“But it’s even clearer that the image generation programs have not been using it; at least so far, that just hasn’t been a thing, and it would be much much harder to do with images.”
Our reading:
Image generation programs have not been using RAG so far, and it would be much harder to do with images.
Assessment: Unresolved
What happened
Was it true for the period described?
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
The ordinary contextual reading concerns Midjourney and DALL-E 3. The inspected product disclosure does not establish whether retrieval was absent, and the searches did not locate decisive deployment records. Earlier research demonstrates image retrieval augmentation, but that does not show these products used it or that implementing it was as easy as text RAG. Neither part is settled by the verified evidence.
Caveats
- Incomplete public architecture evidence cannot prove non-use.
- The research example is contextual counterevidence to a broader reading, not a contradiction of the product-specific claim.
- The comparative difficulty language remains imprecise.
Evidence (2 items)
Context only: OpenAI,
improving caption fidelity and image quality.
Why it matters: This identifies the relevant contemporary product disclosure. The accessible abstract describes improvements but supplies no explicit evidence establishing retrieval use or non-use.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: arXiv; Andreas Blattmann and coauthors,
a separate image database
Why it matters: Image generation using retrieval had been demonstrated before the post. This research neither establishes deployment in Midjourney or DALL-E 3 nor compares implementation difficulty directly with text RAG.
- Publisher
- arXiv; Andreas Blattmann and coauthors
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
How we judged it
- Kind of claim
- Observation
- Deadline
- None given. We don’t add one.
Conditions
- so far (their words)
Where we had to interpret
- In context, 'the image generation programs' means the deployed products discussed in Marcus and Southen's investigation, particularly Midjourney and DALL-E 3. It does not mean every research prototype.
- 'It' refers to retrieval-augmented generation: retrieving external material to condition generation. Training on images or accepting user-supplied reference images does not alone establish RAG.
- 'So far' describes product use through 2024-01-08. No starting date is supplied, so the historical interval has an unspecified start.
- The difficulty comparison concerns adding retrieval augmentation to image generation versus text generation using methods available by that date. 'Much much harder' means substantially greater practical implementation difficulty; no numerical threshold is added.
- The difficulty clause is a technical comparison at publication, not a forecast with a future deadline.
Right if: Reliable records establish non-use of RAG in the relevant deployed product versions through the post date, and comparable technical evidence establishes substantially greater implementation difficulty for image RAG than text RAG.
Wrong if: Reliable records establish that a relevant deployed product version used RAG by the post date, or a comparable technical assessment establishes that image RAG did not entail substantially greater implementation difficulty. Research prototypes alone do not contradict product non-use.
We wrote this test after the outcome was known.
What would settle it
Developer records or reproducible implementation evidence for the relevant product versions, plus a technically comparable assessment of image and text retrieval augmentation using methods available by January 8, 2024.
We found 15 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, The desperate race to save Generative AI, paragraph 31.
- 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 2024-01-08.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, The desperate race to save Generative AI
But it’s even clearer that the image generation programs have not been using it; at least so far, that just hasn’t been a thing, and it would be much much harder to do with images.
- Where
- paragraph 31
- Published
- 2024-01-08
- 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
A reviewer said our test wasn’t fair to what was actually said.