Are LLMs able to notice the “gorilla in the data”?

That's such a classical human behaviour in technical discussions, I wouldn't even be mad. I'm more surprised that picked up on that behaviour from human generated datasets. But I suppose that's what you get from scraping places like Stackoverflow and HN.

Now the society's time and energy has shifted from general scientific progress to gaining expertise in the growing patchset used to rationalize the theory that the LLM possesses intelligence.

The plot would turn when Picard tries to wrest a phasor from a rogue non-believer trying to assassinate the Queen, and the phasor accidentally fires and ends up frying the entire LLM patchset.

Mr. Data tries to reassure the planet's forlorn inhabitants, as they are convinced they'll never be able to build the warp drive now that the LLM patchset is gone. But when he asks them why their prototypes never worked in the first place, one by one the inhabitants begin to speculate and argue about the problems with their warp drive's design and build.

The episode ends with Data apologizing to Picard since he seems to have started a conflict among the inhabitants. However, Picard points Mr. Data to one of the engineers drawing out a rocket test on a whiteboard. He then thanks him for potentially spurring on the planet's next scientific revolution.

Fin

Lots of front ends will do tricks like partially loading the file or using a cached version or some other behavior. Plus if you presented the file to the same “thread” it is possible it got confused about which to look at.

These front ends do a pretty lousy job of communicating to you, the end user, precisely what they are pulling into the models context window at any given time. And what the model sees as its full context window might change during the conversation as the “front end” makes edits to part portions of the same session (like dropping large files it pulled in earlier that it determines aren’t relevant somehow).

In short what you see might not be what the model is seeing at all, thus it not returning the results you expect. Every front end plays games with the context it provides to the model in order to reduce token counts and improve model performance (however “performance gets defined and measured by the designers)

That all being said it’s also completely possible it missed the gorilla in the middle… so who really knows eh?

This is, quite literally, not something that ChatGPT has the capability to do -- reporting on its thought process, that is. This is a hallucination.

Eg, from uploading the gorilla scatterplot to gpt4o and asking "What do you see?"

"The image is a scatter plot of "Steps vs BMI by Gender," where data points are color-coded:

Blue (x) for males

Red (x) for females

The data points are arranged in a way that forms an ASCII-art-style image of a "smirking monkey" with one hand raised. This suggests that the data may have been intentionally structured or manipulated to create this pattern.

Would you like me to analyze the raw data from the uploaded file? "

I have custom instructions that would influence its approach. And it does look more like a monkey than a gorilla to me

But imagine a new employee eager to please - you could easily imagine them OK’ing the document and making the same assumption the LLM did - “why would you randomly throw in that word if it wasn’t relevant”. Maybe they would ask about it though…

Google search has the same problem as LLMs - some meanings of a search text cannot be de-ambiguified with just the context in the search itself, but the algo has to best-guess anyway.

The cheaper input context for LLMs get, and the larger the context window, the more context you can throw in the prompt, and the more often these ambiguities can be resolved.

Imagine in your gorilla in the step example, if the LLM was given the steps, but you also included the full text of slack/notion and confluence as a reference in the prompt. It might succeed. I do think this is a weak point in LLMs though - they seem to really, really not like correcting you unless you display a high degree of skepticism, and then they go to the opposite end of the extreme and they will make up problems just to please you. I’m not sure how the labs are planning to solve this…

Try sending this graph to an actual human analyst. His response, after you paying him will probably be to cut off any further business relationship with you.

[1]: https://en.wikipedia.org/wiki/Anscombe's_quartet [2]: https://en.wikipedia.org/wiki/Datasaurus_dozen

And even when explicitly prompted to look at the plot, they only brush up against the data anomalies rather than properly analyzing the plot.

It's not silly at all.

o1 pro: It appears the data have been deliberately arranged to form a stylized humanoid figure. In terms of actual trends, there is no obvious linear or nonlinear correlation between steps and BMI, and the female–male color split seems fairly uniform across the shape. Visually, BMI spans roughly 15–30, while steps extend up to about 15,000, with no clear clustering or separation by gender. Within reasonable confidence (e.g., an estimated R² near zero if one were to attempt a linear fit), there is no discernible predictive relationship between steps and BMI in this plot.

The contents of the CSV might be entirely unknown at inference time.

> I asked the model to closely look at the plot, and also uploaded a png of the plot it had generated.

The conclusion is quite reasonable and the article was IMO well written. It shares details of an experiment and then provides a thoughtful analysis. I don’t believe the analysis is overly broad.

A corollary if this that is my personal pet peeve is attributing everything you can’t explain to “seasonality” , that is such a crutch. If you can’t explain it then just say that. There is a better than not chance it is noise anyway.

In general, I’ve steered clear of current LLMs for data analysis/description because they seem so highly influenced by choice of prompt and wording. They tend to simply affirm any language I use to describe the data initially.

To be fair, I’ve attended conferences and lab meetings where humans will refer to a any vaguely concave curved distribution as “mostly normal” :P

But I guess GPT-4o results are more funny to look at.

They do seem to generally have legs and head, which is an improvement over 4o. Still pretty unimpressive.

Not the end of the world but it does seem like AI gets fixated, like people, and can’t see anything else.

It has no problem identifying our other two dogs as actual dogs who bark and move like dogs.

The gorilla is just an extreme example of that.

Albeit perhaps an unfair example when applied to AI.

In the original experiment with humans, the assumption seemed to be that the gorilla is fundamentally easy to see. Therefore if you look at the graph to try to find patterns in it, you ought to notice the gorilla. If you don’t notice it, you might also fail to notice other obvious patterns that would be more likely to occur in real data.

Even for humans, that assumption might be incorrect. To some extent, failing to notice the gorilla might just be demonstrating a quirk in our brains’ visual processing. If we expect data, we see data, no matter how obvious the gorilla might be. Failing to notice the gorilla doesn’t necessarily mean that we’d also fail to notice the sorts of patterns or flaws that appear in real data. But on the other hand, people do often fail to notice ‘obvious’ patterns in real data. To distinguish the two effects, you’d want a larger experiment with more types of ‘obvious’ flaws than just gorillas.

For AI, those concerns are the same but magnified. On one hand, vision models are so alien that it’s entirely plausible they can notice patterns reliably despite not seeing the gorilla. On the other hand, vision models are so unreliable that it’s also plausible they can’t notice patterns in graphs well at all.

In any case, for both humans and AI, it’s interesting what these examples reveal about their visual processing, which is in both cases something of a black box. That makes the gorilla experiment worth talking about regardless of what lessons it does or doesn’t hold for real data analysis.

"AI can't see gorilla because wokification"?

/s

Edit: Adding /s Thought "wokification" already signalled that.

You just assumed it’s similar to the Google situation a few years back where they banned their classifier from classifying images as gorillas. It isn’t.

"Google promised a fix after its photo-categorization software labeled black people as gorillas in 2015. More than two years later, it hasn't found one."

Companies do seem to have developed greater sensitivity to blind spots with diversity in their datasets, so Parent might not be totally out of line to bring it up.

IBM offloaded their domestic surveillance and facial recognition services following the BLM protests when interest by law enforcement sparked concerns of racial profiling and abuse due in part to low accuracy in higher-melanin subjects, and Apple face unlock famously couldn't tell Asians apart.

It's not outlandish to assume that there's been some special effort made to ensure that datasets and evaluation in newer models don't ignite any more PR threads. That's not claiming Google's classification models have anything to do with OpenAI's multimodal models, just that we know that until relatively recently, models from more than one major US company struggled to correctly identify some individuals as individuals.