Month: January 2025

You Still Need to Learn to Write Code in the Age of LLMs

Can we really delegate most or all of our coding tasks to LLMs? Should we tell our kids not to study computer science? What are the reasons we should still learn to write code in the age of LLMs?

There is a chorus of voices telling us that soon we will be able to hand all of our coding tasks over to an AI agent powered by an LLM. It will do all the tedious boring things for us, and we can focus on the important stuff. While the new generative AIs are amazing in their capabilities, I for one think we shouldn’t be so quick to dismiss the value of learning to code. Granted, I make my living teaching people to write software, so maybe I should call this “Why I don’t quit my job and tell everyone to use ChatGPT to write their code”, because I believe that learning to code is still necessary and good.

In early 2024 the founder and CEO of NVIDIA Jensen Huang participated in a discussion at the World Governments Forum that inspired countless blog posts and videos with titles like “Jensen Huang says kids shouldn’t learn to code!”. What he actually said is a bit different but the message is essentially the same [click here to watch for yourself, it’s the last 4 minutes or so of the interview]: “It’s our job to make computing technology such that nobody has to program … and that the programming language is human. Everybody in the world is now a programmer…“

Photo of Jensen Huang speaking at the 2024 World Governments Forum.

He suggests that instead of learning to code, he says we should focus on the Life Sciences because that’s the richest domain for discovery, and he thinks that developing nations that want to rise on the world stage should encourage their children to do the same. Oh, and he says we should build lots of infrastructure (with lots of NVIDIA chips of course).

There is a core part of his message I actually agree with. At Diller Digital, and at Enthought where our roots are, we have always believed it’s easier to add programming skills to a scientist, engineer, or other domain expert than it is to train a computer scientist in one of the hard sciences. That’s why if you’ve taken one of our courses, you’ve no doubt seen the graphic below, touting the scientific credentials of the development staff. And for that reason, I agree that becoming an expert in one of the natural sciences or engineering discipline is personally and socially valuable.

Image from the About Enthought slide in Enthought Academy course material showing that 85% of the developers have advanced degrees, and 70% hold a PhD.

At Enthought, almost no one was formally trained as a developer. Most of us (including me) studied some other domain (in my case it was Mechanical Engineering, the Thermal and Fluid Sciences, and Combustion in particular) but fell in love with writing software. And although there is a special place in my heart for the BASIC I learned on my brother’s TI/99 4A, or Pascal for the Macintosh 512k my Dad brought home to use for work, or C, which I self taught in High School and college, it was really Python that let me do useful stuff in engineering. Python has become a leading language for scientific computing, and a rich ecosystem has developed around the SciPy and NumPy packages and the SciPy conference. One of the main reasons is that it is pragmatic and easy to learn, and it is expressive for scientific computing.

And that brings me to my first beef with Huang’s message. While the idea of using “human language”, by which I believe he means “human language that we use to communicate with other humans” otherwise known as natural language, to write software has some appeal, it ignores the fact that we already use human language to program computers. If we were writing software using computer language, we’d be writing with 1s and 0s or hexadecimal codes. Although there are still corners of the world where specialists do that, it hasn’t been mainstream practice since the days of punch cards.

Image of human hands holding a stack of punch cards. Image originally appears on IBM's web page describing the history of the punch card.

Modern computer languages like Python are designed to be expressive in the domain space, and they allow you to write code that is clear and unambiguous. For example, do you have any doubts about what is happening in this code snippet borrowed from the section on Naming Variables in our Software Engineering for Scientists & Engineers?

gold_watch_orders = 0
for employee in employee_list:
    gold_watch_orders += will_retire(employee.name)

Even a complete newcomer could see that we’re checking to see who’s about to retire, and we’re preparing to order gold watches. It is clearly for human consumption, but there are also decisions about data types and data structures that had to be made. The act of writing the code causes you to think about your problem more clearly. The language supports a way of thinking about the problem. When you give up learning the language, you inevitably give up learning a particular way of thinking.

