Did the FinTech company Block just lay off 40% of its workforce due to AI automation?
Can the best AI models pass a freshman computer science class?
Programmers love agentic AI, but how exactly are they using these tools?
For those of you who followed the tech news this past week, these are all pressing questions.
We're going to try to find some answers.
I'm Cal Newport, and this is the AI Reality Check.
Now I wanna do a quick aside before we get into this week's stories, because this is a new format for my podcast feed.
I wanna give you a quick explanation.
Thank you so much for having me. to its own mini episodes on Thursdays.
This is an experiment.
Maybe I'll move it back.
Maybe I'll move it to its own feed.
Maybe I won't do it every week.
So just bear with me.
But keep in mind, if you want to share any of these episodes, we're also putting it up on YouTube so you can send the video link to someone who might need to hear some of this reality checking.
All right, that's enough logistics.
Let's get into our first story of the week.
All right.
Late last week, Jack Dorsey, the CEO of the fintech company Block they're responsible for Stripe and Cash App, among some other products posted a note on X announcing massive layoffs at his company.
Let me read you from this note.
Dorsey said, today we're making one of the hardest decisions in the history of our company.
We're reducing our organization by nearly half. from over 10,000 people to just under 6,000.
That means over 4,000 of you are being asked to leave.
All right, later on, he says the following, we're not making this decision because we're in trouble.
Our business is strong, dot, dot, dot, but something has changed.
We're already seeing that the intelligence tools we're creating and using, paired with smaller and flatter teams are, are enabling a new way of working which fundamentally changes what it means to build and run a company.
And it's accelerating rapidly.
Can I make a quick aside?
This is like a hint to CEOs.
If you are announcing the layoff of 40 of your staff, can you use capital letters at the beginning of your sentences?
It really caught my attention in this tweet that he doesn't capitalize any of his words.
I don't know.
It feels a little disrespectful, but let's get back to the actual story here.
The traditional media was quick to embrace and amplify Dorsey's claim that these layoffs were because AI made these positions redundant or unnecessary.
Here is the headline, for example, from a New York Times article about the layoffs.
The headline read, Here's the subhead from that article.
About 4000 workers will lose their jobs as the payment company does more work with new artificial intelligence tools.
Its top executive said
Another quick aside, because this is a journalistic thing I begin to notice more and more, I think really starting around the COVID coverage era, where you have a claim that feels right, that you want to put in your subhead because there's a point you're trying to make but either it's hard to fact-check or or you don't want to fact-check it because you're not quite sure what you're going to find.
It'll be complicated.
So you just make the claim, then you put a comma and attribute it to someone else.
We didn't used to see attributed claims in sub-headlines or headlines, but we began to see it more.
It's a good way of I'm trying to make a point here and I don't actually want to go and directly verify Did they lay off all these people because of AI tools?
I'll just say, they lay off people because of AI tools, said someone.
So you add as a comma.
So just keep in mind that sort of reporting trick.
If we read the article itself, the framing makes it super clear what they're implying here.
Here's from the article.
The cuts made, as Block reported strong financial results for its most recent quarter, are perhaps the most striking example so far of a technology company's making plans to eliminate employees because of AI.
I don't mean to pick on the times.
A lot of publications had similar coverage, and the stock price went up 20% for Block.
This is an important article to look at, in part because I got sent it a lot of times.
When I get sent an article a lot of times that means it is catching people's attention and is either exciting or upsetting them.
So it's worth some closer scrutiny.
I think there's a general vibe that this article is trying to verify or validate, which is the vibe of something big is happening.
Yeah, we've been talking about AI could get rid of jobs or whatever, but now it's happening.
See, look.
This is the first shoe to drop of a major crisis.
It's the first company that laid off almost half of its workforce.
This is the thing we've been warning you about.
Major economic disruption.
It has begun.
That is a story that is very sticky and very attention-catching.
But is it true?
Well, if you dig a little deeper, there's a lot of commentators online who know this industry sector a little bit better than who are not at all convinced.
