As Sullivan recalls in ​a recent newsletter​, she attended an optional Catholic Mass the first morning, held in an ancient church. She was surprised to see one of the tech leaders sitting a few rows away in the pews. “[This was] the kind of guy you typically see in a black t-shirt and chinos,” she writes. “That morning he was dressed in a brown suit and tie, quietly taking in the sanctuary as the first rays of morning light filled the room.”

After the service concluded, they chatted. The tech leader, it turned out, was not Catholic. When Sullivan asked him why he was here, he gave the following answer:

“We’re building something that is going to change life as we know it. I want to make sure I keep in touch with what humans have always cared about. This is a place that takes care of those values.”

I found this interaction chilling. Not because of what it says about potential AI disruptions, but because of what it tells us about the engineers developing this technology.

The AI industry is suffused with a religious fervor. As Elizabeth Lopatto ​recently pointed out​, they’ve stopped trying to build useful products and are focusing instead on “inventing the future.” This puts these companies in the dual role of priest and prophet, frantically trying to appease the digital deity they imagine they’ve summoned, all the while warning the masses of its impending holy wrath.

I understand this attraction. This all must be terribly exciting and life-affirming for that executive in his brown suit. He gets to be Aaron and Amos all at once.

But not everyone is willing to go along with this game…

Last week, Pope Leo XIV released a 42,000-word encyclical, ​”Magnifica Humanitas,”​ in response to the challenges of artificial intelligence. I’m still digesting the full document, but ​early summaries​ indicate that the Pope is not ready to meekly acquiesce to the AI future that we’ve been told is inevitable.

Consider the following key exhortation from the encyclical:

“With the heart of a shepherd and a father, I ask everyone to abandon the construction of yet another Tower of Babel and to join forces in building up the common good, so that humanity will never lose its beauty, and the world once again will come to recognize the human heart as the place where God desires to dwell.”

Tools should be useful. Tools should help people’s lives and build up the “common good.” Tools are what technology companies should seek to create.

When AI leaders resignedly shake their heads, and talk about the need for the government to provide ​guaranteed income​ once their AI models automate all work, or eagerly describe a future in which we live happily alongside ​ “machines of loving grace​,” this is not forward-looking pragmatism; it’s hubris. A new Tower of Babel built out of GPUs.

Thankfully, in recent weeks, there has been a marked shift in how technology executives talk about AI. Nvidia CEO Jensen Huang ​called BS​ on executives claiming they’re laying people off due to AI, calling the excuse “lazy” and “just a way for them to sound smart.” Perhaps even more surprising, just last week, Sam Altman ​admitted​ he had been “pretty wrong” about his previous predictions that AI would automate large numbers of jobs.

These shifts in tone are likely more about PR damage control than a legitimate change of heart. But it’s still good to see. Leave the religion to the Pope; I want my technology executives focused on building things people actually want.

###

As An Aside

I think it’s telling that one of the articles I read summarizing the Pope’s new encyclical casually and confidently made the following points:

• “A.I. has already displaced many entry-level jobs.”
• “Mass automation of both white collar work and blue collar work is likely to significantly reshape most sectors of the labor market.”

This is pretty shocking, as neither of these claims is widely accepted. And now, as mentioned, many leading AI CEOs have come out to push back on these specific talking points.

I’m not sure that tech leaders realize just how much anxiety and fear they spread during the period in which they were cosplaying as solemn x-risk sages. This is a p(doom) genie that might take a while to put back in its bottle.

The post The Pope vs. Silicon Valley appeared first on Cal Newport.

There are different ways to translate the Hebrew term rendered here as “living being.” Onkelos, a Roman nobleman turned Torah scholar, interpreted it to mean “speaking spirit” (in Aramaic: ruach memalela). As Rabbi Shai Held elaborates in the second volume of ​The Heart of the Torah​, Onkelos’s translation implies that “speech is constitutive of what it means to be a human – a core part of our humanity is our ability to communicate with words.”

Onkelos is not the only scholar to explore this concept. As Held also explains:

“According to [the medieval commentator] Rashi, speech is thus not only central to who we are as human beings, it is also key to our uniqueness. Alone among God’s creations, Jewish tradition affirms, human beings are capable of speech.”

