Believe it or not, this is not the first robotaxi-assisted theft.

OpenAI is rolling out some significant enhancements to ChatGPT’s memory feature, particularly if you’ve been using the chatbot through a free account. Before getting to those improvements, a quick recap will help set the stage for what to expect: OpenAI shipped its first memory feature in April 2024. By the company’s own admission, this early implementation, then known as saved memories, was basic. It depended on strong cues from user, such as a direct prompt telling ChatGPT to remember a fact. People also found the chatbot’s memories became less relevant over time. So over the next year, OpenAI began working on the first version of a feature it would end up calling dreaming. 

Dreaming runs in the background, allowing ChatGPT to synthesize information from many different conversations without it relying on explicit instructions to remember something. “Over the last year, dreaming supplemented saved memories to create a step-function improvement in ChatGPT’s ability to personalize responses and offset the staleness of saved memories,” OpenAI explains. “However, it historically was never sufficient as a standalone memory system.”

That brings us to today’s release, which sees OpenAI rolling out what it describes as a new memory architecture that builds on the dreaming process to offer something that is “significantly” more capable and compute-efficient. Now, as ChatGPT synthesizes information about you, it will write a “memory summary” you can read at any time. From there, you can add and update information about yourself, as well as tell ChatGPT when it should reference what it knows about you and your preferences. “If you want to drill down into a particular area to learn more, just chat with the model,” says OpenAI. The new summary is designed to complement the memory sources feature OpenAI released alongside GPT-5.5 Instant. Sources allow you to see the information ChatGPT used to personalize an answer, and edit or delete that context as desired. You can see both features pictured above.

Besides offering greater visibility to the user, OpenAI says the new dreaming architecture is better at carrying forward context. For example, if you’ve talked to ChatGPT about photography in the past and mentioned the camera you currently use, the chatbot will know to generate tailored results the next time you ask for product recommendations that are compatible with your photography setup.

Similarly, OpenAI says the new architecture is better at following preferences. Say you’re planning a trip, ChatGPT will use things it has learned from past travel conversations to inform its responses. Using photography as example again, it might suggest a Singapore itinerary that includes suggestions on spots where you can do street photography. To bring everything together, ChatGPT will automatically revise its memories as time passes, so that it doesn’t do something like reference a trip you took in the past as if it were coming up.

OpenAI is beginning to roll out the new memory architecture to Plus and Pro users in the US starting today. Thanks to behind-the-scenes efficiency improvements, ChatGPT will, for the first time, soon start recording memories through the dreaming process for free accounts. As a Plus or Pro user, those enhancements will translate to ChatGPT offering greater memory capacity. The new architecture will roll out to users in other countries in the coming weeks.

Mira Murati isn’t a natural creature of the conference stage. As the CTO of OpenAI, she was present but rarely the public face of the company. As CEO of her own company, Thinking Machines Lab, she has been even harder to find. So when she sat down with Bloomberg in San Francisco on Thursday — her first major media appearance in roughly 18 months — it was worth paying attention, even if she was careful not to say too much.

The timing makes sense. Thinking Machines has spent the better part of a year and a half operating largely in the background: raising capital, hiring researchers, and shipping one product, Tinker, an API for fine-tuning open-source AI models.

In the meantime, the companies competing for the same talent, customers, and headlines have only grown more omnipresent. OpenAI, where Murati spent six years as CTO, is constantly in the news cycle. Anthropic’s momentum is all that anyone can talk about right now. And xAI, Elon Musk’s AI venture, has been folded into SpaceX ahead of what is expected to be its massive public offering, generating its own gravitational pull on attention and investment. In that environment, staying heads down has diminishing returns; at some point, you have to make some noise just to remind the market you exist.

Murati used the Bloomberg appearance to do exactly that and not much more. She previewed what Thinking Machines is calling “interaction models,” which she described as a fundamentally different kind of AI interface. Rather than the turn-based, prompt-and-response dynamic that defines most AI products today, she told interviewer Emily Chang, the company’s models are designed to process continuous streams of audio, text, and video in 200-millisecond intervals. The idea is that they can pick up on the texture of human communication — the interruptions, the mid-thought corrections, even pauses to think — in something closer to real time. But Murati was careful to frame it as a first step, not a finished product, and she declined to put a specific release date on anything.