This brings me to my second beef with the idea that we don’t need to learn programming. Using what Huang calls a “human language” in fact devolves pretty quickly into an exercise called “prompt engineering”, where the new skill is knowing how to precisely specify your goal to a generative model using a language that is not really designed for that. You end up needing to work through another layer of abstraction that doesn’t necessarily help. Or that is useful right up to the point where it isn’t, and then you’re stuck.

I often point my students to an article by Joel Spolsky called “The Law of Leaky Abstractions“, in which the author talks about “what computer scientists like to call an abstraction: a simplification of something much more complicated that is going on under the covers.” His point is that abstractions are useful and allow us to all sorts of amazing things, like send messages across the internet, or to our point, use a chat agent to write code. His central premise is there is no perfect abstraction.

All non-trivial abstractions, to some degree, are leaky.

Joel Spolsky

By that, he means that eventually the abstraction fails, and you are required to understand what’s going on beneath the abstraction to solve some tricky problem that eventually emerges. By the time he wrote the article in 2002, there was already a long history of code generation tools attempting to abstract away the complexity of getting a computer to do the thing you want to do. But inevitably, the abstraction fails, and to move forward you have to understand what’s going on behind the abstraction.

… the abstractions save us time working, but they don’t save us time learning.
And all this means that paradoxically, even as we have higher and higher level programming tools with better and better abstractions, becoming a proficient programmer is getting harder and harder.

Joel Spolsky

For example, I’m grateful for the WYSIWYG editor WordPress provides for producing this blog post, but without understanding the underlying HTML tags it produces and the CSS it relies on, I’d be frustrated by some of the formatting problems I’ve had to solve. The WYSIWYG abstraction leaks, so I learn how HTML works and how to find the offending CSS class, and it makes solving the image alignment problem much much easier.

But it’s not only the utility of the tool. There’s a cognitive benefit to learning to code. In my life as a consultant for Enthought, and especially during my tenure as a Director of Digital Transformation Services, I would frequently recommend that Managers, and even sometimes Directors, take our Python Foundations for Scientists & Engineers, not because they needed to learn to code, but because they needed to learn how to think about what software can and can’t do. And with Diller Digital, the story is the same. Especially in the Machine Learning and Deep Learning classes, managers join because they want to know how it works, what’s hype and what’s real, and they want to know how to think about the class of problems those technologies address. People are learning to code as a way of learning how to think about problems.

The best advice I’ve heard says this:

Learn to code manually first, then use a tool to save time.

I’ll say in summary, the best reason to learn to code, especially for the scientist, engineers, and analysts who take our classes, is that you are learning how to solve problems in a clear, unambiguous way. And even more so, you learn how to think about a problem, what’s possible, and what’s realistic. Don’t give that up. See also this article by Nathan Anacone.

What do you think? Let me know in the comments.

Managing Pandas’ deprecation of the Series first() and last() methods.

Have you stumbled across this warning in your code after updating Pandas: “FutureWarning: last is deprecated and will be removed in a future version. Please create a mask and filter using `.loc` instead“? In this post, we’ll explore how that method works and how to replace it.

I’ve always loved the use-case driven nature of methods and functions in the Pandas library. Pandas is such a workhorse in scientific computing in Python, particularly when it comes to things like timeseries data and dealing with calendar-labeled data in particular. So it was with a touch of frustration and puzzlement that I discovered that the last() method had been deprecated, and its removal from Pandas’ Series and DataFrame types is planned. In the Data Analysis with Pandas` course, we used have an in-class exercise where we recommended getting the last 4 weeks’ data using something like this:

In [1]: import numpy as np
   ...: import pandas as pd
   ...: rng = np.random.default_rng(42)
   ...: measurements = pd.Series(
   ...:    data=np.cumsum(rng.choice([-1, 1], size=350)),
   ...:    index=pd.date_range(
   ...:        start="01/01/2025",
   ...:        freq="D",
   ...:        periods=350,
   ...:   ),
   ...:)
In [2]: measurements.last('1W')
<ipython-input-5-ec16e51fe7ce>:1: :1: FutureWarning: last is deprecated and will be removed in a future version. Please create a mask and filter using `.loc` instead
  measurements.last('1W')
Out[2]:
2025-12-15   -7
2025-12-16   -8
Freq: D, dtype: int64

This has the really useful behavior of selecting data based on where it falls in a calendar period. Thus the command above usefully returns the two elements from our Series that occur in the last calendar week, which begins (in ISO format) on Monday, Dec 15.

