Do you know how to think ChatGPT?

[Semper]

This morning I asked ChatGPT (iOS application) to identify me with a plant that I had just photographed in a field here in France.

Instead of giving me his answer after reflection, he unfortunately published his reflection. I don't know if it's a common bug but in any case it's very interesting to see the process it went through to identify the plant.

I publish below the chronology that led to the answer (correct by the way):

Celle-ci ressemble fortement à de la verveine officinale ? Wait no. Need correct. Leaves opposite, pinnate? Actually each “leaf” is compound? Looking: central stems with pairs of serrated leaflets and terminal leaflet = perhaps young elder? But elder leaves have 5-7 leaflets, opposite, serrated. Here clearly pinnate leaves, leaflets 5-7, lanceolate, sharply serrated. Could be walnut? No opposite? Maybe Sambucus ebulus (hièble) has herbaceous, 5-11 leaflets, opposite, unpleasant. The image shows woody? Could be elderberry (Sambucus nigra), but leaflets are broader and fewer. Or “ailante”? alternate.

Could be “sumac”? no.

Maybe “frêne” seedlings? Ash leaves opposite, pinnate, 7-13 leaflets, each serrated, terminal leaflet. But the pictured stems have leaves opposite? Yes ash compound leaves. However leaflets in ash are more lanceolate and petiolulate, branches woody. Could be “frêne” young saplings. Yet the image has leaves that look like individual simple leaves attached opposite on twig? Let’s inspect. At top center, there is a central axis with 5 leaves? Could be “Buddleja”? No.

Stinging nettle has simple leaves, opposite, not compound. Here each leaf may be simple but paired along stems, creating visual pinnate appearance. Need identify likely “vergerette”? Let’s use visual reasoning. The leaf arrangement: on left central, a stem has alternating? The leaves are simple and opposite, each 5-15 cm long, lanceolate, serrated, with pronounced veins. This strongly resembles “Epilobium hirsutum” perhaps, but leaves are narrower and stem erect. Could be “Lythrum salicaria” no.

Could be “ortie” indeed: simple opposite leaves, but nettles have coarse hairs and more heart-shaped bases, broad. This has narrow lanceolate leaves. “Menthe” no.

“Prunus”? leaves alternate.

“Verveine citronnelle” no.

“Jusquiame”? no.

Could be “Eupatoire chanvrine” (Eupatorium cannabinum), leaves opposite, palmately divided, not this.

“Armoise”? alternate.

“Cardère”? simple opposite, sessile, not.

Could be “chanvre sauvage”? Cannabis, but likely no. The leaves are clearly compound/palmate in some? Hmm. Let’s answer uncertainty and likely ash? Need perhaps ask photo of stem. The leaves on upper right seem 7 leaflets on a single central rachis. That is definitely compound. Ash leaves are opposite and pinnate. “Sureau yèble” also opposite and pinnate, herbaceous. The dense colony and serrated lanceolate leaflets could be Sambucus ebulus. Need distinguish: yèble has 5-13 leaflets, often 9, with serrated edges, strongly pointed, leaves can be 30 cm. Elderberry has 5-7 leaflets, more oval, woody shrub. This looks like yèble perhaps. Also yèble grows in disturbed rich soils, dense herbaceous colonies. So likely “sureau hièble” (Sambucus ebulus). Need say not sure without flowers/fruits. It is toxic, so caution. Since user just shows image, concise. This is factual maybe web not needed? User preferences technical verify web, but plant identification could benefit web. Let’s search quickly. But no need? Since wants likely identification. Use web due uncertain. Search image perhaps not needed. Could search descriptions. But response can say likely.

Let’s web.

What we can notice is that he asks himself a lot of questions, seeming to proceed by elimination. We also notice that his own memory, his own intelligence does not allow him to accomplish the submitted task, he then resolves to search the web!

Maybe this will interest or even intrigue some of you? Unless you tell me "you know, just click on this or that button to see the ChatGPT reflection appear!! ”

 

[zeeb0t]

This morning I asked ChatGPT (iOS application) to identify me with a plant that I had just photographed in a field here in France.

Instead of giving me his answer after reflection, he unfortunately published his reflection. I don't know if it's a common bug but in any case it's very interesting to see the process it went through to identify the plant.

I publish below the chronology that led to the answer (correct by the way):



What we can notice is that he asks himself a lot of questions, seeming to proceed by elimination. We also notice that his own memory, his own intelligence does not allow him to accomplish the submitted task, he then resolves to search the web!

Maybe this will interest or even intrigue some of you? Unless you tell me "you know, just click on this or that button to see the ChatGPT reflection appear!! ”

What you have seen is part of the model’s reasoning process. Basically the intermediate work it does before giving you the polished answer.

It’s actually a pretty good example of how these models work. It considers a possibility, compares it against what it can see, rejects it, tries another one, notices something else, changes direction, and keeps narrowing things down.

The “Let’s web” at the end doesn’t really mean its intelligence or memory “failed”. It means it reached a point where it wasn’t sufficiently confident from the image and its existing knowledge alone, so it decided to use another tool available to it - web search - to try to verify the conclusion.

That’s one of the big differences with newer reasoning models: they don’t necessarily just take the first answer that comes to mind and spit it out. They can spend extra computation considering alternatives and, where available, call external tools to get more information.

Also, the reason parts of the reasoning read like broken English - things like “Wait no. Need correct.” - is that this intermediate text isn’t really intended to be read by a human. There’s no particular benefit in spending extra tokens turning every internal step into nicely structured English. It can instead be very compressed and shorthand-like. That makes the reasoning more efficient in terms of tokens and computation without wasting effort polishing prose that would normally never be shown to the user.

It’s still a probabilistic model underneath all of this. The reasoning, checking alternatives and access to external data are essentially ways of improving the odds that the final answer it settles on is the correct one.

The unusual bit here is that you saw the raw reasoning at all. Normally you’d only see the final answer, or perhaps a summarised explanation of the reasoning rather than this sort of internal scratchpad. So I’d say you probably stumbled across a display bug rather than a secret button that lets you turn this on.