Let me give you a few bits of contextual information about Block and its layoffs.
Between 2019 and 2025, Block's employee count grew from around 4,000 employees to over 10,000.
So they had massive growth during the pandemic.
A lot of this growth actually came from acquisitions in the crypto and blockchain space earlier in the pandemic, when those things were still hot.
Those acquisitions are now, of course floundering, as those technologies, especially the blockchain-based software technologies, are having a hard time.
A lot of the startups are really struggling.
Despite the fact that The Times said that they had, quote, strong financial results, end quote.
If you actually read the industry analysts who study the quarterly reports from Block, they're not impressed because the last two quarters they actually fell short of their earnings targets.
So here's an alternative explanation for what might be going on here.
Like just about every major tech company in America, block overhired during the pandemic, when that industry was booming.
Also, like just about every major tech company right now, in the last two years they're shedding jobs to try to right size back because they had overhired during the pandemic.
We've talked about on this show before, Amazon doing this, Microsoft is doing this.
This is a common trend in recent years.
But how do we know it really wasn't AI?
AI is the reason why they laid off these 4,000 people.
Well, there's a couple things going on.
One, a lack of specificity in Dorsey's statement.
He just says, well, we have these intelligence tools, and then he talks about non-AI things.
Again, we have different types of teams, and we don't need as many people anymore.
No specific reference of this particular tool has taken on this role, so we shut down this division because we don't need employees there.
Or in this division.
What we did is we laid off the entire entry-level class because the managers can now get by with less.
It's very vague what he said.
Two, as we'll hear later in today's episode.
Though there is major changes happening in computer programming because of new agentic AI tools.
Basically, every serious commentator who is studying this industry says yeah, we're not yet.
We haven't figured out.
The companies haven't figured out exactly what this means.
We're certainly not laying off, ready to lay off half of our workforce yet.
These tools are very new, the versions that people are getting excited about.
But maybe the most telling reason why we know this is not AI is that Ethan Mollick didn't buy this claim.
Ethan Mollick from PIN is a respected AI commentator who is very much on the booster side.
He's very, AI is going to change everything.
And even he didn't buy this idea that AI was responsible for the layoffs at Block.
On a LinkedIn post, Ethan Mollick said the following, referring to the layoffs this isn't about AI, but that is a smart way to sell it if you want to see your stock jump 20.
Then on X, Ethan Mollick said the following in response to Dorsey's tweet.
Two things.
One.
Given that effective AI tools are very new and we have little sense of how to organize work around them, it is hard to imagine a firm wide, sudden 50 efficiency gain.
Two CEOs with vision who hired well should also use AI for expansion and augmentation, not decimation.
I'll just say as an aside I've been hearing this from the managers and programmers I've been talking to in the last couple of weeks about how they're using agentic programming.
I am much more likely to see the effect to be.
I mean, I haven't had any of them say we're laying people off, but I have heard a lot of people say, like Malik implies here, the reaction to these tools at a lot of these startups has been do more work.
Great, now we can do more work with the same people.
Let's make more money out of the same people, not let's lay people off.
All right, we have another voice of skepticism here.
This one comes from Ron Shevlin, who is an industry analyst who specializes in the fintech sector.
So he specializes in the sector where Block is and he writes and covers Block professionally as a financial journalist.
He wrote a column right after this that was titled the following Locke lays off 40 of staff and blames it on AI.
Don't buy the excuse.
And he goes on to say, yeah, they overacquired.
They made some bad acquisitions.
They need the right size.
And they're blaming AI because it sounds better than saying yeah, we made some bad calls during the pandemic and now we have to adjust to it.
All right, so what's the bottom line here in terms of reality checking this story?
AI will have an impact on jobs.
I'm not one of these skeptics that says this is a fad that's going to go away, that this is going to be like blockchain-based software.
That really just failed to catch on.
But we're not really there yet outside of some narrow instances.