The deprecation warning says “FutureWarning: last is deprecated and will be removed in a future version. Please create a mask and filter using .loc instead.” Because .last() is a useful feature, I wanted to take a closer look to see if I could understand what’s going on and what the best way to replace it would be.

Poking into the code a bit, we can see that the .last() method is a convenience function that uses pd.tseries.frequencies.to_offset() to turn '1W', technically a designation of period, into an offset, which is subtracted from the last element of the DatetimeIndex, yielding the starting point for a slice on the index. From the definition of last:

 ...
    offset = to_offset(offset)

    start_date = self.index[-1] - offset
    start = self.index.searchsorted(start_date, side="right")
    return self.iloc[start:]

Note that side='right' in searchsorted() finds the first index greater than start_date. We could wrap all of this into an equivalent statement that yields no FutureWarning thus:

In [3]: start = measurement.index[-1] - to_offset('1W')
In [4]: measurement.loc[measurement.index > start]
Out[4]:
2025-12-15   -7
2025-12-16   -8
Freq: D, dtype: int64

There’s a better option, though, which is to use pd.DateOffset. It’s a top-level import, and it gives you control over when the week starts, which to_offset does not. Remember we are using ISO standards, so Monday is day 0:

In [5]: start = measurements.index[-1] - pd.DateOffset(weeks=1, weekday=0)
In [6]: measurements.loc[measurements.index > start]
Out[6]:
2025-12-15   -7
2025-12-16   -8
Freq: D, dtype: int64

Slicing also works, even if the start point doesn’t coincide with a location in the index. Mixed offset specifications are possible, too:

In [7]: measurements.loc[measurements.index[-1] - pd.DateOffset(days=1, hours=12):]
Out[7]:
2025-12-15   -7
2025-12-16   -8
Freq: D, dtype: int64

The strength of pd.DateOffset is that it is calendar aware, so you can specify the day of the month, for example:

In [8]: measurements.loc[measurements.index[-1] - pd.DateOffset(day=13):]
Out[8]:
2025-12-13   -7
2025-12-14   -6
2025-12-15   -7
2025-12-16   -8
Freq: D, dtype: int64

There’s also the non-calendar-aware pd.Timedelta you can use to count back a set time period without taking day-of-week or day-of-month into account. Note: as with all Pandas location-based slicing, it is endpoint inclusive, so 1 week yields 8 days’ measurements:

In [9]: measurements.loc[measurements.index[-1] - pd.Timedelta(weeks=1):]
Out[9]:
2025-12-09   -9
2025-12-10   -8
2025-12-11   -7
2025-12-12   -8
2025-12-13   -7
2025-12-14   -6
2025-12-15   -7
2025-12-16   -8
Freq: D, dtype: int64

You may have noticed I prefer slicing notation, whereas the deprecation message suggests using a mask array. There’s a performance advantage to using slicing, and the notation is more compact than the mask array but less so than the .last() method. In IPython or Jupyter, we can use %timeit to quantify the difference:

In [10]: %timeit measurements.loc[measurements.index[-1] - pd.DateOffset(day=13):]
45.7 μs ± 2.36 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

In [11]: %timeit measurements.last('4D')
56.3 μs ± 14.9 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

In [12]: %timeit measurements.loc[measurements.index >= measurements.index[-1] - pd.DateOffset(day=13)]
89.2 μs ± 6.31 μs per loop (mean ± std. dev. of 7 runs, 1,000 loops each)

After spending some time with git blame and the Pandas-dev source code repository, the reasons for the deprecation of the first and last methods make sense:

  • there is unexpected behavior when passing certain kinds of offsets
  • they don’t behave analogously to SeriesGroupBy.first() and SeriesGroupBy.last()
  • they don’t respect time zones properly

Hopefully this has been a useful exploration of pd.Series.last (and .first), their deprecation, and how to replace them in your code with the more-explicit and better-defined masks and slices. Happy Coding!