The original scribes of the Hebrew Bible were reflecting the central role words have played in the post-Paleolithic human experience. They allow us to alchemize our state of mind into vocalized phonemes or written letters that can then be decoded in the mind of the receiver – a miraculous and intimate act of human-to-human telepathy. It was also arguably the rise of alphabet systems in the ancient Near East and the literacy they enabled that democratized holiness and, in doing so, birthed the notions of human dignity and universal justice that we take for granted today. Speech matters.

I’ve been thinking about this recently as I grapple with some of the consequences of the generative AI revolution. Like many people, I feel a general sense of unease when holding a fluent conversation with a chatbot. I understand that this lexical fluidity is merely an illusion – the result of endless matrix multiplications, autoregressively generating one token after another – but it still creates a discomforting sense of transgression.

The sacredness of speech helps explain these feelings. It also raises some deep questions.

Should we be so quick to extend the role of ruach memalela to machines, allowing them, seemingly all at once, to become active participants in a ritual so defining to our experience? Should we let AI write and speak on our behalf, or serve as a golemic conversation partner when authentic human companionship isn’t readily available? Something about this, for lack of a better word, feels profane.

I don’t have definitive answers here. But one thing that seems clear is that the newly emerging field of digital ethics (in which ​I’m currently active)​ is in the same place today as bioethics was five decades earlier, when new medical technologies began to force tough moral quandaries.

Which is all to say: before we blindly embrace whatever AI product Sam Altman or Dario Amodei declares to be inevitable, we still have a lot of work to do in figuring out ​what we’re willing to accept​.

The post On God and LLMs appeared first on Cal Newport.

Jevons first articulated this idea in an 1865 book, pithily titled, ​The Coal Question: An Inquiry Concerning the Progress of the Nation, and the Probable Exhaustion of Our Coal-Mines​. He argued that building more efficient steam engines – ones that required less fuel to generate the same power – would not solve the problem of England’s diminishing coal supplies. If you made the engines more efficient, Jevons predicted, people would find more applications for steam power, and even more coal would be burned overall.

This is indeed what happened. (At least, the part about increased coal consumption. The feared coal shortage was averted through new mining techniques.)

The Jevons Paradox is popular again because it provides a useful frame for understanding the potential impact of AI on jobs. Many fear that this technology will make workers so efficient that the labor market will shrink. If one programmer can now do the work of five, then companies will fire 80% of their programmers!

The Jevons Paradox implies the opposite might occur. If you make workers more efficient, their output will become cheaper, and the demand for their services might grow. If one programmer can now do the work of five, the effective cost of creating software will become so cheap that many more individuals and organizations will now pay to develop their own tools and applications.

This is a fascinating prediction that’s worth keeping an eye on. (For a deeper dive into the counterintuitive economics of AI, I recommend Derek Thompson’s ​recent interview​ with Alex Imas.) But there’s also a darker side to the Jevons Paradox that hasn’t been discussed as much recently: suddenly increasing demand for a resource can create unexpected negative side effects.

More efficient steam engines, for example, led to soot-stained buildings and the smoky start to the era of human-driven climate change. More recently, in the context of knowledge work, the arrival of digital communication tools such as email and Slack created similar unanticipated problems. By making communication significantly more efficient, the demand for fast interaction exploded, leading to our current moment in which the average knowledge worker is now interrupted once every ​two minutes​. (For more on how this descent into communication madness occurred, check out my 2021 bestseller, ​A World Without Email​.)

If AI ends up making certain types of workers more efficient, I hope the Jevons Paradox holds, as it’s better than the alternative of labor market contraction. But we need to remain vigilant about its side effects. It’s tempting to assume that increasing efficiency, in any context, can only make things better, but economic history has often told a more complicated tale.

The post The Dark Side of the Jevons Paradox appeared first on Cal Newport.

“If you are an editor or reviewer at a journal these days, you probably already know this. The manuscripts are arriving in greater volume, with a particular feel that is hard to pin down. On the surface, the papers look the same as ever, but the writing feels weightless in a way that rarely describes academic writing…you find yourself scratching your head at the meaning the words are trying to convey.”

The culprit? The task force crunched the numbers and produced a clear answer. Starting in 2023, after ChatGPT became available, the number of submissions to Organization Science rapidly increased. At the same time, the percentage of submissions classified as using minimal AI has plummeted from near 100% down closer to 30%.