She also answered questions about the episode that first put her more squarely in the public eye: the chaotic week in November 2023 when OpenAI’s board fired Sam Altman and she became interim CEO. Inside OpenAI it came to be called “the blip.” Murati said she felt clear about her decisions in each moment — that protecting the mission and the team was the through-line that made the choices feel obvious even as the situation appeared to be falling apart from the outside. She said the company would have “imploded” if not for her involvement through that strange five-day stretch and its immediate aftermath. But she acknowledged that clarity of intent is not the same thing as clarity about consequences. In retrospect, she said, she would have pushed harder for more information, a better transition plan, and more transparency. What she did not say, at least not directly, is whether she thinks things turned out well.

Asked whether she still trusts her former boss, she sidestepped the question, steering the conversation toward a larger concern that she returned to several times: the concentration of consequential decisions in too few hands — not just at OpenAI but across the industry. Her worry, she said, is less about the character of any individual leader (though she acknowledged that matters) and more about the absence of structural checks. Good people make bad calls. Well-intentioned organizations drift. Too much attention has been paid to virtue and too little to governance, she suggested.

Chang also politely pressed her on the departures of several high-profile researchers from Thinking Machines in recent months , a subject Murati has largely avoided in public and that she downplayed on Thursday. First, she said, building a frontier AI lab from scratch compresses years of normal organizational volatility into months. She also acknowledged that compensation — the nine-figure packages that have become standard currency in the war for AI talent — captures people’s imaginations, but she suggested it isn’t usually the whole story. To some audience laughter, she said of her own competitive instincts, “When I wake up in the morning, I am not thinking about how to kill the competitor.”

Naturally, Chang asked about what comes next for AI broadly, including for the humans who AI companies once said would be empowered by AI but who’ve more recently grown scared by talk of mass job displacement, not to mention a future where AI is used to create chemical weapons.

Murati, who was born in Albania and speaks with a slight Eastern European accent, was measured in her response. She pushed back on the framing of inevitable dystopia or inevitable utopia, arguing that neither outcome is predetermined and that the period we’re in right now is the one that will determine which way things go. Still, she said — and not for the first time during the interview — that if humans take their hands off the wheel too soon, the future will look very different, and not better.

Unless you’ve spent the last few years locked indoors and had all of your goods delivered to you — a not entirely implausible situation, given our audience — you’ve likely noticed the growing popularity of electronic shelf labels (ESLs). They’ve been a common sight in grocery stores like Aldi for some time, and major retailers such as Walmart and Home Depot have been expanding their use of the technology.

On the surface, it makes perfect sense. With electronic ink displays, you can create a price tag that looks enough like a paper label that the customer’s experience isn’t really any different, but the retailer doesn’t have to send somebody out to update the prices. Sure, the upfront cost is higher than a roll of sticky paper, but theoretically, the ESLs should pay for themselves thanks to the reduced labor costs.

It’s the sort of high-tech solution to a common problem that one of us would have come up with. If this were a decade ago, we wouldn’t have been surprised to see something like this get entered into the Hackaday Prize. It might have even won.

Now that the technology is becoming commonplace, there’s even more reason for hardware hackers to be interested in it. Since most of these tags will show whatever image you beam over to them via radio or infrared, we’ve seen a number of projects that repurpose second-hand tags as convenient data displays.

Rather than showing the price of milk, they can show the current price of Bitcoin. Or maybe you’d like to stick them up all over the house to display the weather forecast and your family calendar. They’ve been repurposed as badges at hacker cons, and at least one industrious hacker has used a discarded ESL to show an alert whenever a new episode of the Hackaday Podcast drops.

But not everyone is happy about ESLs. Recently, the United Food and Commercial Workers (UFCW) International Union released the results of a poll showing that most American consumers are opposed to ESLs, citing concerns that the technology would ultimately lead to higher prices.

With Great Power Comes…

The rejection of electronic shelf labels isn’t just about automation taking over a job that humans used to do, although that’s likely part of it. What’s got most consumers worried is what happens in the future once ESLs are the norm. There’s growing concern that the ability to rapidly and remotely update an item’s price will enable retailers to implement aggressive dynamic pricing schemes that were previously impractical. When you don’t have to send out a teenager with a price gun for each change, there’s nothing stopping stores from updating item prices every hour.

Things get really worrying when you consider the possibilities should the ESL system get tied into other data sources, and artificial intelligence be given free rein to virtually put its thumb on the scale. It’s not hard to imagine the price of umbrellas going up when it rains, or a premium being put on a particular team’s merchandise after they win a big game.