Arrays and Lists

I often get questions about the difference between the Python list and the NumPy ndarray, and we talk about this a lot in Python Foundations for Scientists & Engineers. But recently I had a question about the array object that is part of the Python language, and why don’t we use it for computations? Why is NumPy necessary if Python has an array data type? Well, this is a great question!

First of all, let’s take a quick look at the Python array. It’s found in the array standard library and usually imported as arr:

>>> import array as arr

Like a NumPy ndarray, it has a data type, and every element of the array has to conform to that type. Set the data type in in the first argument when creating the array. 'l' in this case specifies a 4-byte signed integer:

>>> a = arr.array('l', range(5))
>>> a
array('l', [0, 1, 2, 3, 4])

It has the same indexing and slicing behavior as a Python list, and we even see that it has the same “multiplication” behavior as lists: i.e. “multiplying” means repetition, not element-wise, arithmetic multiplication as with the NumPy ndarray.

With that in mind, computations look the same for arrays as for lists:

>>> b = arr.array(a.typecode, (v + 1 for v in a))
>>> b
array('l', [1, 2, 3, 4, 5])

Let’s use IPython’s %timeit magic to see how array computations fare in comparison with Python lists and NumPy ndarrays. For the list and array, we’ve specified a list comprehensions as the most efficient way to do the computations and a simple vector computation for the NumPy ndarray, and our test is to increment sequence of 10 integers:

In [1]: import array as arr
   ...: import numpy as np
   ...:
   ...: %timeit [val + 1 for val in list(range(10))]
   ...: %timeit arr.array('q', [val + 1 for val in arr.array('q', range(10))])
   ...: %timeit np.arange(10) + 1
279 ns ± 45.8 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
825 ns ± 5.22 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)
1.03 μs ± 98.7 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

It’s surprising to note that list comprehension is fastest, beating array computation by a factor of 3 or so, and beating ndarray computation by a factor of 3.7! Suspecting that a 10-integer sequence may not be representative, let’s bump it up to 1000:

In [2]: %timeit [val + 1 for val in list(range(1000))]
   ...: %timeit arr.array('q', [val + 1 for val in arr.array('q', range(1000))])
   ...: %timeit np.arange(1000) + 1
28.4 μs ± 3.47 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)
82.4 μs ± 9.05 μs per loop (mean ± std. dev. of 7 runs, 10,000 loops each)
1.82 μs ± 175 ns per loop (mean ± std. dev. of 7 runs, 1,000,000 loops each)

With more numbers to chew on, NumPy clearly pulls ahead, suggesting the overhead of creating an ndarray may dominate computation time for small arrays, but the actual computation time is relatively small. The Python array is still relatively inefficient compared to the Python list. Let’s make some helper functions to compare how long it takes to increment each value in lists, arrays, and NumPy ndarrays accross a broad range of different sizes. Each function will accept a list, array, or ndarray and return the time (in ms) required to increment each value by 1.

In [3]: from datetime import datetime
   ...:
   ...: def inc_list_by_one(l):
   ...:     start = datetime.now()
   ...:     res = [val + 1 for val in l]
   ...:     return (datetime.now() - start).total_seconds() * 1000
   ...:
   ...: def inc_array_by_one(a):
   ...:     start = datetime.now()
   ...:     res = arr.array(a.typecode, [val + 1 for val in a])
   ...:     return (datetime.now() - start).total_seconds() * 1000
   ...:
   ...: def inc_numpy_array_by_one(arr):
   ...:     start = datetime.now()
   ...:     res = arr + 1
   ...:     return (datetime.now() - start).total_seconds() * 1000

