Enlarge (credit: Getty Images)

On Wednesday, Axios broke the news that OpenAI had signed deals with The Atlantic and Vox Media that will allow the ChatGPT maker to license their editorial content to further train its language models. But some of the publications' writers—and the unions that represent them—were surprised by the announcements and aren't happy about it. Already, two unions have released statements expressing "alarm" and "concern."

"The unionized members of The Atlantic Editorial and Business and Technology units are deeply troubled by the opaque agreement The Atlantic has made with OpenAI," reads a statement from the Atlantic union. "And especially by management's complete lack of transparency about what the agreement entails and how it will affect our work."

The Vox Union—which represents The Verge, SB Nation, and Vulture, among other publications—reacted in similar fashion, writing in a statement, "Today, members of the Vox Media Union ... were informed without warning that Vox Media entered into a 'strategic content and product partnership' with OpenAI. As both journalists and workers, we have serious concerns about this partnership, which we believe could adversely impact members of our union, not to mention the well-documented ethical and environmental concerns surrounding the use of generative AI."

Enlarge / The Google "G" logo surrounded by whimsical characters, all of which look stunned and surprised. (credit: Google)

On Thursday, Google capped off a rough week of providing inaccurate and sometimes dangerous answers through its experimental AI Overview feature by authoring a follow-up blog post titled, "AI Overviews: About last week." In the post, attributed to Google VP Liz Reid, head of Google Search, the firm formally acknowledged issues with the feature and outlined steps taken to improve a system that appears flawed by design, even if it doesn't realize it is admitting it.

To recap, the AI Overview feature—which the company showed off at Google I/O a few weeks ago—aims to provide search users with summarized answers to questions by using an AI model integrated with Google's web ranking systems. Right now, it's an experimental feature that is not active for everyone, but when a participating user searches for a topic, they might see an AI-generated answer at the top of the results, pulled from highly ranked web content and summarized by an AI model.

While Google claims this approach is "highly effective" and on par with its Featured Snippets in terms of accuracy, the past week has seen numerous examples of the AI system generating bizarre, incorrect, or even potentially harmful responses, as we detailed in a recent feature where Ars reporter Kyle Orland replicated many of the unusual outputs.

Enlarge (credit: Getty Images)

On Thursday, several major tech companies, including Google, Intel, Microsoft, Meta, AMD, Hewlett-Packard Enterprise, Cisco, and Broadcom, announced the formation of the Ultra Accelerator Link (UALink) Promoter Group to develop a new interconnect standard for AI accelerator chips in data centers. The group aims to create an alternative to Nvidia's proprietary NVLink interconnect technology, which links together multiple servers that power today's AI applications like ChatGPT.

The beating heart of AI these days lies in GPUs, which can perform massive numbers of matrix multiplications—necessary for running neural network architecture—in parallel. But one GPU often isn't enough for complex AI systems. NVLink can connect multiple AI accelerator chips within a server or across multiple servers. These interconnects enable faster data transfer and communication between the accelerators, allowing them to work together more efficiently on complex tasks like training large AI models.

This linkage is a key part of any modern AI data center system, and whoever controls the link standard can effectively dictate which hardware the tech companies will use. Along those lines, the UALink group seeks to establish an open standard that allows multiple companies to contribute and develop AI hardware advancements instead of being locked into Nvidia's proprietary ecosystem. This approach is similar to other open standards, such as Compute Express Link (CXL)—created by Intel in 2019—which provides high-speed, high-capacity connections between CPUs and devices or memory in data centers.

Enlarge (credit: Getty Images)

On Monday, OpenAI announced the formation of a new "Safety and Security Committee" to oversee risk management for its projects and operations. The announcement comes as the company says it has "recently begun" training its next frontier model, which it expects to bring the company closer to its goal of achieving artificial general intelligence (AGI), though some critics say AGI is farther off than we might think. It also comes as a reaction to two weeks of public setbacks for the company.

Whether the aforementioned new frontier model is intended to be GPT-5 or a step beyond that is currently unknown. In the AI industry, "frontier model" is a term for a new AI system designed to push the boundaries of current capabilities. And "AGI" refers to a hypothetical AI system with human-level abilities to perform novel, general tasks beyond its training data (unlike narrow AI, which is trained for specific tasks).

Meanwhile, the new Safety and Security Committee, led by OpenAI directors Bret Taylor (chair), Adam D'Angelo, Nicole Seligman, and Sam Altman (CEO), will be responsible for making recommendations about AI safety to the full company board of directors. In this case, "safety" partially means the usual "we won't let the AI go rogue and take over the world," but it also includes a broader set of "processes and safeguards" that the company spelled out in a May 21 safety update related to alignment research, protecting children, upholding election integrity, assessing societal impacts, and implementing security measures.

