Hacker News Reader

Comment by consumer451 | Hacker News Reader

Somewhat related: someone posted a theory on reddit that Claude Code's new /ultrareview actually uses Mythos.

Does that seem plausible to anyone else? It runs on their cloud. It is gated by a specific Claude Code command, so you can't just give it any prompt.

View on HN

Something in favor of this is the fact that it runs in their cloud and literally tells you that it costs I think $10 to $25 per run

Why would they use their most expensive model when sonnet or opus can do the job as well?

In my experience sonnet<opus by a long shot for code review. Sonnet often flags things as errors that are not, because it fails to grasp the big picture… and also fails to grasp structural issues that are perfectly coded and only show up as problems at the meta scale.

I have no reason to believe that the next generation won’t offer similar gains in verification, and there is some evidence to support that the cybersecurity implications are the result of exactly this expansion of ability.

It depends on how you review. In an orchestrated per-task review workflow with clearly defined acceptance criteria and implementation requirements, using anything other than Sonnet (handed those criteria and requirements) hasn’t really led to much improvement, but it drives up usage and takes longer. I even tried Haiku, but, yeah, Haiku is just not viable for review, even tightly scoped, lol.

Siccing Sonnet on a codebase or PR without guidance does indeed lead to worse results than using Opus, though.

That makes sense, if your scope is tight enough, good enough is good enough. I’ve got the expected specifications and code style guides, including some aerospace engineering ones, but in complex systems I still run into difficult to sus out corner cases where the code works but the system breaks, usually due to unresolved conflicts in operational requirements.

There’s definitely a ceiling for what LLMs are capable of, and I think aerospace engineering might just currently be it, haha.

Lol yeah, I don’t think I’m ready to ride in the jet that Claude built lol. I should clarify that I use the code guidelines because they are solid guardrails for making things that perform predictably, not because I’m building MCAS lol. Let’s hope that “vibe aerospace engineering” is a way off for now.

It would be pretty simple to see what API they're calling.

That's what I meant to get at by "it runs on their cloud."

They can name that user-facing ultrareview API endpoint whatever they want, and we have no way to see what model endpoint it calls internally once running on their cloud, right?

Introduce intentional and increasingly subtle vulns and test against Sonnet, Opus, etc? Should give statistical evidence of its power.

Reply on news.ycombinator.com