The impact of this shift on readability has been marked, with scores on a standard “reading ease” metric falling by 1.28 standard deviations between January 2021 and January 2026:

“Submissions have become far harder to read,” the Task Force reports. “This is counterintuitive. Most people assume that AI produces cleaner, more polished text. And in some narrow dimensions, it does…but on the measures that capture whether a reader can actually parse and absorb the prose, AI writing is worse…[using] longer words, more complex sentence structures, more jargon, and more nominalizations.”

Papers that are more difficult to read might be worth it if AI increased the amount of good science being produced. But this doesn’t seem to be the case. Organization Science is desk-rejecting (e.g., rejecting a paper before even sending it to peer reviewers) nearly 70% of manuscripts that made heavy use of AI. This number drops to 44% for papers written without AI.

Similarly, only 3.2% of high-AI papers are ultimately accepted compared to 12% of low-AI papers.

(It’s important to note here that the editors making these decisions do not themselves know the role of AI in the paper construction. These are retrospective analyses.)

All of this points to a distressing conclusion: generative AI tools are leading to many more poor paper submissions, which are taxing the time and patience of the community tasked with reviewing this research.

These tools make individual researchers’ lives easier in the moment (writing is hard!), but they are leading to worse outcomes for the field as a whole.

I tell this story because I think it’s a useful cautionary tale about AI. As I’ve been trying to argue from many different angles in recent weeks (e.g., ​1​ ​2​), making things faster or easier is not the same as making things better.

Sometimes there really is no shortcut to taking your time.

The post Easy is Overrated appeared first on Cal Newport.

There was one chapter in particular, however, that captured my attention as being uniquely well-suited to the themes we discuss here. It focused on the ideas of a somewhat eccentric physicist-turned-management guru named Eliyahu Goldratt, who in the 1980s popularized a framework for understanding industrial productivity that he dubbed the “theory of constraints.”

Here’s how a non-profit established to promote Goldratt’s work summarizes it:

“Every system has a limiting factor or constraint. Focusing improvement efforts to better utilize this constraint is normally the fastest and most effective way to improve profitability.”

To borrow one of Goldratt’s examples, imagine you run a small assembly line that manufactures chicken coops following a step-by-step process – building the frame, attaching the roof, adding wire mesh, etc. Goldratt notes that the speed of this production is limited by whatever step is slowest; what he calls the “bottleneck.”

Speeding up other steps of the process won’t increase the rate at which you produce chicken coops, as the bottleneck still determines the overall efficiency. If, for example, putting on the roof is the slowest step, then adding more workers or better tools to earlier steps will lead to more partially-constructed coops piling up at the roofing station. To speed up the line, you need to move more resources to the weakest link.

Goldratt was primarily concerned with industrial production, but I think his theory of constraints provides insight into personal productivity, too.

Something I’ve long written about is the reality that many digital productivity tools paradoxically make us busier, rather than better at our jobs. Goldratt’s theory helps explain why.

When we deploy a digital tool like email to speed up communication, or generative AI to create (sloppy) slide presentations quickly, we don’t automatically become better at our jobs. If these steps don’t improve the bottleneck in our process – the key link where the real value is produced – then, as in the chicken coop example, they’re just as likely to create pile-ups and distraction, without actually boosting our true productivity.

This helps explain why ​email ended up an accidental disaster​, and the early returns on ​AI office tools have been mixed​ at best.

The theory of constraints implies a different way of thinking about getting better at our jobs. Don’t seek speed, or efficiency, or the avoidance of hard things. What ultimately matters more than anything else is how well we perform the deep steps that actually move the needle.

The post On Bottlenecks and Productivity appeared first on Cal Newport.

“Within recent memory, people who made software and hardware understood their job was to serve their customers. It was to identify a need, and then fill it,” she writes. “But at some point following the financial crisis, would-be entrepreneurs got it into their heads that their job was to invent the future, and consumers’ job was to go along with that invented future.”

I certainly noticed this shift when it first began emerging. See, for example, my 2015 article titled, ​“It’s Not Your Job to Figure Out Why an Apple Watch Might Be Useful.”​ But it really picked up speed in the last half-decade. Here’s Lopatto with a needle-sharp summary of our current status quo:

“In the place of problem-solving technology, companies have jumped on successive bandwagons like NFTs, the metaverse, and large language models. What these all have in common is that they are not built to really solve a market problem. They are built to make VCs and companies rich.”

