Chinese AI Labs ARE COPYING Claude?! — Superbash Learn

Superbash Learn

Click to load the YouTube player

Key takeaways

Switch non-reasoning workloads to MiniMax 2.5 or Kimi if you can tolerate a ~30% failure rate on logic puzzles — at ~5% of Claude Opus cost, the savings outweigh the misses for bulk code-gen and drafting tasks.
Keep Claude Opus on the critical path for any prompt that requires multi-step logic — the car-wash test (~95% vs ~30%) is a decent proxy for which tasks to route where.
Avoid treating Anthropic's IP complaints as moral high ground in any public thread — the Harry Potter test (~95% of the books) is now a common counter-example and will end the argument against you.
Pin DeepSeek V4 to your evaluation queue the week it drops — the creator explicitly framed it as the test of whether Chinese labs close the intelligence gap, not just the price gap.
If you build developer tools, assume Chinese models will be plugged into your stack via raw API spec paste — design i18n and doc rendering to handle untranslated Chinese inputs on day one.
Watch the next 12–24 months for a developer-pool-driven performance catch-up: China's annual engineering output exceeds the US purely on population (~1.4B), so Anthropic's 'more developers on board' moat has a known expiry.
Stop paying full Opus rates for tasks that don't need it — route the cheap 95% to MiniMax/Kimi and reserve the expensive 5% for Opus, since that cost split is already the creator's own usage pattern.

Video summary

Companion notes

Anthropic says Chinese AI labs ran a 24,000-account distillation attack against Claude — and the creator thinks the accusation is fair.

The Accusation

Anthropic alleges that Chinese model makers (DeepSeek, MiniMax, Kimi) operated ~24,000 accounts to query Claude, harvest its outputs, and distill that knowledge into their own models. The attack is "extremely sophisticated" and "growing in intensity," and Anthropic is now publicly calling for "rapid coordinated action among industry players, policy makers and the broader AI community."

Creator's Read: Both Sides Steal

The host says he "completely agrees this is happening" — China is "very, very good at copying things" and will "do whatever it takes to win this AI race." But he also points out the hypocrisy: a tester pulled ~95% of the Harry Potter books out of Claude, suggesting Anthropic itself ingested copyrighted material during training. The community's response on X was blunt — "circle stealing, so what?" — with figures like Elon Musk and Ash Crypto calling out the double standard.

Performance Gap Is Real, Cost Gap Is Bigger

In the creator's internal tests, Claude Opus still beats MiniMax 2.5 and Kimi on reasoning. On the classic "walk vs. drive to the car wash" prompt, Opus got it right ~95% of the time while MiniMax 2.5 still fails ~30% of the time. But MiniMax runs at roughly 5% of Claude's cost — a gap the creator calls "a lot of savings" for builders. DeepSeek V4 is "coming soon" and he's watching to see if it closes the intelligence gap, not just the price gap.

Why The Race Tightens

Anthropic's edge on coding comes from "more developers on board" feeding it usage patterns. China produces more engineers per year than the US by sheer population (~1.4 billion), so the developer-pool asymmetry may close within "a few years." The creator's framing: this is "the first shot that's fired and this will not be the last."

What It Means For You

Pricing pressure benefits every developer regardless of which lab "wins." The creator's own bill on Opus is still high, but routing workloads to MiniMax or Kimi at 5% cost is already a viable split. Paste a Chinese API spec into any model and it works — "AI doesn't care what language you use" — so the AGI sprint is now genuinely global.

Watch on YouTube

Prefer the native player? Open it on YouTube: https://www.youtube.com/watch?v=8ihU_y1QET8