The following is a manually-constructed structured debate map. Hover over icons/symbols for their meaning.
- The exception is "⯈", which is rendered by the browser-engine and indicates an openable branch.
1. 🗶 LLMs definitely cannot think.
- 🞛 🡇
1.1 ✔ Discussion of whether LLMs can think is meaningless without an objective definition of "think", which we do not have here: see argument #A.1.
🡅 1.2 🗶 LLMs assemble words without thinking: essay/01
🡇 1.2.1 This is primarily a reassertion of the root conclusion (#1.).
- 🛈 All it adds is a description of the process ("assembling words" being a summary of the process described much more completely in the essay). This seems to be making the implicit argument that "if you can understand it or describe it in a way that is objective, logical, or deterministic, then it can't be 'thinking'."
🡇 1.3 ✔ Assembling a test for a negative is logically impossible; there is no way to prove the definite absence of a phenomenon.
- 🛈 We can, however, either demonstrate that the preponderance of evidence overwhelmingly suggests its absence, if we have a test for it: see argument #
3.
🡅 1.4 🗶 It's not thinking because it's an algorithm that evaluates to zero.
- 🞛 🡇
1.4.1 This implies an agreed-upon definition of "thinking" which is mutually exclusive with "an algorithm that evaluates to zero".
🡅 1.5 🗶 Thinking has emergent properties [therefore LLMs aren't thinking].
🡅 1.5.1 🗶 (implied) LLMs do not have emergent properties.
- 🞛 🡇
1.5.1.1 Clearly LLMs do have emergent properties, or we'd know how to reliably train them to do what we want.
2. 🗶 We have not seen any evidence that LLMs are able to think.
- 🞛 🡇
2.1 ✔ Discussion of whether LLMs can think is meaningless without an objective definition of "think", which we do not have here: see argument #A.1.
3. ✔ The only way to agree on whether or not an LLM can "think" is to have an agreed-upon definition of "thinking".
🡇 3.1 🗶 Assembling a test for a negative is logically impossible, as there is no way to prove the definite absence of a phenomenon; this is therefore not a good way to determine the absence of "thinking".
- 🞛 🡇
3.1.1 We are not trying to prove the definite absence of "thinking", but rather to demonstrate that it does not appear to be happening.
- 🞛 🡇
3.1.2 With an agreed-upon test or definition, we can nonetheless establish whether there is any evidence to support the idea that the phenomenon is occurring.
- Absence of evidence in sufficient quantities becomes evidence of absence. If you repeatedly test for something and the test always comes up negative, you've at least shown that the phenomenon is unlikely, and that any claim that it is happening goes against the available evidence.
- This "claiming against the available evidence" has proven to be a clear red-flag for bad-faith discussion. If I haven't written about this on Issuepedia, then I need to...
- 🞛 🡅
3.2 It is pointless to discuss the existence of something if the participants don't all agree on what the thing is.
|
Oft-repeated sub-argument
A.1. ✔ The discussion of whether LLMs can think is meaningless without an agreed-upon definition of "think", which we do not have.🝢⯯
🡅 A.1.1🝢⯮ It is a logical error to assert the existence or nonexistence of a phenomenon without defining what is required in order to detect its presence.
- 🛈 IOW: You can't say either "it's doing $X" or "it's not doing $X" without defining $X.
- 🞛 🡅
A.1.2🝢⯮ None of the arguments so far advanced have specified a clear definition of "to think", much less a test to detect its presence.
|
Subpages
- essay/01: 3rd party argument against AI thinkage
Definitions
- of cognitive terms:
- "thinking":
- "thought": Wikipedia
- "understand"
- "know"
- "believe"
- "concept"
- "idea"
- "cognition"