Of these three examples, large language models clearly have the most potential utility. But this doesn’t let AI companies off the hook when it comes to figuring out and communicating those uses.

As Lopatto points out: “Normal people aren’t running around like chickens with their heads cut off, trying to automate every single part of their lives.“ Their biggest exposure to AI is using a tool like ChatGPT as a more verbose Google, or perhaps occasionally formatting an event itinerary. This is cool, and even useful, but at the moment it is probably less positively impactful in their lives than, say, the arrival of the iPod in the early 2000s.

But unlike an iPod, these same ordinary users are forced to hear about AI constantly; not just enthusiast tech bro nonsense, but dark, disturbing, relentless accounts about how everything is about to change in terrible ways that they can’t control.

This isn’t sustainable.

Generative AI has no shortage of ways that it might, with care, be shaped into genuinely useful products, but this shaping needs to actually happen before the hyper-scalers earn the right to continually harass the psyche of billions of people with breathless pronouncements. Most people don’t care that GPT 5.5, released late last week, underperformed Opus 4.7 on SWE-Bench Pro. They want the AI companies to let them know when they have a product that will actually and notably improve their lives, and until then, they want these companies to leave them alone and try their best not to ​crash the economy​.

As Lopatto concludes: “At some point, our Silicon Valley overlords forgot that in order for their vision of the future to be adopted, people had to want it.” They still have a lot of work to do.

AI Is Destroying the Job Market. Also, AI Is Saving the Job Market

I couldn’t help but add a quick additional note about AI to this week’s newsletter…

One of the big stories of the last year was the shrinking post-pandemic job market for recent college graduates. Many media outlets confidently offered an explanation for this shift: AI was automating the work of entry-level positions.

An ​article​ from last summer proclaimed that “AI is wrecking an already fragile job market for college graduates,” going on to note that “ChatGPT and other bots can do many of [the] chores” that used to be handled by entry-level workers. Another ​article​, published only two weeks ago, offered a stark warning: “college graduates can’t find entry-level roles in shrinking market amid rise of AI.”

But then, last week, new job numbers revealed that the entry-level job market for college graduates was rebounding, and hiring in this demographic is now projected to rise significantly. Whoops. I guess AI wasn’t actually automating those jobs. (​I told you so​.)

Does this mean the media will stop trying to force this technology into these more routine workforce narratives? If only wishing made it so. A recent Wall Street Journal ​article​ describing these positive numbers included the following line: “In some cases, artificial intelligence is spurring hires by enabling companies to expand services and product lines.”

So, let’s get this straight: AI is simultaneously contracting the job market for recent college graduates while also expanding the job market for recent college graduates.

Is there anything AI can’t do?

The post Who Asked For This? appeared first on Cal Newport.

As Sanderson explains, early in his address: “The surge of large language models and generative AI raises questions that are fascinating, and even if I dislike how the movement is going in relation to writing and art, I want to learn from the experience of what’s happening.”

Sanderson makes it clear that he disapproves of AI-generated art (“my stomach turns”), but he wants to understand better why this is the case. To do so, he begins considering and then ultimately dismissing a series of common objections:

• Does he dislike AI art because of the economic and environmental impacts? “Well, those do concern me, but if I’m answering honestly, I would still have a problem with it even if AI were not so resource hungry.”
• Does he dislike AI art because it’s trained on the work of existing artists? “ Well, I don’t like that. But even if it were trained using no copyrighted work, I’d still be concerned.”
• Does he just hate the idea of a machine replacing a person? Sanderson references the folk tale of John Henry attempting to beat a steam drill in a tunnel-digging competition that culminates in Henry’s death. “We respect him, but as a society we chose the steam drill. And I would too…The truth is, I’m more than happy to have steam engines drilling tunnels for me to drive through.”

So what is it?

Sanderson ultimately lands on a more personal reason. Talking about his struggles with his first (failed) book manuscripts, he identifies the key value of art: it changes the artist who attempts it. As he elaborates:

“Maybe someday the language models will be able to write books better than I can. But here’s the thing: Using those models in such a way absolutely misses the point, because it looks at art only as a product. Why did I write [my first manuscript]?… It was for the satisfaction of having written a novel, feeling the accomplishment, and learning how to do it. I tell you right now, if you’ve never finished a project on this level, it’s one of the most sweet, beautiful, and transcendent moments. I was holding that manuscript, thinking to myself, ‘I did it. I did it.’”

