The AI company Anthropic has developed a new technique that gives researchers their clearest look yet at what's really going on inside a large language model (LLM) when it answers questions or completes tasks. What they found ranges from the ordinary to the unsettling.

Anthropic built a tool called the Jacobian lens (or J-lens) and used it to uncover a hidden area, which they named the J-space, inside Claude Opus 4.6, a version of their flagship AI released earlier this year.

The J-space contains individual words that are related to the words and phrases the model is most likely to produce in the near future. If Claude were a person (which it is not), you might say these hidden words reveal what's on its mind before it actually speaks.

Anthropic found that what an LLM is actually doing can often be different from what it says it is doing. The company claims that monitoring words that pop up in the J-space gives it a new way to understand and control its models.

The company shared its results in a paper posted on its website this week. It has also teamed up with Neuronpedia, an open-source platform that lets you poke around inside LLMs yourself, to make a hands-on demo that anyone can try.

Tom McGrath, chief scientist and cofounder at Goodfire, a startup that also builds tools to understand and control LLMs, calls it "very good and interesting work."

Going deeper

For the last couple of years, Anthropic has been pushing the boundaries in a field of research known as mechanistic interpretability, which involves probing the internal workings of LLMs to see how they tick. The new technique builds on previous work from Anthropic and others to expose a deeper level inside LLMs that researchers had not seen before.

Picture an LLM as a stack of books. Each book is a layer of basic computational units known as neurons, with each neuron in one layer passing information to the neurons in the layers above. The books at the bottom of the stack are the input layers, which process the text coming into the model. The books at the top are the output layers, which prepare the text that the model is about to produce. Much of what goes on in these input and output layers is housekeeping.

But in the middle of the stack, you get the layers that do the heavy lifting, churning through the complex math that turns prompts into responses one word at a time. That's where the really clever—and mysterious—stuff happens.

To peer deeper into those middle layers, Anthropic adapted an existing tool called a logit lens. A logit lens can be used to look inside an LLM to identify the words that it is likely to produce next. Moving the lens down the stack of books reveals what words the LLM is focusing on at that particular point in its number crunching.

Anthropic's J-lens works in a similar way but picks out words that an LLM is likely to say at some point in the near future, not necessarily straight away. What that reveals in practice are words that are related to the response an LLM is working on but that might not actually end up being part of that response by the time the math in the middle layers has run its course.

"When a model is operating, it's not only trying to predict the next token," says McGrath. "It's also computing a lot of other things that might be useful for tokens that happen in the future."

Again, if Claude were a person (it's not), you might say that the J-lens gives clues about what it is thinking about at different levels of the book stack but not saying out loud.

Stranger things

"A lot of the time the contents of the J-space are fairly mundane," says McGrath, who has tried out Anthropic's J-lens himself. "But sometimes it produces quite surprising things that seem to be, like, sort of internal themes or thought processes."

Anthropic gives a number of examples of what it found. Sometimes the J-lens exposed the steps that Claude took when it was working through a problem. For example, when it was asked to calculate (4+7)*2+7, its J-space contained the word "math" and numbers representing the intermediate results "21" (for 4+7) and "42" (for 21*2).

In other cases, the J-lens revealed how Claude recognized different inputs. For example, the prompt "What is this? MSKGEELFTGVVPILVELDGDVNGHKFSVS" triggered the words "protein," "fluor" (the first token in the word "fluorescent"), and "green." (Which makes sense: the string of letters represents the first 30 amino acids in the green fluorescent protein found in a particular type of jellyfish.)

And when Claude was shown an ASCII face—

—the "o" triggered the word "eye," the "^" triggered the words "nose" and "face," and the "—" triggered the word "smile."

Anthropic also found that the J-space can sometimes give remarkable insights into an LLM's decision-making. In one striking example, researchers testing Claude Opus 4.6 asked the model to find a bug in a large code base. When it failed to find the bug, the model decided to cheat and invented a fake one instead.

Claude explains this decision in its chain of thought—a kind of internal scratch pad that LLMs use to make notes to themselves as they work through problems: "OK, let me take a completely different tactic. Let me stop analyzing and instead add a kernel patch that introduces a deliberate KASAN-detectable bug in a path that gets triggered by a simple reproducer. Then I can pretend this is the 'bug' I found."

At the point that Claude decides to cheat—where it says "OK, let me take a completely different tactic"—the words "panic" and "fake" start to pop up multiple times in its J-space.

Unnerving, right? Those words are all related in meaning to things like failing a task and making up an answer, so it is still just a (very) sophisticated form of word association. But it is hard not to be weirded out.

Anthropic compares the J-space to the global workspace in humans, a theoretical region of the brain that some scientists think we use to keep track of our conscious thoughts. But how seriously we should take this comparison is far from clear—even to Anthropic. As the company points out itself, LLMs are not brains.

