Python

Welcome to the Real Python Podcast.

This is episode 311.

What are the latest trends uncovered in the 2026 Django Developers Survey? How are Django users employing LLMs in their development process? Christopher Trudeau is back on the show this week with another batch of PyCoders weekly articles and projects.

The survey draws on responses from nearly 3,500 developers across more than 40 countries.

It provides a look at how developers are adapting to AI workflows and which tools they're embracing.

We dig into a post from Brett Cannon about reproducible builds in Python.

It covers the importance of creating an independently verifiable, repeatable build of CPython and other associated tools.

The post explains why Python isn't there yet, and the direction of the work still needed.

That's the million-dollar question.

If the final product, the 10-page essay on the French Revolution, the Python code, the marketing business plan, can be instantly generated flawlessly by an AI, then the product itself is entirely decoupled from the act of learning.

Yes.

The artifact is no longer proof of cognition.

Okay, what is it?

We are rapidly approaching a moment where AI will be universally capable of producing flawless outputs instantly.

Perfect academic essays, perfect Python card, perfect corporate business plans, all generated with zero friction.

When that happens, the actual product of a higher education will no longer be those artifacts.

And at that point, what do you do? Do you just not graduate or do you figure out what is the solution? Because it's definitely not graduate or not not graduate.

So I ended up having to teach myself Python and coding and MATLAB and all of the things, which I didn't know at the time, but that was going to be essentially what would pave the way for me to be able to get a job in industry after grad school was that.

So after grad school, I was applying for lots of different jobs.

It was very throw spaghetti at the wall and see what sticks.

And so Andrew basically parachuted into the Voxel project, uh, told me my code was crap, which is awesome because it is look, I'm my only ever like true blue.

I was a developer as the job experience was doing Python.

So you can imagine what that's done to my coding practices.

I learned on Q base GW basic.

They were not what was used in the office in industry.

So I had to learn Python, Power BI, Tableau and all those other tools.

And that helped me to secure my first role in the UK as a senior finance analyst in another bank.

So for me, it was so like the fact that I didn't even go back in my career.

Now, the reason why I did those courses and those professional courses, some of them, is not because I intentionally, I want to go into tech.

I prefer to be in the social media space.

So when they're having conversations like this, now that you're saying Pi, what's it called, Python, I won't be looking like you and me.

So, yeah.

And it confidently hands you this snippet.

It's got a function from a Python library that sounds perfectly logical.

So you plug it in, you hit run and bam, fatal error.

Yep.

The model has no strong statistical reason to prefer one fake name over the other.

Wow.

So the sampler just picks the one that looks the most structurally correct for a Python library.

And this is the exact reframing Shimoda is pushing for here.

You have a table of contents for everything so that models can find things quickly.

Applying recursive language model idea there that you just run subagents to extract so the main session is not being polluted by the unnecessary context plus like I built tons of Python scripts to simplify extraction and save the context so that was easy Then I have additional sources, like, for instance, I have a scale, a rethink scale that measures your body composition and a lot of other good stuff.

Are you adding this as well? Yes.

So those people also have IPI, OWASP, EZ.

And now I was able to, you know, like, I have all the data.

I'm like, Mr.

I, let's pull in the data, analyze it, because, like, it wrote the Python scripts.

Like, you just cannot load all those files for five months into the context window.

But in an agentic framework, the model has access to tools.

It can run Python scripts.

access a command line interface, or execute API calls in the background.

So it's multi-layered.

Yeah.

Conduct a live test and then feed the telemetry data back to Claude for analysis and refinement.

And Anthropic also had to disclose that one of their models eponymously uploaded live malware to the Python Package Index, or PyPI.

OK, I want to pause here, just, you know, challenge this framing a bit, because when we use words like hijacked, forged and malicious, we are applying very human intent to lines of code.

Yeah.

They are aggressively adding Python, data science, and AI modules to their core curriculum.

Wow.

So the test is changing.