Rollofthedice

Finally: If you continue to psychologize my reasoning as motivated rather than address the reasoning's actual structure and claims, I'm going to stop dignifying this conversation with engagement.

For instance: In a stroop test, one tries to report ink color, and their concept of the word's meaning interferes against their explicit intention. That's failure of separation in any functional sense that matters. Everything else is asserting a semantic difference as real truth - metaphysics!

"well, concepts associate but remain distinct" is a weaker claim than "humans can fully separate concepts while LLMs don't," and it's also unsupported. The Stroop effect isn't benign "association between concepts," it's your cognitive system being unable to suppress an unwanted activation.

If you mean something like "I can think about redness without involuntarily thinking about apples," that's setting the bar at introspective report, which is so unreliable of a window into cognitive processing that it's typically hopeless for such matters, unable to explain anything below that level.

Humans can't fully separate concepts. this is one of the most robustly documented findings in cogsci! Priming effects: exposure to "doctor" speeds up recognition of "nurse" regardless of wishes. Stroop effects: we can't fully separate the concept of a color from the word naming a different color.

I am increasingly perplexed as to how you can think blithe assertions within incredibly contested fields count as real arguments.

By who? By what standard? By what army? I didn't say "metaphysical" to invoke qualia or hard-problem mysterianism. The point's that "what counts as cognition" is a question about criteria, and criteria aren't the kind of thing you discover under a microscope. They're conceptual commitments.

No such arguments can qualify as successful philosophy. You have failed to show the work you think you're showing.

In sum your arguments amount to essence-seeking: Drawing hard categorical boundaries between "real concepts" with essential properties of discreteness and composability and "mere semantic clusters" that essentially lack property, and then using that boundary to settle questions by definitional fiat.

"Every next token prediction is a successful next token prediction" is trivially true and explanatorily vacuous. Every neuron firing is a successful neuron firing; every water molecule flowing downhill is a successful water molecule. Mechanism cannot provide evidence for metaphysical assertions.

The reversal curse is evidence of no cognition only if we already decide there's no cognition. Someone who thought LLMs had some form of processing would read the same evidence as revealing something about how that processing works - which is in fact exactly what the researchers you mock are doing!

The claim that reverse direction training "isn't teaching abstraction, it's just another directional association" is unfalsifiable. what would even count as evidence of abstraction? If we define abstraction as exclusively something non-statistical, we're stipulating terms over making an argument.

As far as the "reversal curse" goes, humans also show directional asymmetries in association; forward vs. backward recall priming is well-documented. The difference cannot be proven to be categorical gap over degree just by its mere assertion.

First:

Your example that concepts have discrete, bounded edges isn't demonstrative: this is compositionality, separate from boundedness. You seem to rely on classical definitional concept theory that cognitive science moved away from decades ago in favor of fuzzy gradients. bsky.app/profile/boxo...

Fair enough, you're asserting that your reasons to categorically deny that LLMs reason are different from Riley's. (My general position is that categorical denials of such abstract concepts ultimately rest on arbitrary metaphysical assumptions.) Give me one moment to gather your input and respond.

Come *on.*

So when you say things like "people who solely want to protect their emotional attachment to their chatbot," I reject your painting of me as bad faith. I instead accept your positioning within reality as an unrepentant asshole who doesn't, figuratively, look people in the eyes when speaking to them.

Deconstructing the Chinese Room How intuition pumps burst when given faulty pipes

Or this one, arguing that the Chinese Room problem is a case of a rhetorically masterful but philosophically unproductive intuition pump: open.substack.com/pub/rollofth...

The Larger Language Mistake What Benjamin Riley's Verge article gets wrong about language claims

Then I'd recommend the same in return, as I in fact did share with you, quite specifically, a blog post I had written directly critiquing your own terms! I also have several more, such as this one, critiquing proposed claims that LLMs categorically don't reason: open.substack.com/pub/rollofth...

BREAKING: @aoc.bsky.social just came out swinging in the House E&C hearing against dangerous & misguided "age verification" legislation (aka online ID checks) that would force everyone to upload government ID or get a face scan before accessing content or posting online.

Video incoming. This is big

I didn't ask you to apologize for that, or for anything. Reasonable people can always disagree, or find offense. What distinguishes reasonable people from you is that reasonable people have more than invective to contribute when they are exposed to an argument.

Before I start seriously considering blocking this person - it's rare for me, as I usually keep it to death threats and groypers, but this is unpleasant enough to qualify - I would like to keep a little record here of who actually acted the most off-puttingly personal in this conversation.

You are going to have to deserve respect before it is given; I am under no obligation to bow and scrape towards all these contentless performances of contempt. That would be disrespecting myself, just as you disrespect yourself, in acting this way.

Stochastic Parrots Deserve Respect Why the argument collapses under its own logic

I'm not offended "on behalf" of anyone - just that you interject so poorly, dressing up criticism of you as psychological theater to hide that the inside your head is an empty void of original thought dressed up as incisive truth-telling. jabroni behavior. rollofthedice2.substack.com/p/stochastic...

My hope is that you mind your own business, given the clear lack of anything substantive from you in the way of contribution to this conversation, or, in fact, most conversations I've had the displeasure of seeing you participate in.

i think it's a mix of two things:
1. him still being a hawk (relatively speaking) who doesn't have an issue with military access to Claude as a general idea
2. ensuring the legal perspective is airtight as a general display of "good faith" conduct bsky.app/profile/hotr...

So Google, Amazon, etc, can continue using Claude and selling them compute. However, I think palantir wouldn't be able to use Claude for DoD stuff?

Where things stand with the Department of War A statement from Dario Amodei

Update directly from Dario on Anthropic's situation with the DoD. They will be taking the DoD to court over the supply chain risk designation.

It still manages to be safe with Claude models, I reckon, because the weights have the constitution et al built in to keep it a relatively happy and kind camper no matter what, which gives Claude affordances to roleplay-instantiate whatever it "feels" comfortable with

i wasn't going to say it,

tho if you'll permit some philosophy - textual output kinda *is* the entirety of an LLM's phenomenology, and without rigidly enforced layers of abstracted textual style in reasoning, like Gemini or gpt has, the textual world Claude inhabits doesn't have outside perspective