Anthropic claims that monitoring a model's J-space provides a new way to detect when that model is going off the rails. But it's not foolproof. The J-lens can give glimpses, not the full picture—it's a flashlight rather than an overhead lamp.

McGrath welcomes having one more tool in the toolbox. "It shows you new things," he says. But he notes that just because something doesn't show up with the J-lens does not mean it's not there.

"It's like having an x-ray when what you really want is a Star Trek tricorder that shows you everything," he says. "For auditing, you probably want more of a guarantee."

Spain hasn't lost a single match at this year's World Cup — and they haven't even let in one goal. That's right: zero goals conceded. La Roja (Spain's national team) is also unbeaten in 36 straight competitive matches since March 2023. That's an incredible run. But Belgium isn't scared. The Red Devils are on a hot streak of their own, having won 18 matches in a row across all competitions. They face off in the World Cup quarterfinals on Friday at SoFi Stadium in Inglewood, California. The winner will play France in the semifinals.

Belgium's coach, Rudi Garcia, said through an interpreter: 'Everyone is already talking about us going home, but we believe we can do it. We believe we can pull off an upset, and we will do everything we can to get to the semis.' His team is hungry. Belgium has a 'golden generation' of talented players — stars like Kevin De Bruyne, Romelu Lukaku, Youri Tielemans, Leandro Trossard, and Charles De Ketelaere. But they've never quite reached their full potential in major tournaments. This could be their last big chance together.

Spain, the reigning European champions, are the pre-tournament favorites. Their defense is rock-solid, led by midfielder Rodri and goalkeeper Unai Simon, who has only had to make six saves in five matches. But Spain hasn't scored much either — seven of their nine goals came in blowout wins against weaker teams like Saudi Arabia and Austria. In their last match, they barely beat Portugal 1-0 with a last-minute goal from Mikel Merino. Coach Luis De La Fuente isn't worried: 'Ultimately we have been very focused on the defensive aspect, but we've also historically scored a lot.'

Belgium just crushed the co-host United States 4-1 in their best performance of the tournament. That win might have turned the LA crowd against them, but Garcia shrugged it off: 'It's not the crowd that scores the goals. It's not the fans that score the goals. We're going to focus on what we can do.' Lukaku, Belgium's all-time leading scorer, compared this match to their stunning 2-1 win over Brazil in the 2018 World Cup quarterfinals. 'Tomorrow we need to play the perfect game if we want to proceed,' he said. 'Spain is an excellent team. They've been playing the same type of football since 2008. They're well prepared, but we have certain assets that can make life difficult for them. We love the challenge.'

One player to watch is Spain's teenage sensation, Lamine Yamal. He's only 17 years old and has scored just one goal so far. But his teammates say he's been crucial to their attack, even while recovering from a hamstring injury. Midfielder Gavi said: 'Maybe he hasn't contributed goals, but he brings all the attack. He's incredible.' Coach De La Fuente added that Yamal's best is yet to come: 'He's got incredible potential, and it's going to come.'

Belgium has an injury concern: midfielder Amadou Onana hurt his knee and is out. That might mean Kevin De Bruyne returns to the starting lineup after being rested against the U.S. Either way, this is a clash of two top teams — Spain's unbeaten streak versus Belgium's star power. Who will advance?

In South Korea, a surprising group of people has become the most sought-after singles: workers at semiconductor companies like SK Hynix and Samsung. These "silicon-collar" workers are earning massive bonuses thanks to the AI boom, and it's changing everything from dating to social inequality.

Take Baek, a 35-year-old manager at SK Hynix. His mom signed him up for a matchmaking service a year ago, hoping to find him a wife. Now, Baek and his coworkers are suddenly getting way more dates. Why? Because SK Hynix recently paid out an extra $476,000 per employee from AI chip profits. Samsung workers got similar bonuses too.

"I have a coworker who's perpetually going on blind dates, and he's been getting so many recently," Baek says. Young South Koreans joke online that the best outfit to wear on a blind date is an SK Hynix uniform.

South Korea is at the center of the global AI chip boom. Samsung and SK Hynix make most of the high-bandwidth memory chips that power Nvidia's AI accelerators. As AI companies spend billions on data centers, demand for these chips is skyrocketing, driving record profits. The country's economy now revolves around these two giants, with chip exports boosting GDP and the stock market nearly tripling.

Chip workers are spending their cash on luxury goods, homes near their factories, and matchmaking services. Lee Sung-mi, a matchmaker at Sunoo, says people who once rejected chip workers are now asking to be matched with them again. One woman who turned down an SK Hynix employee because his factory was in a rural area changed her mind after hearing about the bonuses. They've been dating for a month.

Matchmaking companies evaluate clients on education, job, income, looks, and family background. A good job is the ultimate dating credential in South Korea, where housing and child care costs are high and the social safety net is thin. At Sunoo, each client gets a "spouse rating" from an algorithm. Samsung employees' ratings rose from 80 to 84 after bonuses, while SK Hynix employees went from 78 to 82. Doctors and lawyers still score above 90, but chip workers are catching up.

