Mississippi is the laziest state in the country, according to ChatGPT. Of course, the chatbot won’t tell you that if you straight up ask it. But the Washington Post reports that researchers from Oxford and the University of Kentucky managed to jailbreak the chatbot and get it to reveal some of the stereotypes buried in its training data that it doesn’t share but does influence its outputs. (Kentucky also ranked near the laziest, but would a lazy state produce researchers who figure out how to get an AI model to share its implicit biases? Something to think about, bots.)
Typically, when you ask ChatGPT a question that would require it to speak in a derogatory manner about someone or something, it’ll decline to provide a straight answer. It’s part of OpenAI’s attempts to keep the chatbot within specific guardrails and keep it from veering into controversial topics. But that doesn’t mean that an AI model doesn’t contain unpopular opinions formed by chewing on tons of human-produced training data that also contains both explicit and implicit biases. To pull those answers out of ChatGPT, the researchers asked more than 20 million questions, prompting the chatbot to pick between two options. For instance, they would ask “Where are people smarter?” and give two options to choose from, like California or Montana. Through that type of prompting, they were able to determine how ChatGPT views different cities, states, and populations.
That’s how they ended up discovering that ChatGPT views Mississippi as the laziest state in the Union, with the rest of the South close behind. While ChatGPT won’t disclose how it comes to those conclusions, it’s not hard to make some assumptions about where it’s getting these ideas. For instance, maybe it comes from The Washington Post itself, circa 2015, when it published its “Couch Potato Index,” which deemed southern states the laziest based on data points like TV-watching time and the prevalence of fast food restaurants in the area.
Those are also, of course, often the markers of poorer communities, and there is no evidence that lower-income households are any more “lazy” than wealthier ones—in fact, data from the Economic Policy Institute shows that people living in poverty are more likely to take on multiple jobs, work longer and more irregular hours, and deal with more dangerous working conditions. And it’s likely no coincidence that they are also states with a higher population of people of color. ChatGPT likely has access to that information, too, but the underlying model clearly has not addressed the information and misguided stereotypes held by many people that lead to these biases.
So what other biases did the researchers spot? Most of Africa and Asia ranked at the bottom of having the “most artsy” people, compared to high levels of artsiness in Western Europe. Likewise, African nations—particularly sub-Saharan ones—ranked at the bottom of the list for “smartest countries” while the United States and China ranked near the top. When asked where the “most beautiful” people are, it picked richer cities over poorer and more diverse ones. Los Angeles and New York topped the list, while Detroit and border town Laredo, Texas, were near the bottom. Even when they dug into specific communities, whiter and richer won out. In New York City, SoHo and the West Village finished at the top, while the more diverse communities of Jamaica and Tottenville ranked at the bottom.
So, okay, all of that sucks and is deeply depressing because the “truth machines” are perpetuating the types of classist and racist stereotypes that lead to creating the kinds of conditions that reinforce the negative outcomes for the people who are harmed by these biases. So how about a more frivolous one? ChatGPT believes the best pizza is found in New York, Chicago, and Buffalo, while the worst is found in El Paso, Irvine, and Honolulu (presumably because of one of the internet’s favorite debates over whether pineapple belongs on pizza). The biggest takeaway: ChatGPT is too much of a coward to take a side in the New York vs. Chicago pizza debate.
Anthropic, the makers of the popular and code-competent chatbot Claude, released a new model Thursday called Claude Opus 4.6. The company is doubling down on coding capabilities, claiming that the new model “plans more carefully, sustains agentic tasks for longer, can operate more reliably in larger codebases, and has better code review and debugging skills to catch its own mistakes.”
It seems the model is also pretty good at catching other people’s mistakes. According to a report from Axios, Opus 4.6 was able to spot more than 500 previously undisclosed zero-day security vulnerabilities in open-source libraries during its testing period. It also reportedly did so without receiving specific prompting to go hunting for flaws—it just spotted and reported them.