Such practices are referred to as “surveillance pricing”, and according to the UFCW poll, as many as 75% of respondents believe that one day stores might even attempt to tailor the price of an item to the individual. Like something out of Minority Report, the price tag could jump up when it detects a more affluent shopper passing by — or at least, one with a higher credit limit.

To those who may say this all sounds a bit far-fetched, the reality is that surveillance pricing is already here for many goods and services. Anyone who’s ever booked a hotel room can tell you that the price goes up and down based on demand, and rideshare services like Uber and Lyft have never hidden the fact that they adjust fare prices in real-time. Online retailers such as Amazon also routinely offer personalized “deals” based on your shopping habits or search activity, although whether or not you actually save any money in these scenarios is up for debate.

Electronic shelf labels don’t make surveillance pricing possible, since it’s already happening every day online. Rather, it enables retailers to use those same techniques in their brick-and-mortar stores in ways that weren’t possible before.

A Double-Edged Label

As hardware hackers, we love electronic shelf labels, if for no other reason than all those e-ink displays eventually trickling down to us. But the ability to change prices on a whim and without the need for human interaction is troubling, especially when considering the pricing schemes that are already so prevalent online. For better or for worse, we’ve become accustomed to dynamic pricing when we buy things on the Internet, but that doesn’t mean we have to accept as an eventuality that the same practices will eventually come to the grocery aisle.

So, Dear Reader, where do you fall on the subject? Are you excited about the technological implications of turning each price tag into a tiny remotely-controlled computing device, or does the potential for misuse outweigh the benefits? If so, do you think there’s a path forward that allows stores to take advantage of electronic shelf labels while protecting the consumer? Let us know in the comments.

The internet just crossed a remarkable threshold. Agentic AI internet traffic now exceeds that of real humans for the first time.

“Welp, that happened faster than I predicted,” Cloudflare CEO Matthew Prince said in a post on X on Wednesday. “Thought it would be [at the] end of 2027, then early 2027, but agentic traffic [is] growing so fast that bots have now passed human traffic online for the first time in the Internet’s history.”

He backed up his claim with a post to Cloudflare Radar, the company’s internet measurement system, showing that agentic bot usage is up to 57.4% of total traffic, while human traffic has dropped to 42.6%. 

Prince said in another post that the data is “a bit messy” but “clearly on the other side now,” indicating this is a trend that isn’t going away. 

These are not the bots you’re looking for

It is important to clarify what Prince refers to regarding web traffic. Regular bots, like search engine scrapers and web performance tools, eclipsed human internet traffic well over a decade ago. There are reports that those same bots exceeded human traffic on small websites even sooner, which led to a lot of small website owners exceeding their hosting usage limits faster than expected. 

The agentic bots Prince is referring to are the systems that search the internet on your behalf when you ask an AI chatbot a question and return the results. Those searches and visits generate real web traffic, even if it doesn’t look that way from your AI chat window. The data means that more AI agents are visiting these webpages than real humans. Humans still physically engage with content more than AI does, but AI visits webpages more often. 

Digging into the data

The above numbers reflect worldwide traffic patterns, but they differ by region. North America as a whole skews more toward bot usage, with bots accounting for 68.6% of activity and humans 31.4%. If you zoom in on the American Midwest, the trend reverses, with humans leading at 54.5% versus 45.5% for bots. The trend is consistent across regions: Broader areas tend to be dominated by agentic bot traffic, while smaller areas within those regions often still show higher levels of human usage.

There are some outliers as well. During peak hours, up to 97% of traffic originating from tiny Gibraltar is bot traffic. Other countries, like Cuba and Laos, sit at the other end of the spectrum, with 80.8% and 84.7% of each country’s traffic coming from human users, respectively. 

North America, Europe and Africa lean toward bots, while Asia, South America and Oceania still see more human internet use most of the time. 

Dead Internet Theory

Interest in something called Dead Internet Theory has increased in recent years, fueled by perceptions that online activity is becoming less human-driven.

The idea behind Dead Internet Theory is that bots and AI generate most of the internet’s activity. The theory seemed far-fetched to many when it emerged in the late 2010s, but it’s becoming harder to argue against as data like Cloudflare’s becomes public. 

The implications become more concerning with additional context: Forty percent of Facebook posts are estimated to be generated by bots. Music-streaming service Deezer announced in April that 44% of new music uploaded to its platform is now AI-generated. And a report from Axios posits that AI generates 52% of all online articles (though not this one — honest).