Not sure if anyone else saw the OpenAI 'Spring Update'/GPT-4o?

Not sure if anyone else watched the live OpenAI introduction/announcement of GPT-4o the other day ?

Introducing GPT-4o (

While I generally appreciate the ‘frankness’ of the demo (i.e. it obviously wasn’t totally canned as others have fallen victim to), and others have gone into more depth in critiquing the new model itself in action (Hot take on OpenAI’s new GPT-4o - by Gary Marcus), there was one point about the demo that really struck me that I have not yet at least seen anyone talk about.

This starts at around 17:10 where they ask it to analyze a Python program and later the graph produced by it. Hallucinations from ‘nowhere’ aside (where it, out of the ether, complements Barrett on his appearance), it is the analysis of the graph I wanted to talk about.

While they do not make reference to it, this is basically some boilerplate code for the widely used Meteostat Python Database/Library, and they preserve the same longitude, latitude, and altitude Meteostat uses in their site example code: Python Library | Meteostat Developers

The city selected turns out to be Vancouver.

Yet it is the moment when they ask Gpt-4o to provide an interpretation from the resulting graph alone we should focus on (starting at 20:08 At first, I would say the general overall interpretation is correct.

However, before they abruptly cut GPT-4o off, it starts to extrapolate and go on to say there appears to be a notable rainfall event at the end of the September.

And at first I thought– What ? Is GPT-4o some ‘magical’ new Willard Scott ? This graph contains data on temperature and nothing else. How is it possibly speaking about rainfall (without more data) ?

So I decided to look it up on year (2018) and month (September) cited and this is what we get:


  1. While there was a warning issued about a large amount of rain in late September (Up to 60 mm of rain to fall in Metro Vancouver by Monday morning | News):

Turns out the total amount of rain that month ( of 91.4 mm is not exceptional considering the results in 2019 (97.4 mm), 2021 (100.2 mm), 2022 (42.3 mm), or 2023 (59.5 mm) [Note: 2020 appears to be something of an outlier at only 13.5 mm].

Nor is it usual for rain patterns to be more heavily distributed towards the end of the month:

  1. Though back to the greater point here: While Meteostat is a widely used database and it may have just ‘memorized’ it at some point during training– How on Earth is it deducing rainfall from temperature? I mean just from a data governance point of view, we did not even give it access to that information.

And from an analysis point of view, how can you possibly deduce rainfall events from temperature alone? It was asked to use solely the graph (not the code) to perform the interpretation. What if I were to say the data in this graph (which is the only thing it is asked to pass interpretation on) came from the desert?

And analyzing the weather of course is an extremely simple case of data analysis.

In any case, I was a little surprised no one else seemed to notice this.

At least I don’t feel the machines are coming for my career as a Data Scientist– (At least not yet–).

1 Like

The silver lining here is that human intelligence still beats artificial intelligence. Who would have thought?

1 Like

@lukmanaj oh I wouldn’t question/wasn’t expecting that.

It is more that:

  1. They’d let this happen in their own demo/introduction to a new product.

  2. I haven’t see anyone else seem to notice the discrepancy here.

Yes, it’s surprising this occurred. In some demos, achieving the same level of output can be challenging, leading one to suspect staging.
This situation with OpenAI is peculiar. Moreover, with comments disabled on YouTube, those who noticed discrepancies couldn’t voice their observations.