Debates over AI benchmarks — and the way they’re reported by AI labs — are spilling out into public view.
This week, an OpenAI worker accused Elon Musk’s AI firm, xAI, of publishing deceptive benchmark outcomes for its newest AI mannequin, Grok 3. One of many co-founders of xAI, Igor Babushkin, insisted that the corporate was in the appropriate.
The reality lies someplace in between.
In a put up on xAI’s weblog, the corporate revealed a graph exhibiting Grok 3’s efficiency on AIME 2025, a group of difficult math questions from a current invitational arithmetic examination. Some consultants have questioned AIME’s validity as an AI benchmark. Nonetheless, AIME 2025 and older variations of the check are generally used to probe a mannequin’s math potential.
xAI’s graph confirmed two variants of Grok 3, Grok 3 Reasoning Beta and Grok 3 mini Reasoning, beating OpenAI’s best-performing out there mannequin, o3-mini-high, on AIME 2025. However OpenAI workers on X had been fast to level out that xAI’s graph didn’t embody o3-mini-high’s AIME 2025 rating at “cons@64.”
What’s cons@64, you may ask? Effectively, it’s brief for “consensus@64,” and it mainly provides a mannequin 64 tries to reply every downside in a benchmark and takes the solutions generated most steadily as the ultimate solutions. As you possibly can think about, cons@64 tends to spice up fashions’ benchmark scores fairly a bit, and omitting it from a graph may make it seem as if one mannequin surpasses one other when in actuality, that’s isn’t the case.
Grok 3 Reasoning Beta and Grok 3 mini Reasoning’s scores for AIME 2025 at “@1” — which means the primary rating the fashions bought on the benchmark — fall under o3-mini-high’s rating. Grok 3 Reasoning Beta additionally trails ever-so-slightly behind OpenAI’s o1 mannequin set to “medium” computing. But xAI is promoting Grok 3 because the “world’s smartest AI.”
Babushkin argued on X that OpenAI has revealed equally deceptive benchmark charts previously — albeit charts evaluating the efficiency of its personal fashions. A extra impartial occasion within the debate put collectively a extra “correct” graph exhibiting practically each mannequin’s efficiency at cons@64:
Hilarious how some folks see my plot as assault on OpenAI and others as assault on Grok whereas in actuality it’s DeepSeek propaganda
(I truly imagine Grok seems good there, and openAI’s TTC chicanery behind o3-mini-*excessive*-pass@”””1″”” deserves extra scrutiny.) https://t.co/dJqlJpcJh8 pic.twitter.com/3WH8FOUfic— Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) (@teortaxesTex) February 20, 2025
However as AI researcher Nathan Lambert identified in a put up, maybe a very powerful metric stays a thriller: the computational (and financial) price it took for every mannequin to attain its greatest rating. That simply goes to point out how little most AI benchmarks talk about fashions’ limitations — and their strengths.