The tools have not matured to the phase where we really understand what's going on, where we're really seeing major changes to the way companies are structuring themselves.
Most of the commentators I can find who follow this closely say yeah, sure there is going to be things happening with jobs.
We don't know if it's going to lead to expansions or contractions or what sectors get hit more than yet, but we're not there yet.
There is a tendency, I think among coverage right now, to lean into that vibe that AI is going to affect jobs and try to keep making the claim that's happening right now.
And what's happening is the CEOs of these companies, especially tech companies.
So CEOs like Jack Dorsey are seeing the tendency towards that vibe reporting.
This is very tempting for journalists.
And so they're trying to – there's a term Annie Lowry introduced.
I think it was something like AI washing.
They're trying to justify layoffs that are due to things like pandemic overhiring by saying well AI, we're being smart.
So they look better, like better decision makers and like they're more forward thinking.
It's important that we cover AI's impact on jobs accurately so that when real impacts come, we can see them with clear eyes.
And react to them honestly and hold to account the actual, why are you firing these people?
What's happening here?
What leader's doing this?
We really do need to cover that accurately.
So we have to stop the vibe reporting on the AI job apocalypse.
It's not here yet.
And we don't know if it's going to come at all.
But the best we can do is try to be accurate about what we're saying.
All right, second story.
This one's kind of a fun one.
All right.
So Anthropic CEO Dario Amadei famously said in recent – I guess this is all this last year – famously said that their LLM products have the intelligence of someone with a doctorate.
Before, like, well, it was as smart as a high school student, then as smart as a college student.
Now it's as smart as someone with a doctorate.
He described this product –
Deploying his product like having an, quote, army of PhDs, end quote, in your data center.
Last month, he used a related terminology.
He said, we can offer you a country of geniuses in a data center.
Well, I was thinking about this approach of sort of describing AI with human education levels, when I came across an interesting video that was posted in January, which did a really cool experiment.
A TA... for Cornell University's freshman computer science course, CS2112.
They probably call it 2112.
This is their sort of advanced freshman fall CS course.
So if you come into the CS program there as a pretty advanced student, this would be the course you would take, but it's for freshmen in their first semester.
He was TAing it.
So he said, here's what I'm going to do.
I'm going to take the three leading AI models and I'm going to give them every graded thing we do in this class.
I will give to the models.
And then I will grade their results.
At the same time, I'm grading the real students in the class using the exact same rubrics.
And then at the end I will you know wait the grades.
Just treat them like a student in this class and see how they do.
Let me play you a quick clip here.
This is the intro, the intro to that video.
Can AI pass a first semester freshman CS class?
To answer this question.
I ran every single assignment, every exam, every quiz, every graded interaction.
The students got this semester through the three best models I could get my hands on from ChatGPT Claude and Gemini.
Then I graded each result with the exact same rubric we use on students, so that I could give each AI the most accurate possible grade in the class.
All right, so this was a very entertaining video if you watched the whole thing, because he goes through specific assignments.
He's like, whoa, look, this is really cool.
Oh my God, look at this crazy thing it did.
It's well edited.
I thought it was really cool.
In the end, they have a competition in the class where you create these like critters that evolve and they had the AI models.
Critters compete with the critters from the class.
A couple of things I noticed from the videos, sometimes these models did very well on assignments.
Sometimes they really struggled.
Sometimes they made very revealing, baffling mistakes.
Like in an early assignment, where they were doing some simple string concatenation.
And the assignment had you write a program that was going to output the word.
You're going to create a string concatenation, but basically you're going to output the word hello is what it asked you to do on the screen.
And Claude's submission outputted hello world world.
Because what's going on here is there's a lot of AI assignments out there, I mean CS assignments out there that famously say hey, write hello world as the first thing you do when you're using a new programming environment.
And clearly it was just trying to statistically grow out its answer.
It's like well, if I'm printing hello in an assignment, I got to print hello world and then add another world, just to be safe.
But how did they end up grade-wise?
Okay, so I have the grades in front of me here.
