Challenged RLC-AI performance claims before NVidia/Humain use

Situation

Peter personally interrogated the RLC-AI 9-10% performance advantage claim by going directly to Damen Knight (engineer who ran benchmarks) and Max Spevack. Discovered gains largely disappear when benchmarking code is properly optimized (uses torch.compile, etc.). Then asked Damen to evaluate whether Brian's marketing write-up is accurate or misleading: 'makes it sound awesome instead of pointless for data center deployments.'

Reasoning

RLC Pro AI GA targets Feb 25 and the board meeting is today -- performance claims will be part of the narrative to both audiences. Google explicitly wants RLC-AI benchmarks, so if claims don't hold up CIQ's credibility with a major customer is at stake. Peter values accuracy over hype -- the 9-10% number only applies to unoptimized code and disappears on real workloads. By going straight to the engineer who did the work, he bypassed telephone-game distortion. The framing ('makes it sound awesome instead of pointless') shows he's trying to make messaging defensible, not kill it.

Additional Context

Damen's benchmarks ran on A30 GPUs in his home lab. Gains come from unoptimized code paths; once torch.compile and vllm cuda graphs are used, RLC-AI is roughly equivalent to Ubuntu. Hardware matters significantly -- results vary across GPU types. Google roadmap meeting happened day before, where Google requested RLC-AI benchmarks. Board meeting happening today (Feb 24).

Observed Evidence

Peter initiated the investigation: 'I want to get a better understanding of the performance testing.' Damen confirmed: 'Later, when we optimised our benchmarking code, most of those gains disappeared.' Peter acknowledged: 'Yup. And lots of places where Ubuntu outperforms us.' Then asked Damen to audit Brian's write-up. Separately told Bjorn that Google wants RLC-AI benchmarks.

Confidence Breakdown

Reasoning Depth Analysis