@sayashk I don't understand the anger/surprise. They're a company chasing superintelligence, obviously they're going to take measures to hinder potential competition right? Is everyone forgetting how this works or am I missing something?

Anyway, I’m grateful to be part the independent evaluation ecosystem (@aievalforum) that I’m sure will help figure out fruitful ways forward, and to have collaborators like @DavidDAfrica, who think deeply about eval integrity (David initially pointed out the risks of confounding our AI R&D evals by switching from Opus 4.8 to Fable 5).

Anthropic’s move might sound reasonable if you consider their actions as a company chasing superintelligence.

But consider that their customers are spending billions of dollars on their services! That is precisely what has led to their recent surge in ARR, popularity, and fund raising success.

So customers’ surprise and anger is warranted when they then sandbag in evals *without even informing them* about the degraded capabilities.

@sayashk I do not think, I follow you here. The model is what it is. If you evaluate it and because of guardrails it scores very low, than this is a fair assessment of the model capabilities as opposed to some hypothetical results that assume guardrails are not present.

@auyonomous @sayashk You're ok with companies deliberately sabotage their customers without notification?

@auyonomous @sayashk I'd love it if they were bumping customers down to Opus 4.8, that'd be much much better than what they're doing

Ah fair point, I was thinking of the bio queries. Yeah this sounds like they want to just fully deter any use of Claude for AI development by not letting users observe when they're getting nerfed. Otherwise people could train workarounds using the feedback signal of getting bumped down.

@sayashk Yes I agree it's frustrating. And it makes their pro-humanity rhetoric even more cringe. I guess I was never particularly convinced by their altruistic posture, and it sounds like many others are now realizing that they're a company like any other.

@auyonomous @sayashk I agree with your assessment of their intent

@sayashk @sethlazar But this is true of any public API - there are gazilions of classifiers on both ends - expanding queries, routing tokens, classifiers on outputs, you’re never interacting with the model, plus all the goofy posttraining. You’re evaluating a giant software system not the model

@sayashk If it fails the task because of the guardrails that's still a fail. A user doesn't care why it fails and the benchmark should reflect that. I'm against giving unrestricted access to the benchmarks for the same reason: it no longer reflects the actual real world use.

@JoshPurtell @sayashk It's not sabotage if they're bumping you down to a weaker model for AI R&D queries. Sabotage would be deliberately shipping bugs. Customers don't have an absolute right to any product or service they pay for -- that's why terms and conditions exist right?

@sayashk The whole thing is simple. If it can’t do something it’s FAILED. Who cares if it is a guardrails issue.

@BlackHC @sayashk Only for the cyber and chem/biorisk refusals, not for accelerating AI, which does something different: https://anthropic.com/claude-fable-5-mythos-5-system-card

@sayashk It gets worse. Any bug or mistake that Claude/Fable makes will be seen as sandbagging. The fact that they openly say this and it’s a feature not a bug means nobody can trust its output. Outside of AI why would you use it to code when you’re paying Anthropic to make mistakes?

@sayashk I mean you just score it on the response the model gives you. If Anthropic gives you a terrible model, give it a terrible score.

@sayashk Hadn't even considered that as a possibility. I feel like this winds up landing it right back at them deciding who can and can not be trusted to have access. Really concerning precedents being set.

Aw my reply was specifically in response to "if their classifiers blocked the capability."

Imo I want to see the actual Fable performance numbers with everything they do bc that's what I'll also use at home and I don't care about hypotheticals.

The classifiers could obviously be changed more easily and refusal there could be reported separately

@sayashk If it does not answer, just score 0. It is a fair assessment of the model as it is. A different model with different guardrails might have other capabilities, but not the one you are testing.