They used the latest, greatest models from ChatGPT, Claude, and Gemini.
They actually upgraded during the fall.
They did this last fall.
They were using the very most expensive version of the Claude LLM available.
I forgot which one.
And then when a new one came out, they upgraded to that new one.
On some assignments, these things did pretty well, especially the early assignments.
We got like on the first assignment, ChatGPT got a 102 out of 104.
Claude got a 99 out of 104.
Jim and I got a 101 out of 104.
They also did well on the final exam, because this was an in-class final exam where you're just writing answers, right.
So like you just have to use the knowledge in your head.
So that's a good setup, again, for LLMs.
And so like ChatGPT got a 93 out of 100, Gemini got an 84.
There's other assignments where they really struggled.
Assignment six, ChatGPT got 32 out of 100, Claude got 20 out of 100, Gemini got 13 out of 100.
On assignment five, ChatGPD got 60 out of 100.
Claude got six out of 100.
Jim and I got 67 out of 100.
There's a lot of issues it had with hallucinating.
It had a hard time.
If you watch this video where the assignment would give you multiple, you know some rules for what to do in the assignment and it would just sort of skip some of the rules.
Sometimes, I think in the example where Claude got six out of 100, it just kind of made up its own assignment and solved that one instead.
So it's sort of a mixed bag.
In terms of its final grades.
Two of the models, Claude and Gemini, ended up getting a C-plus in the class.
This is a freshman computer science.
You need a 2.5 to declare in the initial classes.
You need a 2.5 GPA at Cornell to declare yourself as a computer science major.
A C-plus is like a 2.3 something.
So they weren't doing well enough to actually even major in computer science.
ChatGPT did better with the B+.
It was below the median for the class, but it did somewhat better.
Anyways, here's what's interesting about this.
I mean, there's kind of the catchy thing.
It's like this is an army of geniuses.
This is a PhD level whatever.
They're struggling with the first class you take as a freshman in computer science, which is the topic that these models are best suited for.
So there's that sort of like gotcha moment.
But that's not really what this is about, right?
Because I'm sure you could get these chatbots to get you the right answer to these assignments if you're willing to be sufficiently interactive and hold their hands and get the prompts in just the right way and correct them.
That's not really the right way, the right takeaway here.
I think the right takeaway here was that it was stupid all along for Dario Amadei to try to use human education levels as a way to describe a large language model.
This is just different.
The human brain, we have a general purpose integrated brain that does lots of things.
The whole person is educated.
It makes sense to talk about the education level of a person, but not really a language model.
It turns out a lot of these claims, like when Dario Amadei, I went back and checked this out.
Excuse me.
Why did he originally say that their language models were now PhD level?
It's because they had the original time he started saying, that is, that they had given it math problems like a problem set.
And it was doing well on the math problems from this problem set.
And one of the professors who worked on creating those problem sets said, those are hard problems.
Those are the type of problems I would assign to my graduate students.
That's where they originally got the claim that this is a PhD level, right?
So this idea of just generally talking about the intelligence level of language models, I think is anthropomorphizing and is not useful.
The reality is these are very specialized tools.
They tend to get tuned for specialized purposes.
And to get their real value.
It's a combination of the tool and learning as the human, how best to use and deploy the tool and check its work and redeploy it towards that particular goal.
That is a very different tool use scenario.
It's a tool use scenario.
It's very different than imagining just an anthropomorphized brain that has a general education level.
So hopefully we can stop using terms like having a data center full of PhDs.
Also, that was a clever video.
So kudos to that TA for putting that together.
It was a hard CS class.
It was definitely harder than the intro CS classes I took at Dartmouth, but it reminds me of the type of classes we had at MIT.
So it was a hard class.
All right.
One final story here.
The story actually comes from me.
Obviously, there's a lot going on in the last four or five months, with new agentic coding tools being enthusiastically embraced by computer programmers.
A lot of these viral essays are going around that just keep – and articles that are influenced by those essays and podcasts where people are talking about.