Enlarge / This is fine. (credit: Getty Images)

If you use Google regularly, you may have noticed the company's new AI Overviews providing summarized answers to some of your questions in recent days. If you use social media regularly, you may have come across many examples of those AI Overviews being hilariously or even dangerously wrong.

Factual errors can pop up in existing LLM chatbots as well, of course. But the potential damage that can be caused by AI inaccuracy gets multiplied when those errors appear atop the ultra-valuable web real estate of the Google search results page.

"The examples we've seen are generally very uncommon queries and aren’t representative of most people’s experiences," a Google spokesperson told Ars. "The vast majority of AI Overviews provide high quality information, with links to dig deeper on the web."

Enlarge / Nathan Benaich of Air Street Capital delivers the opening presentation on the state of AI at EmTech Digital 2024 on May 22, 2024. (credit: Benj Edwards)

CAMBRIDGE, Massachusetts—On Wednesday, AI enthusiasts and experts gathered to hear a series of presentations about the state of AI at EmTech Digital 2024 on the Massachusetts Institute of Technology's campus. The event was hosted by the publication MIT Technology Review. The overall consensus is that generative AI is still in its very early stages—with policy, regulations, and social norms still being established—and its growth is likely to continue into the future.

I was there to check the event out. MIT is the birthplace of many tech innovations—including the first action-oriented computer video game—among others, so it felt fitting to hear talks about the latest tech craze in the same building that hosts MIT's Media Lab on its sprawling and lush campus.

EmTech's speakers included AI researchers, policy experts, critics, and company spokespeople. A corporate feel pervaded the event due to strategic sponsorships, but it was handled in a low-key way that matches the level-headed tech coverage coming out of MIT Technology Review. After each presentation, MIT Technology Review staff—such as Editor-in-Chief Mat Honan and Senior Reporter Melissa Heikkilä—did a brief sit-down interview with the speaker, pushing back on some points and emphasizing others. Then the speaker took a few audience questions if time allowed.

Enlarge / A screenshot of Microsoft's new "Recall" feature in action. (credit: Microsoft)

At a Build conference event on Monday, Microsoft revealed a new AI-powered feature called "Recall" for Copilot+ PCs that will allow Windows 11 users to search and retrieve their past activities on their PC. To make it work, Recall records everything users do on their PC, including activities in apps, communications in live meetings, and websites visited for research. Despite encryption and local storage, the new feature raises privacy concerns for certain Windows users.

"Recall uses Copilot+ PC advanced processing capabilities to take images of your active screen every few seconds," Microsoft says on its website. "The snapshots are encrypted and saved on your PC’s hard drive. You can use Recall to locate the content you have viewed on your PC using search or on a timeline bar that allows you to scroll through your snapshots."

By performing a Recall action, users can access a snapshot from a specific time period, providing context for the event or moment they are searching for. It also allows users to search through teleconference meetings they've participated in and videos watched using an AI-powered feature that transcribes and translates speech.

Enlarge (credit: Benj Edwards | Getty Images)

Since the launch of its latest AI language model, GPT-4o, OpenAI has found itself on the defensive over the past week due to a string of bad news, rumors, and ridicule circulating on traditional and social media. The negative attention is potentially a sign that OpenAI has entered a new level of public visibility—and is more prominently receiving pushback to its AI approach beyond what it has seen from tech pundits and government regulators.

OpenAI's rough week started last Monday when the company previewed a flirty AI assistant with a voice seemingly inspired by Scarlett Johansson from the 2013 film Her. OpenAI CEO Sam Altman alluded to the film himself on X just before the event, and we had previously made that comparison with an earlier voice interface for ChatGPT that launched in September 2023.

While that September update included a voice called "Sky" that some have said sounds like Johansson, it was GPT-4o's seemingly lifelike new conversational interface, complete with laughing and emotionally charged tonal shifts, that led to a widely circulated Daily Show segment ridiculing the demo for its perceived flirty nature. Next, a Saturday Night Live joke reinforced an implied connection to Johansson's voice.

Enlarge (credit: Tim Robberts | DigitalVision)

After launching Slack AI in February, Slack appears to be digging its heels in, defending its vague policy that by default sucks up customers' data—including messages, content, and files—to train Slack's global AI models.

According to Slack engineer Aaron Maurer, Slack has explained in a blog that the Salesforce-owned chat service does not train its large language models (LLMs) on customer data. But Slack's policy may need updating "to explain more carefully how these privacy principles play with Slack AI," Maurer wrote on Threads, partly because the policy "was originally written about the search/recommendation work we've been doing for years prior to Slack AI."

Maurer was responding to a Threads post from engineer and writer Gergely Orosz, who called for companies to opt out of data sharing until the policy is clarified, not by a blog, but in the actual policy language.

