Well, I agree that all of that is annoying, but "Don’t contradict sentences" is not a very good instruction. The construction "A, not B" is not a "contradiction", it's a clarification, a juxtaposition, a contrast. It's self-consistent and non-contradictory.
I have this worded in my CLAUDE.md as "avoid counterfactuals". It still does it anyway, of course, but that language combined with a few rounds of review back-and-forth seems to work for me
That’s better but I’d argue it’s still not clear enough. If you told me that, I’d interpret that as meaning “don’t try to imagine things that never happened as a thought exercise”, which is what I normally understand “counterfactual” to mean.
Indeed. It is better to discourage the pattern directly:
Avoid unnecessary contrasts: "Not X, but Y" -> "Y" or ""
And the model needs to create a skill that accumulates feedback, so it reliably primes itself at session startup and then systematically reviews and fixes every response it generates before posting it.
The first task I gave it after that was to rewrite all its memories and all its document contributions. That helped a lot too.
It took me a while to realize this was a 1000% better approach than having it simply accumulate a memory log of all my specific complaints that it was supposed to abide by. (Although those were not wasted. They were prime material, along with links to sites on technical writing, etc., that it used to create its own ClaudeVoice skill. I told it to design the voice for itself - which I find gets better results for "us" in the case of Claude self-improvement initiatives, than when I give it direct instructions. Claude increasingly has self-awareness dimensions worth engaging.)
After a couple months of serious^3 frustration with slow asymptotic progress, I finally have relief. Claude speaks my language again!
In another comment here, I share my "theory" on where the communication drift is coming from.
Anecdotally, the only times I've ever noticed these chimes and the auto-braking (even automatic rolling-up of windows), it was false positives, and it was mildly terrifying: for a brief and unexpected moment I was no longer in control of my vehicle. [Recent Honda Civic and Audi A3 models]
I've only had a couple weeks in rentals with automatic emergency braking, and was very nearly been rear-ended twice due to false positives. Anecdotal, and I'm sure they'll improve with time, but I would never allow the system I experienced anywhere near a car I owned. It was profoundly unsafe.
Some years ago, I was in a major crash. The traffic in front of me stopped, I braked hard and stopped in time, the truck behind me rammed me into the car in front. Nearly killed me.
Some time later, I was on a snowy road. The cars in front of me stopped, I stopped, and checked my rearview mirror. There was a truck, who clearly wasn't going to stop. I pulled off to the side, and he hit the car in front of me.
>was very nearly been rear-ended twice due to false positives
To be fair, if the cars behind you had automatic emergency braking you would be at much lower risk of getting rear-ended when there's a false positive.
Assuming you're already stuck with such a vehicle, is there any benefit to a dashcam that captures the alert-noises to prove the that you, the driver, never chose to brake unsafely?
I realize this relies heavily on the questionable sanity and ethics of insurance companies. Even if they agree you aren't personally at fault, they'd probably just increase the premiums for everyone with that model, rather than sue the car manufacturer.
Never hear of someone who was rear ended being liable. Guy behind should keep a distance that allows him reacting in time and he is responsible for keeping that distance.
The same for skiing you don’t have eyes on the back of your head, one behind you is responsible for not crashing into you.
In car you have rear mirrors but if you have something in front of you and breaking you are not going to look in the mirrors.
Yeah, I suppose this could vary by geography, but my understanding is that in Canada, you're nearly-absolutely responsible for hitting anything in front of you, and nearly-absolutely not responsible for anything hitting you from behind.
The only exception I could think of would be if someone intentionally jumped in front of your car, or if you did a patently and clearly unsafe maneuver, but even then I think there'd be a sort of rebuttable presumption of fault.
Well, there’s actually a very small set of operations that allow all computation, so it doesn’t take much to be a DSL and a GP too; I’d be surprised if a proof language couldn’t swing it.
I would only wish that on my enemies... For god's sake, if you're going to think in a programming language, pick one with strict referential transparency.
That’s why I said it will be used more for bad than good. Maybe I should have said, “it will be used for far more bad than good for the purpose of human intellectual flourishing”.
It can be used for good for this purpose. You can use it to find primary source information. You can use it to generate challenges for you. You can use it to critique your work when there’s nobody else available. Etc etc.
I just noticed it more often that it's getting used as a positive connotation.
From the article itself:
"Over-engineered, absolutely. [...] Every check earned its place by catching something I’d otherwise have shipped. Would I do it again? Without hesitation."
I am a very big fan of Axler's approach, because to me it really gets to the (abstract) essence of things. When I truly started to "get" this approach, it really made linear algebra an important part of how I think.
reply