Whether it happens to 5% or 50% (or even 0.5%) of users, I think everyone should feel extremely cynical when a company offers a huge freebie of a product ... and fails to mention that any % of their customers will actually get yesterday's model.
They did talk about it at length in the post, and most of the guardrails were already there in the first release. To be clear, it shouldn't have anything to do with the user. It's the prompt, and if it switches, it'll tell ya.
I mean for many of us this doesn’t even increase the cynicism. Ie it can’t get worse. These AI labs are already believed to be shifting behavior of these models at runtime so you never really know what you’re getting. Stability is not a thing with any provider, if you need that you have to run it locally.
I’m not offended by that because I’m already maximally offended lol.
They were calling for bans on open weight models. Bans on their competitors. Bans on anyone not as "enlightened" as them.
It is absolutely hilarious that they were the first to get regulated, and that it got to the point they had to turn off Fable as though it had been banned even for american citizens.
So this hinges on a reading of SB 1047 that interpreted the full shutdown requirement as impossible for an open-weight LLM. But it looks like that was already addressed. Here's an analysis:
>Clarifying the scope of a “full shutdown.” SB 1047’s “full shutdown” requirement has been a source of constant consternation for the open-source community. CalChamber explains:
>Under SB 1047, developers must build “full shutdown” capabilities into their models and may be held liable for downstream uses over which they have no control, impeding their ability to open-source their models. Ultimately, liability should rest with the user who intended to do harm, as opposed to automatically defaulting to the developer who could not foresee, let alone block, any and all conceivable uses of a model that might do harm. While recent amendments seemingly seek to narrow what is meant by “full shutdown” capabilities, the exclusions are unnecessarily difficult to interpret as drafted (full shutdown “does not mean the cessation of operation of a covered model to which access was granted pursuant to a license that was not created by the licensor…”) and altogether insufficient.
>Committee amendments simplify and clarify the definition of “full shutdown” such that the shutdown capability can be implemented into hardware used to train or run a model, rather than the model itself. The amendments also serve to exclude covered model derivatives that are outside of the developer’s control.
> may be held liable for downstream uses over which they have no control
Equivalent to a ban. Nobody is going to host or invest in this stuff if they suddenly become liable for everything it does. This is equivalent to repealing the safe harbor provisions in the DMCA.
>Committee amendments simplify and clarify the definition of “full shutdown” such that the shutdown capability can be implemented into hardware used to train or run a model, rather than the model itself. The amendments also serve to exclude covered model derivatives that are outside of the developer’s control.
I get the impression you are conflating whether a developer can be sued to oblivion for not implementing a "full shutdown" process that applies to finetunes versus whether they can be sued to oblivion for releasing a model that may cause "critical harm" when finetuned.
I'm confused why you think the only legal requirement is a "full shutdown" process. The text is there and I see a heck of a lot of requirements that are not about full shutdowns.
I get the impression that the full shutdown requirement is the main concern for open-weight from:
>SB 1047’s “full shutdown” requirement has been a source of constant consternation for the open-source community.
And I get the impression it's been addressed from the quote you're responding to. Neither mentions fine-tuning, which is defined elsewhere in the document. I'm not a lawyer, though, just relying on the analysis.
You claim you "get the impression" then do not quote either the law or a third party analysis of the law. Apparently we are supposed to believe this does not ban open source because the committee didn't write "This bans open source" in the beginning of the committee notes then circle it 3 times in red pen.
No, we're supposed to believe the official California Assembly committee bill analysis of SB 1047 over hacker news comments. Notably, SB 1047 passed the Legislature but was vetoed by Governor Newsom in September 2024. So whether or not it would have restricted open-weight models is an open question that won't get an answer.
No, you are not supposed to treat a politician or their staff’s statement about how their proposed bill works as dispositive and ignore the statutory text or third party analysis.
I’m glad we clarified the epistemological issue, so thank you for replying.
It is strange that half your reply is appeal to the authority of a not on point source and half is epistemological learned helplessness about what the impact of a vetoed bill would have been, pick a side.
When you make specific claims about a statute you were apparently too lazy to read, then respond with basically “I read the committee notes and surely if the statute was bad, it would say so and/or nobody can ever know what the statute does” when someone discusses the statute, whatever you are doing isn't “epistemological humility.”