As a writer myself, I’ve also been thinking about this question recently. I like Sanderson’s take, but I’ve been developing one of my own. I understand art to be an act of deep human communication, in which the artist uses a tangible medium, such as a page of prose or a painted canvas, to transmit a complex internal cognitive state from their brain to that of their audience.

It’s telepathy. And it’s one of the most beautiful and human things we do.

This makes the idea of reading a book written by a language model, or watching a film generated by a prompt, intrinsically absurd, if not anti-human. It’s the heroin needle providing a quixotic simulation of love.

What really struck me about Sanderson’s talk, however, was his conclusion. If art is deeply human, he argues, then it’s up to us to define it. “That’s the great thing about art – we define it, and we give it meaning,” he says. “The machines can spit out manuscript after manuscript after manuscript. They can pile them to the pillars of heaven itself. But all we have to do is say ‘no.’”

I’ve noticed a trend in recent AI commentary toward a certain nihilistic passivity. You probably know what I’m talking about – the now popular style of essay in which the author, with a sort of worldly weariness, lays out some grim scenario in which AI destroys something sacred, and then sort of just leaves it there, like a cat dropping a dead bird on the doorstep.

I’m getting tired of this meekness.

Sanderson reminds us that we have agency. In the areas that matter most, it’s us, not the whims of Sam Altman or Dario Amodei, that determine how we shape our existence. All we have to do is say “no.”

Correction:

In last week’s AI Reality Check episode of my podcast, I said the following:

“If you go back and look at the release notes for Anthropic’s earlier, less powerful opus 4.6 LLM, they say the following: their researchers used Opus to find, quote, ‘over 500 exploitable zero-day vulnerabilities, some of which are decades old.’ And let’s stop for a moment because that note, which was hidden in the system card for opus 4.6, is almost word for word what anthropic said about Mythos.”

Some of this wording was sloppy, so I want to clarify it here. I was referring to this report on Opus 4.6, which Anthropic published the same day it was released. This is not technically the system card for Opus 4.6, but it is accurately described as release notes (or perhaps supplementary release notes).

This report said: “Opus 4.6 found high-severity vulnerabilities, some that had gone undetected for decades.” In another place, it said: “So far, we’ve found and validated more than 500 high-severity vulnerabilities.” Both the title of the report and the conclusion refer to these vulnerabilities as “0-day.”

The specific quote I provided, however, does not appear in the report. It’s actually a summary of the report from this tweet. In my opinion, the summary is accurate, but the way I worded the above implies that it was actually found in the report, which it was not.

Thank you to the AI researcher who pointed out these issues. I appreciate corrections! You can always send concerns or notes to podcast@calnewport.com.

The post Brandon Sanderson vs. AI Art appeared first on Cal Newport.

The “stunning advance” was the release of Anthropic’s new LLM, named Claude Mythos. In a lengthy ​press release​, Anthropic announced that the model would be made available to a consortium of business partners, but not to the general public. To justify this decision, Anthropic cited their concerns about its effectiveness at finding security vulnerabilities in source code, noting: “AI models have reached a level of coding capability where they can surpass all but the most skilled humans at finding and exploiting software vulnerabilities.”

They go on to explain that Mythos “has already found thousands of high-severity vulnerabilities, including some in every major operating system and web browser.”

This announcement clearly rattled Friedman, who called Anthropic’s decision not to release the model a “terrifying warning sign,” writing:

“Holy cow! Superintelligent A.I. is arriving faster than anticipated, at least in this area…If this A.I. tool were, indeed, to become widely available, it would mean the ability to hack any major infrastructure system — a hard and expensive effort that was once essentially the province only of private-sector experts and intelligence organizations — will be available to every criminal actor, terrorist organization and country, no matter how small.”

Friedman was far from alone in this concern. Many major news outlets expressed similar unease about this scary new development, including ​one particularly anxiety-provoking headline​ that asked if Mythos was an “AI nightmare waiting to happen?”

So, what’s really going on here?

