“New Ways to Corrupt LLMs”
We recorded 6 claims from this post. 2 were checked against evidence. 4 can’t be checked. 3 more are held. Read the post at its source.
Claims from this post (6)
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
“The problem with generative AI has always been that large language models associate patterns together without really understanding those patterns; it’s statistics without comprehension.”
What we tested: Large language models associate patterns without really understanding them; they are statistics without comprehension.
Assessment: UnresolvedSaid
“A lot of hallucinations are borne of exactly this kind of overgeneralization.”
What we tested: Many LLM hallucinations result from overgeneralizing word correlations such as semantic leakage.
Assessment: Can’t be checkedSaid
“As Evans noted, this is no joke. A bad actor could easily use this techniques to do nasty things:”
What we tested: Bad actors could easily use subliminal learning techniques to corrupt LLMs for harmful purposes.
Assessment: Can’t be checkedSaid
“And again their is an avenue that could be easily exploited.”
What we tested: Weird generalization in LLMs is an avenue that bad actors could easily exploit.
Assessment: UnresolvedSaid
“There is no way in Darwin’s green earth that we are ever going to be able to patch what is likely to be an endless list of vulnerabilities.”
What we tested: LLM vulnerabilities from semantic leakage are likely endless and will never all be patched.
Assessment: Can’t be checkedSaid
“Putting society in the hands of giant, superficial correlation machines is not going to end well.”
What we tested: If society is put in the hands of giant, superficial correlation machines such as LLMs, it is not going to end well.
Held from this post (3)
These records did not pass every check, so nothing about them is published. Held is not a finding.
- The two reviews reached different assessments. Our reading, not assessed: If you extract weird correlations from one machine, you can feed them into another and bend it to your will.
- The criteria are ambiguous. Our reading, not assessed: In errors like semantic leakage, LLMs are not picking up on real correlations in the world; they are learning weird nth-order correlations between words rather than concepts.
- The record does not meet the current rules. Our reading, not assessed: Suno's meagre copyright defenses can be circumvented by adversarial use of statistical correlates.