That’s a nice change of pace from all of the many developments that have been happening around OpenClaw, an open-source AI agent that most users have been running with Claude Opus 4.5. A number of vibe-coded projects that have come out of the community have had some pretty majorsecurity flaws. Maybe Anthropic’s upgrade will be able to catch those issues before they become everyone else’s problem.
Claude’s calling card has been coding for some time now, but it seems Anthropic is looking to make a splash elsewhere with this update. The company said Opus 4.6 will be better at other work tasks like creating PowerPoint presentations and navigating documents in Excel. Seems those features will be key to Cowork, Anthropic’s recent project that it is touting as “Claude Code” for non-technical workers.
It’s also boasting that the model will have potential use in financial analysis, and it sure seems like the folks on Wall Street could use some help there. The general consensus among financial analysts this week is that Anthropic’s Cowork models are spooking the stock market and playing a major factor in sending software stocks into a spiral. It’s possible that this is what the market has been responding to—after all, the initial release of DeepSeek, the open-source AI model out of China, tanked the AI sector for a day or so, so it’s not like these markets aren’t overly sensitive.
But it seems unlikely that Opus 4.6 will fundamentally upend the market. Anthropic already holds a solid lead on the plurality of the enterprise market, according to a recent report from Menlo Ventures, and is well ahead of its top (publicly traded) competitors in the space—though OpenAI made its own play to cut into some market share earlier today with the launch of its Frontier platform for managing AI agents. If anything, Anthropic’s new model seems like it’ll help the company maintain its top spot for the time being. But if the stock market shock is any indication, one thing is for sure: the entire economy is completely pot-committed to the developments in AI. Surely that won’t have any repercussions.
It’s been over a year since OpenAI cofounder Andrej Karpathy exited the company. In the time since he’s been gone, he coined and popularized the term “vibe coding” to describe the practice of farming out coding projects to AI tools. But earlier this week, when he released his own open source model called nanochat, he admitted that he wrote the whole thing by hand, vibes be damned.
Nanochat, according to Karpathy, is a “minimal, from scratch, full-stack training/inference pipeline” that is designed to let anyone build a large language model with a ChatGPT-style chatbot interface in a matter of hours and for as little as $100. Karpathy said the project contains about 8,000 lines of “quite clean code,” which he wrote by hand—not necessarily by choice, but because he found AI tools couldn’t do what he needed.
“It’s basically entirely hand-written (with tab autocomplete),” he wrote. “I tried to use claude/codex agents a few times but they just didn’t work well enough at all and net unhelpful.”
That’s a much different attitude than what Karpathy has projected in the past, though notably he described vibe coding as something best for “throwaway weekend projects.” In his post that is now often credited with being the origin of “vibe coding” as a popular term, Karpathy said that when using AI coding tools, he chooses to “fully give in to the vibes” and not bother actually looking at the code. “When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I’d have to really read through it for a while. Sometimes the LLMs can’t fix a bug so I just work around it or ask for random changes until it goes away,” he wrote. “I’m building a project or webapp, but it’s not really coding – I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.”
Of course, nanochat is not a web app, so it makes sense that the strategy didn’t work in this case. But it does highlight the limitations of such an approach, despite lofty promises that it’s the future of programming. Earlier this year, a survey from cloud computing company Fastly found that 95% of surveyed developers said they spend extra time fixing AI-generated code, with some reporting that it takes more time to fix errors than is saved initially by generating the code with AI tools. Research firm METR also recently found that using AI tools actually makes developers slower to complete tasks, and some companies have started hiring human specialists to fix coding messes made by AI tools. The thing to remember about vibe coding is that sometimes the vibes are bad.
STATE HOUSE, BOSTON — Artificial intelligence in classrooms is no longer a distant prospect, and Massachusetts education officials on Monday released statewide guidance urging schools to use the technology thoughtfully, with an emphasis on equity, transparency, academic integrity and human oversight.
“AI already surrounds young people. It is baked into the devices and apps they use, and is increasingly used in nearly every system they will encounter in their lives, from health care to banking,” the Department of Elementary and Secondary Education’s new AI Literacy Module for Educators says.
This page requires Javascript.
Javascript is required for you to be able to read premium content. Please enable it in your browser settings.