“One can imagine an alternative AI paradigm that could learn much of what it needed to know to be effective in the world without going through all your personal data, but that does not yet exist.”
Our reading:
An AI paradigm that could learn much of what it needed to be effective in the world without going through all of a person's personal data is imaginable but does not yet exist.
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 categorical nonexistence claim remains unsettled. DreamCoder and I-JEPA show that alternative learning approaches existed, but their documented applications do not establish the broadly effective replacement described in context. Gato addresses broader capabilities without establishing the required alternative paradigm and personal-data independence. These findings neither demonstrate a qualifying counterexample nor justify concluding that none existed.
Caveats
- The search is not an exhaustive census of AI research.
- The cited records provide historical context rather than decisive outcome evidence for the post-date snapshot.
- A narrower interpretation of effectiveness would change the test and could unfairly turn specialized systems into counterexamples.
Evidence (3 items)
Context only: arXiv,
drawing pictures and building scenes.
Why it matters: DreamCoder documents an alternative learning approach with several demonstrated applications. These results do not establish the broad effectiveness and personal-data independence required to contradict the contextual reading.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- About events on
- Published
Context only: arXiv / Transactions on Machine Learning Research,
generalist policy.
Why it matters: Gato is relevant to the breadth requirement, but the paper describes extending the approach used in large-scale language modeling. Its generalist designation does not establish a broadly effective alternative independent of extensive personal information.
- Publisher
- arXiv / Transactions on Machine Learning Research
- Kind of source
- Peer-reviewed paper
- About events on
- Published
Context only: arXiv,
self-supervised learning from images.
Why it matters: I-JEPA documents an alternative approach within image learning. That domain-specific demonstration does not establish a broadly effective replacement meeting the claim's personal-data requirement.
- Publisher
- arXiv
- 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
- much of what it needed to know (their words)
- without going through all your personal data (their words)
- not yet (their words)
Where we had to interpret
- The firm assertion concerns the nonexistence of a demonstrated alternative as of September 7, 2023.
- The ordinary contextual reading is a broadly effective alternative to prevalent data-intensive AI. A useful system confined to a narrow task does not by itself contradict it.
- I read 'all your personal data' as extensive use of personal information, rather than literally every datum about a particular person. Public availability alone does not establish that training data contain no personal information.
- Research has no regular reporting period. I use the post date as the snapshot for 'not yet'; earlier demonstrations provide historical context.
- Marcus supplies no numerical threshold for breadth, effectiveness, or personal-data dependence. I do not impose one.
Right if: A sufficiently comprehensive contemporary assessment establishes that available alternative approaches had not demonstrated broad effectiveness without extensive dependence on personal information. Merely failing to find a counterexample is insufficient.
Wrong if: A documented alternative existing by the post date demonstrates broadly effective learning across varied domains, with training-data evidence establishing that it learned much of what it needed without extensive use of personal information. An isolated narrow success is insufficient.
We wrote this test after the outcome was known.
What would settle it
Documentation of an alternative existing by September 7, 2023 that combines broad demonstrated effectiveness with adequate training-data provenance, or a comprehensive contemporary assessment establishing that available alternatives lacked that combination.
We found 4 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, The dirty underbelly of AI, paragraph 10.
- 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 2023-09-07.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, The dirty underbelly of AI
One can imagine an alternative AI paradigm that could learn much of what it needed to know to be effective in the world without going through all your personal data, but that does not yet exist.
- Where
- paragraph 10
- Published
- 2023-09-07
- 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.