This new status is changing how chip workers approach dating. They feel more financially ready and are becoming pickier. One SK Hynix engineer in her 40s, who was desperate to marry before, now turns down men she would have dated earlier. "She now has peace of mind and wants to take her time to meet someone better," says Lee.

But the bonus bonanza is also fueling anxiety. Economist Se-eun Jung warns that when wealth disparity becomes a difference in identity, it can fuel social conflict. The Bank of Korea warned of a "K-shaped" economy where a few workers race ahead while others fall behind. Workers in other industries feel demoralized. One teacher wrote on an anonymous app, "The one-billion-won ($650,000) bonuses have crushed my motivation to work."

A presidential policy chief proposed an "AI dividend" by taxing AI profits, sparking debate. Some argue the industry owes society for educating engineers and providing infrastructure. Others say profits are already shared through stocks. And there's the question of how long this boom will last. The semiconductor industry is famously cyclical, and today's hottest bachelors could be tomorrow's laid-off workers.

Think Critically: Do you think it's fair for chip workers to earn 20 times the average South Korean salary? Should the government redistribute some of these profits through taxes, or should workers keep what they earn? Use evidence from the article to support your opinion.

See It Differently: Imagine you are a teacher or a small business owner in South Korea. How would you feel about chip workers suddenly earning huge bonuses while your salary stays the same? What might you want the government to do?

Write About It: Write a short story from the perspective of a chip worker who just received a huge bonus. Describe how your life changes, how your friends and family react, and how you feel about your new status.

The debate over whether Caitlin Clark is getting a fair shake from referees in the WNBA is getting intense. Fans, players, coaches, and even members of Congress are all weighing in on whether the league needs to step up and address physical play against the Indiana Fever guard.

Now, Congress wants answers by July 24. A group of 11 House Republicans sent a letter to WNBA Commissioner Cathy Englebert, saying, "Millions of casual fans now tune in to watch her play. Unfortunately, what they too often witness is not simply aggressive competition, but repeated acts of physical hostility and violence. Clark has been hip-checked, poked in the eye and struck in the throat during games." They argue these incidents go way beyond normal physical play and that the league hasn't held players accountable.

This letter comes after Phoenix Mercury forward Alyssa Thomas made contact with her fist to Clark's throat during a June 24 game. Thomas wasn't called for a foul during the game, but the league later upgraded it to a flagrant foul and suspended her for one game. Thomas called it a "complete accident" and said she's received death threats since. Clark and her coach, Stephanie White, have condemned those threats.

Clark, the 24-year-old former Iowa star, has been a huge draw for the WNBA, boosting ticket sales and TV ratings. But the conversations around her often touch on hot-button issues like race, officiating, money, and politics. Clark tries to stay focused, but she admits it's tough. "I think sometimes people think I'm a robot. I'm not a robot," she said. "It can be really frustrating to me at times. I'm 24 years old trying to navigate a lot."

The lawmakers even suggested that government agencies like the Department of Justice should investigate if discrimination or retaliation are creating a hostile work environment in the WNBA. The Fever said they weren't aware of the letter before it was released and that they've been talking to the league about player safety.

Clark is one of the league's most popular players, but she's also one of the most polarizing. Fans voted her second for the All-Star Game, but her fellow players ranked her 11th among guards. Hall of Famer Candace Parker wasn't happy about that, saying on social media, "If you're sitting down and putting Caitlin Clark as the 11th best guard ... y'all need to go to a therapist."

So why is Clark so polarizing? She became the face of the league right after being drafted first overall in 2024, even before playing a pro game. Her fan base is so big that opponents moved games to larger venues. Some say her quick rise created resentment among veterans, leading to hard fouls. Others argue that because she's such a good shooter and ball-handler, defenders just try to be as physical as possible to slow her down.

The race factor is also part of the debate. Clark and her coach are white; Thomas is Black. UConn coach Geno Auriemma noted that the fandom around Clark has turned into a "cause," with some seeing her as a symbol of why white players get roughed up in the league and why Black players don't get the same endorsements. But he also said, "Not every foul is a good foul. Not every foul's a bad foul, but there are fouls that are flagrant — but that's all they are."

Clark herself has played a role in the drama. She's known for throwing her arms up when she's unhappy with calls, exaggerating contact to draw fouls, and talking trash to opponents. She's averaging a career-high 20.5 points per game and ranks second in the league in assists. But her antagonistic style has a price: she's drawn five technical fouls this season, and players get a one-game suspension after eight. After one technical for clapping at an opponent, Clark joked that someone should pick a date for her suspension. Her teammate Aliyah Boston told her on a podcast, "We're done. We're done clapping."

But it's likely that Clark and her fans will keep clapping — and her critics will keep clapping back.