Horse rides astronaut
We recorded 11 claims from this post. 6 were checked against evidence. 5 can’t be checked. 6 more are held. Read the post at its source.
Claims from this post (11)
In the order they appear in the post. Each quotation is the speaker’s words in this post. A label such as “Supported” or “Can’t be checked” is the claim’s current assessment; its page has the evidence and the reasons.
Assessment: Can’t be checkedSaid
Observation
“Judging by all the premature declarations I have seen from the machine learning community of late, I am not sure AI has yet fully learned it.”
Our reading: The machine learning community has recently made premature declarations and may not have learned the Clever Hans lesson about overclaiming.
No horizon
Assessment: UnresolvedSaid
Observation
“As Ernie Davis, Scott Aaronson, and I showed, a week or so after the release, it’s easily fooled, especially with complex sentences and words like “similar” that would require a deep understanding of language to be interpreted correctly”
Our reading: DALL-E 2 is easily fooled, especially by complex sentences and words requiring deep language understanding.
No horizon Evidence as of
Limitation: Unresolved does not mean the claim was false.
Assessment: Can’t be checkedSaid
Observation
“DALL-E 2 is, I dare say, not as smart as it seems.”
Our reading: DALL-E 2 is not as smart as it appears.
No horizon
Assessment: Can’t be checkedSaid
Recommendation or value judgment
“Against a long history of overclaiming in AI, such strong statements demand an awful lot of evidence. Because Google won’t allow skeptics to try the system, we can only make guesses. But from the few clues that have filtered out, there is enormous reason for skepticism.”
Our reading: From the few clues that have filtered out, there is enormous reason to be skeptical of Google's claim that Imagen has a deep level of language understanding, though without access to the system skeptics can only make guesses.
No horizon
Qualifier: from the few clues that have filtered out
Assessment: UnresolvedSaid
Observation
“(In reality, Imagen has no idea what weight, mass, or density are.)”
Our reading: Imagen has no understanding of weight, mass or density.
No horizon Evidence as of
Limitation: The behavioral interpretation makes the claim testable but does not settle philosophical questions about understanding.
Assessment: UnresolvedSaid
Causal claim
“And, no, it wasn’t about common sense at all.”
Our reading: Imagen's failure on 'a horse riding an astronaut' was not due to common sense.
No horizon Evidence as of
Limitation: The demonstration provides no repeated-sample success rates or controlled comparison separating wording from scene plausibility [e5].
Assessment: SupportedSaid
Observation
“That’s where the deep understanding of language should live—but, despite the PR, it isn’t there.”
Our reading: Imagen lacks deep understanding of language despite Google's PR.
No horizon Evidence as of
Limitation: The result depends on interpreting deep understanding through the role-and-relationship behavior discussed in the post.
Assessment: Can’t be checkedSaid
Observation
“The answer to the second question—does Imagen have a deep understanding of language—is almost certainly “no”: if a system can draw a given picture with incantation X, but not with perfectly sensible alternative incantation Y, it’s pretty hard to say that the system understands Y.”
Our reading: Imagen almost certainly does not have a deep understanding of language.
No horizon
Assessment: UnresolvedSaid
Causal claim
“Since we know that Imagen can draw images of horses riding astronauts (if the instructions are specific enough) we know that the failure to draw a horse riding an astronaut given the prompt “a horse riding an astronaut” can only have one explanation: the system doesn’t understand what the prompt “a horse riding an astronaut” means.”
Our reading: Since Imagen can draw horses riding astronauts when the instructions are specific enough, its failure to draw one from the prompt 'a horse riding an astronaut' can only be explained by the system not understanding what that prompt means.
No horizon Evidence as of
Limitation: The available demonstration is not a controlled causal experiment.
Assessment: Can’t be checkedSaid
Causal claim
“it casts doubt on “an astronaut riding a horse”, and suggests success there is not because the network knows how to extract meanings from individual words”
Our reading: Imagen's failure on 'a horse riding an astronaut' casts doubt on its success with 'an astronaut riding a horse' and suggests that this success comes not from extracting word meanings and combining them by syntax, but from something more holistic and approximate, a bit like keyword matching.
No horizon
Qualifier: casts doubt / suggests
Assessment: UnresolvedSaid
Causal claim
“Furthermore, there is no step in the process to try to derive such structures; so it is just wishful thinking to expect them to be fully sensitive to semantics as it relates to syntactic structure.”
Our reading: Because the systems under discussion (Imagen, DALL-E 2 and similar architectures) represent phrases as strings and have no step to derive hierarchical syntactic structures, it is just wishful thinking to expect them to be fully sensitive to semantics as it relates to syntactic structure.
No horizon Evidence as of
Limitation: The supplied architecture evidence concerns Imagen and does not independently document DALL-E 2's pipeline (e1).
Held from this post (6)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The record does not meet the current rules. Our reading, not assessed: Imagen fails to draw a horse riding an astronaut from that prompt.
- The record does not meet the current rules. Our reading, not assessed: It remains to be seen whether Imagen can in general be coaxed into producing particular images by sufficiently fluent users, and the answer will help determine whether artists stick with these systems after the novelty wears off.
- The criteria are ambiguous. Our reading, not assessed: Imagen's failures on compositional block-stacking prompts would likely be easy to replicate.
- The two reviews read the claim differently
- The record does not meet the current rules. Our reading, not assessed: There is no principled reason to think that Imagen's architecture is profoundly and relevantly different from other architectures, such as DALL-E 2, that have already been shown to have similar problems.
- The criteria are ambiguous. Our reading, not assessed: No one yet knows how to build a system that derives semantics of wholes from parts as a function of syntax, which is what is needed.