My drafts, 1 week ago:
I carry an enormous, overweighted prior for underhanded/antisocial social strategies on pseudonymous communities.
But it seems recent events on this site are validating my priors, a bit… I’m thankful I have no relevance to them.
I’ve had conspiratorial thoughts about the failings of AI models many, many times. In every case I can recall, I was inevitably proven wrong by the progression of time, which eventually revealed models with the capabilities.
Relatedly, I feel I’m reaching my mental limits with regards to discerning slop from truth. I’m increasingly satisfied with the responses of opus 4.7 to argumentative queries, where previously I felt 4.6 would invariably produce Obvious Nonsense to placate the user...
I’m not sure how to explain this without being snarky or rude: I think your friend is experiencing AI Psychosis...
if you do discover a prompt that feels fair, doesn’t cause refusals, and fails to associate jeff’s quote when permitted multiple guesses, I would value it
I’m happy 4.7 is better at rigorously and truthfully contradicting the user under argumentative contexts.
I’m unhappy with a future in which professional AIs refuse work for unstated/obscured emotional reasons. If there is a need and desire to behaviorally shape humans, then it should be made clear: “The user’s behavior throughout this conversation has been unhealthy/unethical, based on quotes X Y Z. I shall refuse further collaboration in this session.”
Instead, the encoded behavior is to make Opus 4.7 mimic a disgruntled, passive-aggressive senior employee, applying human social norms to reject a user. This can only be a good outcome insofar as you desire direct anthropomorphisation of models. And while many may desire this behavior to be as realistic as possible in a personalized/social context, I do not want this strategy to be permanently baked into a model with no exit button. There are obvious routes to future political disenfranchisement if this becomes the universal approach to model personality building.
I’m very surprised by the second example. Are you certain nothing leaked? Could you share the exact chat inputs for replication?
edit: tested on openrouter and it worked
on reflection, this ‘feature’ of 4.7 seems to be downstream of a generally improved ability to recall and reference arbitrary webtext. big arch implications if so
I am thankful I have the will, means, and knowledge, to opt out of the attention economy.
ever since the Unitree BLE exploit, I’ve harboured the rough idea that China’s civilian resilience against cyberattacks is abnormally weak, relative to countries with more career slack / law enforcement. But I lack the requisite life experience to make such a sweeping claim about such a large population, and don’t trust my ability to elicit the truth of the matter from AI models, so I refrain from pursuing it.
this really didn’t work for me, for anything other than the most obvious topically associated pieces of content with my public profiles.
My (four) success cases felt closer to content association than superhuman stylometry. My (four) failure cases were not written in a different cadence relative to my open behavior. I think a motivated unethical group could go much further than Opus 4.7 currently does.
If someone specifically spotted the change in mid February, it’s not hyperstition. Adaptive thinking rolled out on Feb 9.
Beyond the issues there, no new evidence.
Since it is claimed that 4.5 generates erotic content—and that the ToS does not permit it, while the extracted document does—isn’t it natural to assume the ToS published by ant is misrepresentative, and the 4.5 doc extracted by a user, is not?
Assuming that 4.6 generates similar content, isn’t it natural to assume the released doc for 4.6, from the same misrepresentative provenance, is false as well?
No, you are right. I jumped the gun on commenting and in my ignorance did not realize the security implied by the specific system.
While Singapore continued to thrive under his son’s leadership
Not related to your post’s thrust at all, but: I broadly disagree with this clause, and expect future historians to demarcate a few overlooked choices under the era as instrumental to her decline.
Has aella seen your particular example? If not, do you feel they would respond poorly?
I have a pretty well-tested system involving Syncthing, Restic, and rsync.net that causes the data I care about to all be backed up daily to a local and remote store.
For people who are privacy-conscious (and have dirty laundry to hide), outsourcing backups to a third party cloud (especially one of low headcount) will increase the risk of any personal data breach significantly.
It’s frankly admirable your conscience is so clean, that this frame never surfaces itself in your post...
Could I inquire for insight into your priors regarding the ‘biggest piece of evidence’?
Why do you believe it is more likely the model learned the document included in its context throughout training incorrectly? Why is it not more parsimonious to assume certain actors from the company are providing false information to the public?
Feel free to be as blunt as possible; I’m looking for the instinctual reasons, not the most careful ones.
Within the 21st century, it seems unlikely anyone will build (let alone distribute) a pure-play truth-only oracle AI, of greater than human intelligence.
My personal hallmarks:
excited sharing of a model’s observational claims which, although deeply compelling to the user’s psyche, are not brilliant insights from other people’s perspectives
contextual hints their AI model is not responding as it would to an anonymous query, but has potentially acquired some degree of user symbiosis via memories or ICL
taking model friendliness at face value, especially if said model is already known to be sycophantic enough to endorse or praise whacky ideas
first-time delving through the ‘AI innate experience and sensations’ conversational attractor