside some Office apps, alongside priority access to the latest OpenAI models and improved image generation.

I’ve been testing Copilot Pro over the past month to see if it’s worth the $20 subscription for my daily needs and just how good or bad the AI image and text generation is across Office apps like Word, Excel, and PowerPoint. Some of the Copilot Pro features are a little disappointing right now, whereas others are truly useful improvements that I’m not sure I want to live without.

The US Patent and Trademark Office (PTO) has denied OpenAI’s application to register the word GPT, which means generative pre-trained transformer, saying GPT is too general a term to register and can prevent competitors from correctly describing their products as a GPT.

OpenAI argued in its application that GPT is not a descriptive word — that GPT isn’t such a general term that consumers would “immediately understand” what it means.

In 2018, a viral joke started going around the internet: scripts based on “making a bot watch 1,000 hours” of just about anything. The premise (concocted by comedian Keaton Patti) was that you could train an artificial intelligence model on vast quantities of Saw films, Hallmark specials, or Olive Garden commercials and get back a bizarre funhouse-mirror version with lines like “lasagna wings with extra Italy” or “her mouth is full of secret soup.” The scripts almost certainly weren’t actually written by a bot, but the joke conveyed a common cultural understanding: AI was weird.

Strange AI was everywhere a few years ago. AI Dungeon, a text adventure game genuinely powered by OpenAI’s GPT-2 and GPT-3, touted its ability to produce deeply imagined stories about the inner life of a chair. The first well-known AI art tools, like Google’s computer vision program Deep Dream, produced unabashedly bizarre Giger-esque nightmares. Perhaps the archetypal example was Janelle Shane’s blog AI Weirdness, where Shane trained models to create physically impossible nuclear waste warnings or sublimely inedible recipes. “Made by a bot” was shorthand for a kind of free-associative, nonsensical surrealism — both because of the models’ technical limitations and because they were more curiosities than commercial products. Lots of people had seen what “a bot” (actually or supposedly) produced. Fewer had used one. Even fewer had to worry about them in day-to-day life.

It’s common knowledge that machine learning consumes a lot of energy. All those AI models powering email summaries, regicidal chatbots, and videos of Homer Simpson singing nu-metal are racking up a hefty server bill measured in megawatts per hour. But no one, it seems — not even the companies behind the tech — can say exactly what the cost is. 

Estimates do exist, but experts say those figures are partial and contingent, offering only a glimpse of AI’s total energy usage. This is because machine learning models are incredibly variable, able to be configured in ways that dramatically alter their power consumption. Moreover, the organizations best placed to produce a bill — companies like Meta, Microsoft, and OpenAI — simply aren’t sharing the relevant information. (Judy Priest, CTO for cloud operations and innovations at Microsoft said in an e-mail that the company is currently “investing in developing methodologies to quantify the energy use and carbon impact of AI while working on ways to make large systems more efficient, in both training and application.” OpenAI and Meta did not respond to requests for comment.)

Our new Thursday episodes of Decoder are all about deep dives into big topics in the news, and for the next few weeks, we’re going to stay focused on one of the biggest topics of all: generative AI. 

There’s a lot going on in the world of generative AI, but maybe the biggest is the increasing number of copyright lawsuits being filed against AI companies like OpenAI and Stability AI. So for this episode, we brought on Verge features editor Sarah Jeong, who’s a former lawyer just like me, and we’re going to talk about those cases and the main defense the AI companies are relying on in those copyright cases: an idea called fair use.