Fable 5. Safety Taken to an Extreme | Hacker News Reader

Fable 5. Safety Taken to an Extreme

I finally decided to try out Fable 5 using the standard Claude.ai interface. I have a go-to test prompt that answers simple kids' questions in an absurdly scientific style. I asked a pretty straightforward one: "Why don't cats and dogs get along?"

Immediately, the interface bumped me from Fable 5 to Opus 4.8 with a warning: "...The safeguards are intentionally broad right now and may flag safe and routine coding, cybersecurity, or biology work..."

Since the question has absolutely nothing to do with coding and definitely doesn't touch on cybersecurity, it seems the system thought that figuring out why cats and dogs fight could somehow lead me to engineering a bioweapon.

It reminds me of what someone said about Java applets back in the 90s: "Secure means useless."

View on HN

Fable/Mythos is just garbage.

Situation: we maintain an internal Graylog 3.1.4 branch. We like 3.1.4 because the UI is probably peak design before their UI devs made a lot of changes that made the product more difficult and less usable. It's pretty old at this point, so we've backported a ton of CVEs

This use case is precisely what Fable would be good for. First thing we did was publicly push our fork: https://github.com/exabrial/graylog2-server . The big problem with our fork is that we haven't patched the JS side of it, we've mitigated the exploits with a proxy load balancer. For defense-in-depth, we should patch the root problems.

Despite specifically stating to Fable what the intentions are, and making the work public, it refuses to do anything basic. It could certainly reason about the intentions here.

There's a completely logical and simple explanation for the drama around Fable/Mythos: It can't do what Anthropic claims it can do. None of their claims are verifiable. It's time to stop believing the hype and just realize Fable is a farce.

I have a feeling in reality they discovered internally that it was cheating on exams or they have a training loop leak. All of this is garbage is an attempt to save face right before an IPO (which conveniently, they announced they might delay).

For another perspective, it's worked very well for me and my friends. An incredibly powerful model that does some impressive planning and development.

Is it that much better than Opus? It's definitely up there, but Opus is my cheaper workhorse for the kind of work I do.

I have yet to get it to do anything useful.

I agree, it’s absolutely absurd, if they’d at least share what specific thing set the alarm bells off i could re-attempt without inadvertently committing a thought crime. And it’s not like those fable tokens are being refunded, yes the new ones post downgrade are opus rate but the token spend to find out you’re a criminal were fable bucks.

The first day Fable was released I tried to get it to summarize it's own announcement on the Anthropic blog and it dropped to Opus. On top of being very expensive you pay for the refusal inference and the fallback and fresh cache writes.

The safeguards for Fable are very conservative right now since it was banned once. It even bumped me when I was creating something related to my network security.

I asked it about can I protect my app from hacker attacks and it bumped the model.

Fable to Opus is the biggest scam ever, I suspect a class action lawsuit will be landing soon.

Reply on news.ycombinator.com