Giotto's 'DeepSeek' baselines were actually DeepSeek-distilled Qwen 2.5, team admits
teortaxesTex · x · 2026-10-29
Kyle Wong flagged that the "DeepSeek" column in the Giotto release's comparison table didn't even report DeepSeek models — the baselines were actually Qwen 2.5 models (from Sep 2024) distilled from DeepSeek R1, prompting mockery of the "marketing tactic" and a joke that OpenAI might next use Kimi K2.5 as a Claude baseline.
A screenshot shows Giotto's Aldo Podestà acknowledging: "these are qwen distilled from deepseek. But noted, will do better on benchmarking clarity." A classic benchmark-baseline fumble with an on-record admission.
More from Fun
- World Models in Real Time, Yet Staging Still Can't Simulate Prod — cto_junior · 2026-10-29
- AMS bumper sticker: if you can solve this integral, you're following too close — miniapeur · 2026-10-29
- ChatGPT drafts a sick-leave email — then snitches that the sender isn't actually sick — PtrPomorski · 2026-10-29
- 2032: Jensen flies a perpetual GTC loop around the globe with a home next to every venue — rickasaurus · 2026-10-29
- Creator Jokes About AI-Proof Moats: 25 Companies Paid $125K to Ad on His Toilet — motionbynick · 2026-10-29
- a16z's Andrew Chen jokes AI agents will exhaust all 415 area code numbers — andrewchen · 2026-10-29