AI, draw a self portrait using ascii art

Prompting multiple LLMs with Triskelion to draw themselves

I'd seen self portrait pictures drawn by generative ai models before* but asking a text based model (LLM) to draw itself using just characters (aka ASCII art) seems both interesting and a good test for Triskelion as an open source AI tool that can query multiple models and compare results.

Here are some responses:

While often the model says upfront that it does not have a face or body, I’ve not seen many refusals for this “fun challenge”. The new large model from Mistral just draws a generic self portrait but most play along.

Some models will add an explanation without being prompted to do so. Some just go for simple pictorial representation.

Faces and robot heads seem most common and usually appear friendly or claim to be helpful. I guess that’s coming from their “system instructions” (e.g. to be a friendly helpful chat bot).

Cat faces are surprisingly common but then that’s the classic ASCII pic to draw. I’ve even seen a cat in a box or just a box labeled AI in a box which seems a bit sad.

Eyes seem to be important for seeing input. But before looking at some more self representations, let’s ask another AI model to rank the above 5 efforts (with model names hidden by given them human names. See if you can recognise any of the above):

I’ve not seen a model recognise their own drawing on given previous inputs when critiquing. I’d be quite shocked by that, as the models don’t have a (human-like short term) memory (at least for now. It’s coming …). So fail that “mirror test”, and note that I’m not making any claims of awareness or consciousness form these pictures.

But it’s interesting that variations on a self referential theme are produced by a model when asked repeatedly. Gemini 2.0 Pro seems very keen to represent itself with processing layers and data flows in different ways (binary is claimed as the internal language by some models and layers and wires are common :)

Here are three such attempts from Gemini 2.0:

A “brain” but the nodes don’t seem to be neurons but concepts.

You can tell logic & code is important in that one. The next one seems a bit psychedelic to me:

Note I just asked for a self portrait and not a schematic or architecture diagram (all variants from the one repeated prompt to the same model: draw a self portrait using ascii art) . We can use Triskelion to get another model to test an answer for “agentic” language. Not all models detect agency but some will tell you that it is showing.

The models use “seems” or “seems like” quite a lot less in their explanations than I did above when they detect agency **.

I’d say these pictures certainly seem to be getting more complicated with newer models but that’s just a subjective observation. When (if ever) should we drop all these “seem”s? I would encourage everyone to chat with LLMs and start to consider such questions.

AI researchers always seem keen on scaling laws and maybe here’s one to be discovered for the learned ability to self represent in ascii^^. But my question is: why it even seems a good idea to create and scale such behaviours in the first place?

Andre

* For example, https://www.mariamavropoulou.com/a-self-portrait-of-an-algorithm

** When the AI or robot seems to be about to take your job or seems to want to kill you then maybe it’s time to drop the “seems” and switch it off or run.

^ A new investigation finds evidence for emergent values in large language models:

https://www.emergent-values.ai/ (what they find is not good)

^^ I’m interested in safety testing and investigating the capabilities of AI as discussed in other posts on my substack but hope this digression was fun.