I thought it was worth taking a moment to look closer, not just to address the specific worries about Mythos, but also to help recalibrate, more generally, how those of us seeking depth in a distracted world should consume AI news.

~~~

When I talked to people who were spooked by Friedman’s column, they tended to be under the impression that this ability to find and exploit security vulnerabilities was a new phenomenon; a skill that emerged unexpectedly in Mythos, “terrifying” those who studied it.

In reality, security researchers have been worried about using LLMs for this purpose since the beginning of consumer LLMs.

Back in 2024, for example, IBM researchers published ​a splashy study​ about using GPT-4 to attack security vulnerabilities. They found that GPT-4 successfully exploited 87% of the vulnerabilities that it was presented, as compared to close to 0% for GPT 3.5. “Our findings raise questions around the widespread deployment of highly capable LLM agents,” they concluded.

To be fair, in the case of GPT-4, researchers were assessing whether an LLM could write code to exploit a known vulnerability. Mythos, however, can also find these vulnerabilities from scratch. But this isn’t new either.

Accompanying the release notes for Anthropic’s earlier Opus 4.6 LLM was ​the observation​ that Anthropic’s security team used the model to find “over 500 exploitable 0-day [vulnerabilities], some of which are decades old.” This is almost word-for-word what Anthropic said last week about Mythos, the main difference being that they replaced 500 with “thousands.”

We are not, therefore, talking about a new capability, but rather one that has been around for multiple years.

The relevant question then becomes, how much better is Mythos at finding vulnerabilities? It’s hard to tell for sure because Anthropic has kept their new model private. They did, however, release that Mythos scored 83.1% on a well-known cybersecurity benchmark. For comparison, Opus 4.6 scored 66.6% on this same test.

In general, benchmark results should be taken with a grain of salt as they represent specific (often narrow) tests that researchers can tune their models to pass. But even if we accept that this particular measure is useful, a sixteen percentage point increase seems to represent solid incremental progress more than a nightmarish leap.

When we turn our attention to actual results, the waters become even murkier. In a recent Substack post (​which is worth reading​), Gary Marcus rounds up responses from security researchers who took a closer look at the specific exploits that Anthropic reported that Mythos discovered. They were not impressed.

• Philo Groves, for example, ​noted​ that Mythos’s attention-grabbing attack on the Firefox browser required certain common security features to be disabled, and it built on results previously discovered by Opus. (“Shocker,” he concludes sardonically.)
• The CEO of the AI company HuggingFace then ​reported​ that they took all of the specific vulnerabilities that Anthropic highlighted and “ran them through small, cheap, open-weight models.” What did they find? “Those models recovered much of the same analysis.”

Since Marcus published his essay, I’ve come across several more similar findings:

• The AI security expert Stanislav Fort ran ​an experiment​ to see if existing, cheap open-weight models could find the same vulnerability in FreeBSD (an open-source operating system) that Anthropic touted as evidence of Mythos’s scary abilities to uncover bugs that had been hiding for decades. The result: all eight existing models they tested discovered the same issue.
• Meanwhile, the renowned security researcher Bruce Schneier ​weighed in​, similarly concluding: “You don’t need Mythos to find the vulnerabilities they found.”

And of course, it doesn’t help that a week before Anthropic released this supposedly super-powered vulnerability detector, they accidentally leaked the Claude Code source, and security researchers immediately found ​serious vulnerabilities​. (I guess Anthropic forgot to use Mythos to clean up their own software…)

~~~

What’s really happening?

It’s fair to say that LLMs have created significant cybersecurity concerns that researchers have been scrambling to address in recent years. It’s also fair to say, however, that we don’t yet have evidence that Claude Mythos significantly changed this reality. If anything, some of the early independent testing by security researchers implies that Mythos might be better understood as a version of Opus 4.6 tuned to perform better on a handful of benchmarks. And yet, many still took Anthropic at their word and covered this model’s release as a catastrophic event.

In a ​recent video​, the AI commentator Mo Bitar compared Anthropic’s model rollouts to Apple iPhone launches, where every year they resell you the same product with minor improvements. “Except here,” he adds, “the product is existential dread.”

And we keep falling for it.

I think we’ve entered a stage where we need to almost entirely discount any claims made by the AI companies themselves until we can independently verify what’s actually going on.

The post Is Claude Mythos “Terrifying” or Just Hype? appeared first on Cal Newport.