Let’s see how long HN takes to ban yet another literal Nazi.
Let’s see how long HN takes to ban yet another literal Nazi.
What really gets me is that we never look past Schrödinger’s version of the cat. I want us to talk about Bell’s Cat, which cannot be alive or dead due to a contradiction in elementary linear algebra and yet reliably is alive or dead once the box opens. (I guess technically it should be “alive”, “dead”, or “undead”, since we’re talking spin-1 particles.)
At risk of being NSFW, this is an amazing self-own, pardon the pun. Hypnosis via text only works on folks who are fairly suggestible and also very enthusiastic about being hypnotized, because the brain doesn’t “power down” as much machinery as with the more traditional lie-back-on-the-couch setup. The eyes have to stay open, the text-processing center is constantly engaged, and re-reading doesn’t deepen properly because the subject has to have the initiative to scroll or turn the page.
Adams had to have wanted to be hypnotized by a chatbot. And that’s okay! I won’t kinkshame. But this level of engagement has to be voluntary and desired by the subject, which is counter to Adams’ whole approach of hypnosis as mind control.
RAG is “Retrieval-Augmented Generation”. It’s a prompt-engineering technique where we run the prompt through a database query before giving it to the model as context. The results of the query are also included in the context.
In a certain simple and obvious sense, RAG has been part of search for a very long time, and the current innovation is merely using it alongside a hard prompt to a model.
My favorite example of RAG is Generative Agents. The idea is that the RAG query is sent to a database containing personalities, appointments, tasks, hopes, desires, etc. Concretely, here’s a synthetic trace of a RAG chat with Batman, who I like using as a test character because he is relatively two-dimensional. We ask a question, our RAG harness adds three relevant lines from a personality database, and the model generates a response.
> Batman, what's your favorite time of day?
Batman thinks to themself: I am vengeance. I am the night.
Batman thinks to themself: I strike from the shadows.
Batman thinks to themself: I don't play favorites. I don't have preferences.
Batman says: I like the night. The twilight. The shadows getting longer.
My NSFW reply, including my own experience, is here. However, for this crowd, what I would point out is that this was always part of the mathematics, just like confabulation, and the only surprise should be that the prompt doesn’t need to saturate the context in order to approach an invariant distribution. I only have two nickels so far, for this Markov property and for confabulation from PAC learning, but it’s completely expected weird that it’s happened twice.
Even better, we can say that it’s the actual hard prompt: this is real text written by real OpenAI employees. GPTs are well-known to easily quote verbatim from their context, and OpenAI trains theirs to do it by teaching them to break down word problems into pieces which are manipulated and regurgitated. This is clownshoes prompt engineering done by manager-first principles like “not knowing what we want” and “being able to quickly change the behavior of our products with millions of customers in unpredictable ways”.
That’s the standard response from last decade. However, we now have a theory of soft prompting: start with a textual prompt, embed it, and then optimize the embedding with a round of fine-tuning. It would be obvious if OpenAI were using this technique, because we would only recover similar texts instead of verbatim texts when leaking the prompt (unless at zero temperature, perhaps.) This is a good example of how OpenAI’s offerings are behind the state of the art.
Not with this framing. By adopting the first- and second-person pronouns immediately, the simulation is collapsed into a simple Turing-test scenario, and the computer’s only personality objective (in terms of what was optimized during RLHF) is to excel at that Turing test. The given personalities are all roles performed by a single underlying actor.
As the saying goes, the best evidence for the shape-rotator/wordcel dichotomy is that techbros are terrible at words.
The way to fix this is to embed the entire conversation into the simulation with third-person framing, as if it were a story, log, or transcript. This means that a personality would be simulated not by an actor in a Turing test, but directly by the token-predictor. In terms of narrative, it means strictly defining and enforcing a fourth wall. We can see elements of this in fine-tuning of many GPTs for RAG or conversation, but such fine-tuning only defines formatted acting rather than personality simulation.
I’m not sure if they’re a decent person or not, but I liked this sneer on the orange site, in the context of this recent rant (previously, on Awful), describing a hypothetical executive-humbling mechanism. I agree with them; I’d read that fanfiction too.
You got Schmidhuber’d! A Gödel machine would fit the bill. Nobody’s built one yet, but the hard part – proof search through something like Metamath (particularly Metamath Zero) – is long-since solved. It wouldn’t take over the world, though; it would just sit in a corner and get really good at maths over the next few centuries.
Why are techbros such shit at Lojban? It’s a recurring and silly pattern. Two minutes with a dictionary tells me that there is {seldikca}
for being charged like a capacitor and {nenzengau}
for charging like a rechargeable battery.
This article motivating and introducing the ThunderKittens language has, as its first illustration, a crying ancap wojak complaining about people using GPUs. I thought it was a bit silly; surely that’s not an actual ancap position?
On Lobsters, one of the resident cryptofascists from the Suckless project decided to be indistinguishable from the meme. He doesn’t seem to comprehend that I’m mocking his lack of gumption; unlike him, I actually got off my ass when I was younger and wrote a couple GPU drivers for ATI/AMD hardware. It’s difficult but rewarding, a concept foreign to fascists.
That thread from Paul is such an eyeroll. “Community”, says the VC-funded startup-builder. But it really does show how far out into the wilderness Dorsey has gone; not even his startup bros want to follow.
Microsoft is legendary for this. In fact, I’ll give you Microsoft’s entire business recipe; it’s not secret:
This is some of the most corporate-brained reasoning I’ve ever seen. To recap:
It’s a fucking pattern, ain’t it.
Standard HN practice. See here for a previous discussion about the phenomenon.
Exactly! Quoting myself from the conversation I linked above:
With [HN] in mind, I don’t really want to give Dan G. any benefit of the doubt. He’s had years to decide whether to open-source the forum, add features like a moderation log, switch the forum to invite-only mode, add reputation tracking, etc. It’s his website, and he chooses to leave these comments up.
At this point, reporting bad posts only serves to let Dan G. and his buddies sweep the fascists under the rug.
To be honest, I think it would be hilarious if Texas were annexed by Mexico, even just a little bit. Texans deserve a competent government who will invest in them and conduct reasonable diplomacy, and they’re not getting that from Texas.
This orange thread is about San Francisco banning certain types of landlord rent collusion. I cannot possibly sneer better than the following in-thread comment explaining why this is worthwhile: