“The problem was not on the image generation side, it was in the connection between language and images.”
Our reading:
Imagen's failure lies in the connection between language and images, not in image generation.
Assessment: Unresolved
What happened
Does the evidence support the cause?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
On May 25, 2022, Neyshabur reported that Imagen could depict a horse riding an astronaut after a small prompt change [e1]. This demonstrates some capacity to produce the scene, but does not establish why the original prompt failed [e1]. The demonstration neither locates that failure in the connection between language and images nor excludes an image-generation limitation [e1]. The claimed cause remains unresolved [e1].
Caveats
- The demonstration reports no sample counts, success rates, seeds, or selection procedures [e1].
- It provides no controlled comparison that separates language-to-image mapping failures from image-generation failures [e1].
Evidence (1 item)
Context only: Behnam Neyshabur,
minimal change
Why it matters: This supports the narrower premise that changing the prompt can produce the intended scene. It does not isolate which process caused the original failure.
- Publisher
- Behnam Neyshabur
- Kind of source
- Demonstration
- About events on
- Published
How we judged it
- Kind of claim
- Causal claim
- Deadline
- None given. We don’t add one.
Where we had to interpret
- The claim concerns the original Imagen and the May 2022 horse-and-astronaut example.
- I read the distinction as a causal claim about translating the instruction into an image versus generating the scene. It does not specifically name the text encoder.
- Success with another prompt establishes some capacity to depict the scene, but does not by itself identify the cause of failure with the original prompt.
Right if: Comparisons on the original Imagen establish that it can depict the requested scene and locate the original failure in language-to-image mapping, excluding an image-generation limitation as the explanation.
Wrong if: Controlled evidence on the original Imagen shows that the prompt's meaning was adequately conveyed and that an image-generation limitation caused the failure.
We wrote this test after the outcome was known.
What would settle it
Controlled tests on the original Imagen using matched prompts, repeated samples, and disclosed selection procedures. Measurements or interventions must establish whether language-to-image mapping caused the original failure or whether generation failed despite adequate conveyance of the prompt's meaning.
We found 13 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, Horse rides astronaut, paragraph 82.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-29. The post may have been edited after 2022-05-28.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, Horse rides astronaut
The problem was not on the image generation side, it was in the connection between language and images.
- Where
- paragraph 82
- Published
- 2022-05-28
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
The final check failed: our summary didn’t match their words.