Now let’s record the computation times for a range of array sizes and plot them using matplotlib and seaborn:

In [14]: import matplotlib.pyplot as plt
    ...: import seaborn as sns
    ...:
    ...: times = []
    ...: for size in range(1, 10_000_001, 200_000):
    ...:     times.append((
    ...:         size,
    ...:         inc_list_by_one(list(range(size))),
    ...:         inc_array_by_one(arr.array('q', range(size))),
    ...:         inc_numpy_array_by_one(np.arange(size)),
    ...:     ))
    ...: times_arr = np.array(times)
    ...:
    ...: sns.regplot(x=times_arr[:, 0], y=times_arr[:, 1], label='Python list')
    ...: sns.regplot(x=times_arr[:, 0], y=times_arr[:, 2], label='Python array')
    ...: sns.regplot(x=times_arr[:, 0], y=times_arr[:, -1], label='Numpy array')
    ...: plt.xlabel('Array Size (num elements)')
    ...: plt.ylabel('Time to Increment by 1 (ms)')
    ...: plt.legend()
    ...: plt.show()
Computation time v array size

…and now it’s clear that computation times for Python lists and arrays scale with the size of the object much faster than NumPy ndarrays do. And it’s also clear that the Python array is no improvement for computations, static typing notwithstanding. Let’s quantify:

In [26]: print(f'Python list {np.polyfit(times_arr[:, 0], times_arr[:, 1], 1)[0]:.2e} (ms/element)')
    ...: print(f'Python array {np.polyfit(times_arr[:, 0], times_arr[:, 2], 1)[0]:.2e} (ms/element)')
    ...: print(f'NumPy ndarray {np.polyfit(times_arr[:, 0], times_arr[:, -1], 1)[0]:.2e} (ms/element)')
Python list 4.02e-05 (ms/element)
Python array 6.50e-05 (ms/element)
NumPy ndarray 2.01e-06 (ms/element)

So if we take the slope of those curves as a gauge of performance, Python arrays are roughly 40% slower in computation than lists, and the NumPy ndarray computations run in about 95% less time, on average for arrays that are not trivially small. This is because NumPy takes advantage of knowing the data type of each element and setting up efficient strides in memory for computations, which can be optimized at the operating system and hardware levels.

What, then, is the advantage of the Python array over a Python list? With helper functions similar to the ones above, we can return sys.getsizeof() sample list and array objects over the same range of sizes. Plotting the results, we see something interesting:

Memory consumption v array size

Memory consumption for both Python arrays and NumPy ndarrays scale exactly linearly with the number of elements (the q typcode specifies a signed integer with a minimum 8 bytes, which corresponds to the default int64 dytpe for integer NumPy ndarrays.) But notice how the Python list memory consumption increases stepwise. This is an implementation detail of the Python list type that allows it to quickly add elements without shifting memory after each append operation. And this is likely the explanation for why lists can be faster and more efficient than arrays. When a new value is written to an array, it likely often occurs that the some or all of the object needs to be shuffled in memory to find contiguous locations, but a list is optimized for appending, so a certain number of additions can be made before any reshuffling happens.

To test this, we can compare the computation times for an array computed with a list comprehension against an array that is copied and overwritten:

def inc_array_by_one_with_copy(a):
    start = datetime.now()
    res = copy(a)
    for i, val in enumerate(a):
        res[i] = val + 1
    return (datetime.now() - start).total_seconds() * 1000

Plotted, we see that the copy-fill approach is marginally faster, and there is less variation, likely due to fewer memory relocations.

Computation time v array size for arrays created with list comprehension and copy-fill strategies

In conclusion, it helps to understand that the array type is really intended as a Python wrapper around a C array, and there are some functions and codes that use it to efficiently return a Python object. But as for computations, I hope I’ve convinced you that the NumPy ndarray is the right type for computations in terms of both memory and computation efficiency. As for Python arrays, now you know what they are, and you can safely ignore them in favor of the ndarray for your scientific computations.