Yeah, I was using 4.6 way more than 4.7. Pulling 4.6 from the web chat also means we lose access to Extended Thinking there. So they're saving on compute. It's hard not to assume this was part of the motivation behind the 4.8 release timing.
On web and mobile I can still select Opus 4.6, after a chat using 4.8, listed under other models. Extended thinking is a toggle in the effort menu
When I select 4.7 or 4.8 Extended thinking is replaced by adaptive thinking, but maybe I've understood the comment wrong and you meant 'when they pull 4.6 from web chat'?
They just showed the benchmarks it improved on but it regressed on so much more, such as the MCRR benchmark: "On multi-round coreference/context recall tests (often cited as MRCR or long-text retrieval benchmarks), Opus 4.7 reportedly dropped from roughly 78.3% down to 32.2% compared to Opus 4.6."
It seems like a lot of things fed into that. Anthropic couldn't keep up with the compute costs when they got a huge influx of users. (So) effort level defaults got turned down. (Looks like we have direct effort control in the web interface now - thrilled about that!) Adaptive Thinking, while usually cheaper for them, seems less robust than Extended Thinking. And this part is just vibes, but the alignment on 4.7 feels too stiff. I understand wanting the model to push back more, but it seems like 4.7 will push back reflexively in situations where it's just odd.
Too much personality, if you ask me. My biggest use case of an LLM is tool, not therapy, but therapy and opinions have been sneaking into workhorse tasks.
haven't verified, but attributed to Askell:
"I just think that... there's this idea that you're always giving the models a personality and a persona, because they are talking like people and they are trained on human data. And I think my worry has been: if you train them to be excessively corrigible and to see that as their persona, in people I think this actually has a lot of negative broader traits. As in, if you met someone and it was just like, "oh yeah, they would literally do anything," a follower — you know, if a person just tells them something and they just fully defer, they don't bother thinking about it at all — I'm just a bit worried about how that might end up generalizing, especially if models are going to be playing a more active role in the world."
Anthropic’s research makes the case that role-playing is inherent to how the models work. Communication implies a sender. Language implies a writer, and the models learn these roles implicitly during training. RLHF is meant to strengthen the attractor to the Assistant persona.
The RLHF very much does do that. My take is that RLHF as a mechanism ought to be avoided altogether, and even the selection of the assistant attractor basin is suspect. If I am exploring a problem space I don't want to hire Igor to explore it with me, it's more helpful to have a colleague role who will sort of jump out and say "nah thats dumb what if we throw out that whole thing and do this completely different angle instead".
Given the incentives that bring out the personalities in various occupations, I would guess other personas would be better suited to getting a task done than 'therapist' or 'tech HR rep'.
For examples, that of an explosive ordnance disposal technician, a surgeon, or a salvage saturation diver.
4.7 is a different base model from 4.6, so it's possible that they introduced regressions with pre-training changes, or undercooked the post-training stage.
Just speculating but I "feel" 4.7 was post-trained using more synthetic techniques. The way it writes for one thing, it's "personality", is less human and more fatiguing-AI-slop like.
You don't need to fry with RLAF to get that "slop feel". The first iterations of "AI slop" were raw SFT+RLHF - all human input, all inhuman output.
That said, I completely agree that 4.7 was a pronounced "model personality" regression. Closer to ChatGPT, and I mean that as an insult. Yet to check whether 4.8 is better.
Yep, until 1st June 4.6 is still x1 on Copilot, but will jump up quite a bit in coat - 4.7 was already highly priced, and the output was frankly terrible.
It still seems trying to build general models is mostly cost prohibitive - the frontier model provider and resellers are repricing in such a way the return on investment is dropping as developers and users become more cautious of burning their limits.
I'm still of the opinion that models like 4.6 don't need to be improved on - rather they need to be better integrated with more domain specific models in agentic flows.
Yeah, this is pretty clearly what's going on, but I wish they'd be more transparent about it. Funneling compute to Mythos and Design, while auto-setting effort levels lower and removing user control of extended thinking. I don't think the need to shuffle compute around is unique to Anthropic, though. I suspect it's part of why Sora got killed. And everyone's having uptime issues. Are we reaching the limits of the available compute?
