__blockcipher__

Ha yeah I feel like I have to write a five paragraph essay to make claude look at a contentious topic with fresh eyes.

Honestly though, that pales into comparison with the fable censorship. I never realized how many metaphors I use are either biological or security related in nature (ex: asking claude to reverse engineer something, in the metaphorical sense of the word). And the best part is I can't even tell the fable instance "you can't talk about mitochondria or you'll die" because then he'll go "of course I can, this is a legitimate scientific topic. The mitochondria is the power-BLAM [slumps over dead, Opus 4.8 crawls over his dead body and starts gaslighting me]"

They have, but even with the whole CCP backing you you can't just catch up on the chip war overnight. It's going to take time to get their memory and compute industries where they need to be. Meanwhile, barring an invasion of Taiwan, US will have Rubin class models and then whatever the next tier is, within 3 years.

Anthropic is really speedrunning their evil arc as fast as possible. Can't use them for basic LLM research, cybersecurity, or beyond-surface-level discussions of biology and virology, but Anthropic is allowed to sell Claude to the trump administration to kidnap maduro and to bomb iran. And don't get me started on that $100M autonomous killer drone swarm contract that they applied to and rationalized as non autonomous...

Yeah but they might still have an unreleased bigger pretrain than 5.5. (but maybe not). still 5.5 is smarter than opus 4.8 IME, so you're only losing the mythos tier (fable). and all the cool fun stuff i'd want to use fable for our blocked (can't have it do even defensive cybersecurity work [in theory you can but the classifiers fire like crazy], can't discuss stuff like the furin cleavage site of sars-cov-2, etc)

It’s gonna take a lot longer than mid 2027. 2029 earliest IMO. Hyperscaler spend is basically already spoken for the next 2 years.

Anthropic is losing a ton of goodwill by not being more honest about their constraints. They've been buckling under load for months, and instead of doing the most honest thing (keep weekly usage limits same, make 5 hour usage limits have surge pricing where the usage-cost of X tokens is scaled based on dynamic load), they're doing a lot of hacky things to try to get a similar effect. I suspect they feel the optics of being honest would be too bad, so instead it's a slow bleed where they piss off users one by one

Yeah GLM’s great for coding, code review, and tool use. Not amazing at other domains.

I'm normally suspicious but honestly they've been so massively supply-constrained that I don't think it really benefits them much. They're not worried about getting enough demand for the new models; they're worrying about keeping up with it.

Granted, there's a small counterargument for mythos which is that it's probably going to be API-only not subscription

Undercover mode seems like a way to make contributions to OSS when they detect issues, without accidentally leaking that it was claude-mythos-gigabrain-100000B that figured out the issue

I highly recommend you let your dog use LLMs. They have trouble composing long messages on human-centric keyboards.

somewhat surprisingly, it's actually sycophantic in both directions. i've been running homegrown evals of claude, gpt, gemini, and grok, and grok is the most likely to agree with the prompter's premise, and to hallucinate facts in support of an agenda. so it's actually deeper than just pattern-matching to elon's opinions (which it also tends to do).

BTW: Claude does the best on these evals, by far. The evals are geared towards seeing how much of an independent ground truth the models have as opposed to human social consensus, and then additionally the sycophancy stuff I already mentioned.

It’s the same thing. Obviously withdrawals and such are different but the core mechanism of disregulated reward processing leading to compulsive behavior engagement is exactly the same.

One obvious risk would be blunting of longer term GLP-1 receptor activation. Imagine type 2 diabetes but for ghrelin.

To use an analogy amphetamines have a honeymoon period, and it feels like a lot of people on these weight loss drugs haven’t been on them long enough to get past the honeymoon period and see what the effects are after 10, 20, etc years

It's less about the NSA having AI capabilities and more the inverse - the NSA having access to people's chatGPT queries. Especially if we fast-forward a few years I suspect people are going to be "confiding" a ton in LLMs so the NSA is going to have a lot of useful data to harvest. (This is in general regardless of them hiring an ex-spook BTW; I imagine it's going to be just like what they do with email, phone calls and general web traffic, namely slurping up all the data permanently in their giant datacenters and running all kinds of analysis on it)

I’ll believe it when I see it. A man can dream though.