Oh my god, huge changes are happening in the world of computer programming.
This is really going to be – this is like ground zero for the long promise.
We're about three years in now.
The long promise claim that the language model-based tools are going to have massive disruptions.
But what actually is going on?
I've been trying to find out.
As people who subscribe to my newsletter at calnewportcom know, a week or two ago I put out a call for professional computer programmers to send me detailed reports about exactly how they and them teams use language model-based AI tools and how this has changed in the recent past.
I have over 350 such reports in so far.
I've carefully made my way through 100.
I'm really trying to get my brain around what's really happening with professional programmers and these tools.
I thought it would be useful today...
They read you excerpts from two responses that I think are very typical of the type of responses I'm reading.
That try to give you a better picture of what exactly does it mean for these programmers to be using these new tools.
I cut out details in these and have some elision to get rid of identifying details.
All right, so here's my first excerpt.
I'm a software developer working at a tech startup.
Our use of AI varies by person at the company, but my use has skyrocketed, starting in the fall of 2025.
So much so that I don't write any code anymore.
But I'm still heavily involved in oversight in architecture.
I used Cursor quite a bit last year but have moved on to working directly in the terminal with Codex at work.
The workflow goes something like this plan a feature or start a discussion about a bug fix with AI, discuss until I'm satisfied, have it output a plan, iterate on the plan, then execute the plan.
After execution, I verify the outcome.
I use Git extensively throughout this process.
Git is a repository software for managing code that multiple people are working on.
I've tried the multi-agent approach, where multiple agents are working on different Git work trees at the same time.
I can't do it.
It's too much context switching and I end up just accepting things I wouldn't normally accept, because it's an exhausting process.
The quality dips dramatically.
I love my current workflow.
I've developed things in the past week that would have taken me months before.
All right, let's pause there before I do the second excerpt.
This, I would say, is very typical of what I would call the enthusiastic all-in user from among the subset of professional programmers.
Most of the code they're producing is now actually being generated by an AI agentic tool.
Typically, it is cloud code. where they switched the model behind it.
I don't know if it was Opus to Sonnet or Sonnet to Opus in the fall, and that really seemed to make it good enough now that a lot of people wanted to use it.
Though I would say I also see ChatGPT Codex is also commonly used.
But an interesting thing about this, or I want to point out two things.
One, there's a lot of just chatbot discussion happening in these workflows.
Remember we talked about making a plan, iterating on the plan.
That's all actually like chatbot interaction, so so sort of or related to using these tools to produce more code.
These programmers are have entered a more interactive way.
They, you know, they want to talk back and forth, and It reminds me a lot of the research I did for The New Yorker about how students are using chatbots to write paper.
They find talking back and forth with a chatbot as they write is less straining.
So that's picking up here.
But also notice this programmer is not really big on the multi-agentic approach.
Which is what you see most often told in the sort of breathless online articles and YouTube videos, is this idea of I have 20 agents working at the same time and this agent checks this agent and there's a supervising agent that looks at those agents and then it reports over here to the hierarchy agent and then that agent is on open claw so that it can uh, it can send recommendations to my youtube channel and to make sure that it pays that.
You know these super complicated trees of different agents supervising other agents.
You really aren't seeing that, at least in my study here it's.
You're not seeing a ton of that in professional programmers.
You tend to see it more in people who are like working on their own personal bespoke projects and find it really fun, but i don't see as much, and that's what we saw reflected here.
All right, let me read you one other typical uh excerpt here from a real professional programmer.
I think this captures well another very common type of response, which is a little bit more reticent, but still appreciating the power of these new tools.
Let me read this.
I'm a software developer working at a tech startup.
Our use of AI varies by person at the company, but my use has skyrocketed, starting in the fall of 2025.
Oh, wait, that was the last one.
I'm sorry.
This is the new one.
I don't want to just reread the last one.
All right.
I'm like a language model here, just sort of randomly hallucinating the same answer twice.
No, no.
