I dunno. The skeeziest people I know would show up squeaky clean on paper, and several of the ones I trust the most have some kind of shit in their past, at least on paper.
Whatever it is that caused "It's not X, It's Y", it's more recent than LLMs as a whole.
As far as I remember, neither GPT3.5, GPT4, nor Claude Instant did it. I think Gemini was the first to really do it, and then out of nowhere, everybody was doing it.
Then again, sometimes a big feature is so comprehensively broken that it’s hard, from the outside, to break it down into specific flaws. Even if you can reproduce the complex circumstances where they manifest.
In the case of the iOS keyboard, I remember one bug that made the rounds (in the popular press!) after somebody recorded their typing in slow motion to validate it [0]. Once they documented it, everybody recognized the feeling and felt vindicated; but it took actual work to substantiate.
That’s the work it seems that Apple engineers should be doing. They have the telemetry, the source access, the design documents, the labs, and the time in their day to make a comprehensive study of it. Just as I can say “my car is handling funny around turns” and let it be the mechanic’s job to diagnose what’s wrong in mechanical terms.
There was a time when this humane aspect was Apple’s particular magic: engineering beyond technical requirements to the point of simplicity, ergonomics, “it just works”…
I guess the hope is that, in a healthy media ecosystem, anything important enough to have a constituency gets reported truthfully and legally fact-checked by somebody.
The public are still free to care which truthful things they care about (or want to pay attention to)—and part of the job of politics is still to try to direct attention toward aspects of truth that favor your political aims. But with sufficiently many truth-motivated reporting organs reflecting sufficiently many constituencies, the work of truth-finding gets done.
When you modify their braking behavior, is that enough to improve their overall driving behavior? Or do forward collisions and rear-enders make up substantially all of what the driver can control, so training the behaviors to reduce that type of near-miss reduces the driver's overall crash risk? To the point that it's similar to the safest tranche?
Is it that hard braking events are broadly indicative of surprises of lots of sorts, and so it happens that the only way to eliminate them all is to develop a full range of defensive driving habits?
Regardless of everything else, forward collisions are most likely to have the driver considered at-fault. Seems like reducing those in your insured population would reduce covered losses more than reducing collisions where your insured may not be at fault.
> Is it that hard braking events are broadly indicative of surprises of lots of sorts,
Of apparent surprises to the driver. And since actual, factual surprises are extremely rare, if a driver is regularly being surprised, they're a bad driver.
Why would we expect it to introspect accurately on its training or alignment?
It can articulate a plausible guess, sure; but this seems to me to demonstrate the very “word model vs world model” distinction TFA is drawing. When the model says something that sounds like alignment techniques somebody might choose, it’s playing dress-up, no? It’s mimicking the artifact of a policy, not the judgments or the policymaking context or the game-theoretical situation that actually led to one set of policies over another.
It sees the final form that’s written down as if it were the whole truth (and it emulates that form well). In doing so it misses the “why” and the “how,” and the “what was actually going on but wasn’t written about,” the “why this is what we did instead of that.”
Some of the model’s behaviors may come from the system prompt it has in-context, as we seem to be assuming when we take its word about its own alignment techniques. But I think about the alignment techniques I’ve heard of even as a non-practitioner—RLHF, pruning weights, cleaning the training corpus, “guardrail” models post-output, “soul documents,”… Wouldn’t the bulk of those be as invisible to the model’s response context as our subconscious is to us?
Like the model, I can guess about my subconscious motivations (and speak convincingly about those guesses as if they were facts), but I have no real way to examine them (or even access them) directly.
Isn’t that what the parent is describing? “Ill and cancelled projects” <==> “luck surface area”, and “trying to go for promotion” <==> “cozy vibes and self-interested comfort”?
reply