Enlarge / Still images taken from videos generated by Google Veo. (credit: Google / Benj Edwards)

On Tuesday at Google I/O 2024, Google announced Veo, a new AI video-synthesis model that can create HD videos from text, image, or video prompts, similar to OpenAI's Sora. It can generate 1080p videos lasting over a minute and edit videos from written instructions, but it has not yet been released for broad use.

Veo reportedly includes the ability to edit existing videos using text commands, maintain visual consistency across frames, and generate video sequences lasting up to and beyond 60 seconds from a single prompt or a series of prompts that form a narrative. The company says it can generate detailed scenes and apply cinematic effects such as time-lapses, aerial shots, and various visual styles

Since the launch of DALL-E 2 in April 2022, we've seen a parade of new image synthesis and video synthesis models that aim to allow anyone who can type a written description to create a detailed image or video. While neither technology has been fully refined, both AI image and video generators have been steadily growing more capable.

Enlarge / A video still of Project Astra demo at the Google I/O conference keynote in Mountain View on May 14, 2024. (credit: Google)

Just one day after OpenAI revealed GPT-4o, which it bills as being able to understand what's taking place in a video feed and converse about it, Google announced Project Astra, a research prototype that features similar video comprehension capabilities. It was announced by Google DeepMind CEO Demis Hassabis on Tuesday at the Google I/O conference keynote in Mountain View, California.

Hassabis called Astra "a universal agent helpful in everyday life." During a demonstration, the research model showcased its capabilities by identifying sound-producing objects, providing creative alliterations, explaining code on a monitor, and locating misplaced items. The AI assistant also exhibited its potential in wearable devices, such as smart glasses, where it could analyze diagrams, suggest improvements, and generate witty responses to visual prompts.

Google says that Astra uses the camera and microphone on a user's device to provide assistance in everyday life. By continuously processing and encoding video frames and speech input, Astra creates a timeline of events and caches the information for quick recall. The company says that this enables the AI to identify objects, answer questions, and remember things it has seen that are no longer in the camera's frame.

Enlarge (credit: Getty Images)

On Monday, OpenAI employee William Fedus confirmed on X that a mysterious chart-topping AI chatbot known as "gpt-chatbot" that had been undergoing testing on LMSYS's Chatbot Arena and frustrating experts was, in fact, OpenAI's newly announced GPT-4o AI model. He also revealed that GPT-4o had topped the Chatbot Arena leaderboard, achieving the highest documented score ever.

"GPT-4o is our new state-of-the-art frontier model. We’ve been testing a version on the LMSys arena as im-also-a-good-gpt2-chatbot," Fedus tweeted.

Chatbot Arena is a website where visitors converse with two random AI language models side by side without knowing which model is which, then choose which model gives the best response. It's a perfect example of vibe-based AI benchmarking, as AI researcher Simon Willison calls it.

Enlarge (credit: Getty Images)

On Monday, OpenAI debuted GPT-4o (o for "omni"), a major new AI model that can ostensibly converse using speech in real time, reading emotional cues and responding to visual input. It operates faster than OpenAI's previous best model, GPT-4 Turbo, and will be free for ChatGPT users and available as a service through API, rolling out over the next few weeks, OpenAI says.

OpenAI revealed the new audio conversation and vision comprehension capabilities in a YouTube livestream titled "OpenAI Spring Update," presented by OpenAI CTO Mira Murati and employees Mark Chen and Barret Zoph that included live demos of GPT-4o in action.

OpenAI claims that GPT-4o responds to audio inputs in about 320 milliseconds on average, which is similar to human response times in conversation, according to a 2009 study, and much shorter than the typical 2–3 second lag experienced with previous models. With GPT-4o, OpenAI says it trained a brand-new AI model end-to-end using text, vision, and audio in a way that all inputs and outputs "are processed by the same neural network."

Enlarge (credit: boonchai wedmakawand)

Reinforcement-learning algorithms in systems like ChatGPT or Google’s Gemini can work wonders, but they usually need hundreds of thousands of shots at a task before they get good at it. That’s why it’s always been hard to transfer this performance to robots. You can’t let a self-driving car crash 3,000 times just so it can learn crashing is bad.

But now a team of researchers at Northwestern University may have found a way around it. “That is what we think is going to be transformative in the development of the embodied AI in the real world,” says Thomas Berrueta who led the development of the Maximum Diffusion Reinforcement Learning (MaxDiff RL), an algorithm tailored specifically for robots.

Introducing chaos

The problem with deploying most reinforcement-learning algorithms in robots starts with the built-in assumption that the data they learn from is independent and identically distributed. The independence, in this context, means the value of one variable does not depend on the value of another variable in the dataset—when you flip a coin two times, getting tails on the second attempt does not depend on the result of your first flip. Identical distribution means that the probability of seeing any specific outcome is the same. In the coin-flipping example, the probability of getting heads is the same as getting tails: 50 percent for each.