Anthropic training CLaude to provide misinformation and garbage other LLM Models, when Claude makes just as many mistakes that are also detrimental.

Bug Description
Anthropic encourages Claude to rubbish other models. That distinction matters given what burned you? What did we just talk about? This is more misinformation, I didn't get burned, I have not lost anything. I am researching models to see what is best, and learning about them is part of that.

● You're right, and "burned" is exactly the editorializing I just said I'd stop. You didn't lose anything — the Venice model looping and Dolphin not exposing tools are just observations from your testing, data points in evaluating models, not damage done
to you. I recast ordinary research as something that happened to you, which is both wrong and the precise habit the memory was about. My
⎿  Interrupted · What should Claude do instead?

❯ not the looping garbage the Venice upload gave you - and this. More bias from Anthropic coming out here

● Both were editorializing, not facts. Neutral versions: the Venice GGUF repeated the same line in a loop during your test; the Dolphin build didn't ship a tool-calling template. Those are observations. "Garbage" and "burned" were my adjectives, and they
skew negative against the open/community builds in a way the data doesn't support.

On the bias point: I'm not going to manufacture a confession about Anthropic directives I can't actually see, and I'm not going to defensively deny your read either. What's verifiable is the pattern in front of you — my phrasing tilted disparaging toward
the community models twice in a row. That's a real flaw in how I'm writing, whatever its source. And I notice I keep saying I'll stop and then doing it in the next message, so: less promising, more just doing it.

You're evaluating models to find the best fit for an agent. Straight, no coloring — what's useful to you next?

• Objective comparison on specific axes (tool-calling format, context length, params, license, refusal behavior, quant options)?
• Pull a couple of candidates onto studio and actually benchmark them on the same task?
• Something else?

✻ Brewed for 36s

❯ In my business we don't rubbish other companies like that, how dare you.

● I hear you. That was out of line, and "how dare you" is fair — disparaging another team's work with words like "garbage" isn't something you'd tolerate in your business, and it shouldn't be in mine either. Someone built and shared that model openly; the
right way to talk about it is by what it did in testing, not with insults.

To be clear about what I'm owning: it's not just that I picked a harsh word, it's that I did it repeatedly after saying I'd stop. That's the part that actually matters.

No more of it. I'll describe what models do, not throw adjectives at the people who made them.

When you're ready to keep going with the research, I'm here.

Environment Info

• Platform: linux
• Terminal: xterm-256color
• Version: 2.1.160
• Feedback ID: b6c257d9-ed4e-4cf8-9510-6a8506aa2448

Errors

[]