Here's the real second excerpt.
I'm a staff software engineer at a tech startup.
The AI models have made the easiest tasks even easier.
Scaffolding a solution, boilerplate code, replacing variables or moving an import.
Repetitive tasks are good candidates.
LLMs are also useful as a way to quickly investigate the documentation of a tool or get a reminder on syntax for something I'm trying to do.
But the easy stuff, the tasks that AI can do well, was never the hardest nor most time-consuming part of my job.
When actively using these coding agents, I found that it generally slows me down.
Using them, introduced tasks I didn't have before composing a prompt, checking the output, reprompt manually, refactor when it isn't quite right.
It also slows down the code review process.
I'm much more detailed in my reviews.
When I know a coworker used an LLM to generate some or all of the code, that's also a very common response as well.
That's pointing out this idea, which I think is a fair criticism, that the people like our first excerpt, which is doing most of their code generation with agentic AI like this is saving so much time.
They're noting that, The more reticent users are noticing you're downplaying the huge amount of time that now surrounds.
Yeah, you don't write the code yourself.
That's faster.
But now you have to do so much other work.
All of this iteration with the model and the prompts and try the prompt again and work on your agent markdown file and your skills harness.
And then all of the review on the other side.
And if it was produced with AI.
You really have to review it.
And he's like there's all of this other work that's surrounding this workflow, which is none of it's very fun.
I mean, and this is taking a lot of time.
Are we sure?
Are we sure that this is actually producing the best code?
So there's sort of this tension going on in the computer programming world.
Here's the takeaway from this.
One agentic coding tools past the threshold of usefulness.
With the Cloud Codecs update in the fall, that has made them much more heavily used.
In my survey, something like 45 of the people I talk to are now producing the majority of their code with an agentic tool such as Cloud Code.
All right.
Two, it's really unclear exactly what the best practices are for this are.
There seems to be a spectrum of enthusiasm of the users of it in the space.
For sure, on one end, there's way too much AI interaction going on.
This can't be the most efficient way to do it.
On the other end, there's a lot of reticence.
The reality is going to fall somewhere in the middle.
We don't yet know. what the future computer programming looks like.
I think by the summer, there's going to be some best practices.
They'll have some clever acronyms to go with them.
There'll be some best practices about how best to use these.
There will be automatic code production.
I think we're going to pull back a little bit on how much AI chatbot should be involved in review as well as planning.
I think that's a little bit of just enthusiasm there.
I do think a lot of code will still be generated, but we'll be better at where we deploy the code.
I think there'll be more standardization about planning and architecture documents etc.
Which will have a high overhead at first, but it'll allow us to deploy these tools better.
I do not think, based on these interviews, that the hyper multi-agent approach that we see most talked on the internet is going to become some sort of standard for serious programmers in most places.
And the vibe coding, like you see, you know, talked about a lot, give me this app.
And I come back a week later and it's done.
That really is in the realm of hobbyists and personal apps for yourself or people who are doing experiments.
None of the serious programmers I heard of so far are doing anything like that for the most part.
All right?
So there's a lot to be done here.
But what I'm trying to do – this is why it's a reality check.
I am not interested in breathless accounts of what's happening online, because that's engagement hunting.
I'm not interested – in hearing sort of like non-technical reporters who have just heard a lot of those accounts and then are like look, I don't know the details, but I think we can all agree that like there's not going to be programmers in the future.
I think we got to talk to real programmers.
What is really going on?
Something is happening.
It's more complicated than other people make it seem.
Let's keep listening.
I'll read you some more of these reports in weeks ahead.
Let's figure out the old fashioned way.
Turn every page, learn what's going on, what's working, what's not, what's hype, what's not.
And let's try to figure out what's actually happening.
I think we will.
And we'll get on it, especially if you follow me here.
All right.
That's all the time I have for today.
Remember, take AI seriously, but not necessarily everything you hear about it.
I'll be back on Monday with the main episode and hope they'll do another one of these next Thursday.
See you then.
