Muse Spark評測:Meta AI解開5道數學懸案 | Muse Spark Review: Meta AI Cracks 5 Open Math Problems
By Kit 小克 | AI Tool Observer | 2026-10-05
🇹🇼 Muse Spark評測:Meta AI解開5道數學懸案
Muse Spark 是 Meta 藏在 meta.ai 聊天介面「Thinking Mode」底下的推理模型。2026年10月2日,它跟一群數學家一起掛名發表了六篇論文,其中五篇解開了業界卡了好幾年、甚至十幾年的公開數學難題。這不是又一次「AI秒解數學」的行銷標題——Meta自己在官方部落格講得很保守:這是人類主導、AI輔助的合作模式,不是AI關起門來自己解題。老實說,這種誠實反而比誇大宣傳更值得寫。
Muse Spark 到底解了什麼?
六篇論文橫跨六個完全不同的數學領域,難度都是「開放問題」等級——換句話說,沒有標準答案可以對,錯了也沒人知道該怎麼改:
- 機率論:找出高維度空間中「用球體擬合橢球」的臨界門檻
- 微分方程:證明了一個2015年就懸而未解的雙調和非線性薛丁格方程解的有限時間崩潰問題
- 群論:用一個384元素的群,推翻了2024年提出的半阿貝爾群猜想
- 最佳化:回答了二元多項式最佳化鬆弛法何時「精確捕捉」原始問題
- 算術物理:串連了數論與p進位弦理論計算之間的關係
- 非結合代數:推翻了一個關於可解進化代數的猜想,並提出替代判定方式
人類數學家才是主角,不是 Muse Spark 自己
這是整件事最重要的細節。每篇論文都清楚標註哪些段落是人類寫的、哪些是AI草擬的。數學家負責選題、把關方向、驗證每一步推論是否正確;Muse Spark負責生成證明草稿、跑計算、寫輔助程式碼。換句話說,這是人機協作,不是「AI取代數學家」。Meta自己也坦承,其中幾個問題在同一時間被其他獨立團隊用不同方法解開了——這提醒我們不要把巧合當成突破。
老實講,這代表什麼?
今年早些時候,類似模型已經在五項數學奧林匹亞競賽拿到「金牌等級」成績,但競賽題目有標準答案、有已知解法套路可循。公開研究問題完全不同——沒有答案本可以抄。這六篇論文的意義,不在於「AI又變聰明了」的宣傳效果,而是證明了AI數學研究目前最實際的用法:當一個會算、會寫證明草稿、不會累的助理,縮短數學家測試想法的時間。如果你是研究者,Muse Spark現在能幫你跑計算、生成候選證明,但審稿、判斷、對錯把關還是得靠人。
好不好用,試了才知道。
🇺🇸 Muse Spark Review: Meta AI Cracks 5 Open Math Problems
Muse Spark is the reasoning model tucked inside meta.ai's chat interface under "Thinking Mode." On October 2, 2026, it got listed as a co-author on six math papers — five of which cracked open math problems that had sat unsolved for years, some over a decade. This isn't another "AI casually solves math" headline. Meta's own writeup is notably restrained: humans drove the research, AI assisted — not the other way around. That restraint is actually the more interesting part of the story.
What did Muse Spark actually solve?
The six papers span six unrelated fields of math, all at the "open problem" level — meaning there's no answer key to check against:
- Probability: found the threshold for fitting Gaussian points to ellipsoids in high dimensions
- Differential equations: proved finite-time blow-up for a biharmonic nonlinear Schrödinger equation question left open since 2015
- Group theory: disproved a 2024 conjecture about semiabelian groups using a 384-element counterexample
- Optimization: answered when a binary polynomial optimization relaxation exactly captures the original problem
- Arithmetic physics: linked number theory to p-adic string theory calculations
- Non-associative algebra: disproved a conjecture about solvable evolution algebras, offering an alternative characterization
Human mathematicians are the lead, not Muse Spark
This is the detail that matters most. Every paper marks which passages were drafted by researchers versus by AI. Mathematicians picked the problems, steered direction, and verified every step. Muse Spark generated proof drafts, ran calculations, and wrote supporting code — this is human-AI collaboration, not "AI replaces mathematicians." Meta itself admits some of these problems were independently solved by other teams around the same time using different methods — a reminder not to mistake coincidence for breakthrough.
Honest take: what does this actually mean?
Earlier this year, similar models hit "gold-medal" scores across five math Olympiads — but competition problems have known answer formats and solution patterns. Open research problems have none of that. The real significance of these six papers isn't "AI got smarter" marketing — it's proof of the most practical current use of AI for math research: a tireless assistant that can crunch calculations and draft candidate proofs, shrinking how long it takes a researcher to test an idea. If you're a working mathematician, Muse Spark can run computations and generate proof candidates today — but review, judgment, and correctness checks still need a human.
好不好用,試了才知道 — you won't know until you try it.
Sources / 資料來源
- Meta Research: Solving Open Research Problems Together
- The Gold Rush in AI4Math: Where Are We Now? (arXiv)
- Meta's Muse Spark Helped Mathematicians Solve Five Open Research Problems - AlphaSignal
延伸閱讀 / Related Articles
- Clef評測:Cloudflare開源決策模型,分類比LLM快10倍 | Clef Review: Cloudflare's Open-Source Decision Model
- Kolibri評測:德國主權AI模型,效能夠用嗎? | Kolibri Review: Germany's Sovereign Open AI Model
- FLUX 3 Image評測:4K修圖新選擇,值得換嗎? | FLUX 3 Image Review: Black Forest Labs' New Editor
AI 工具觀察站 — 每日精選 AI Agent 與工具趨勢
AI Tool Observer — Daily curated AI Agent & tool trends
留言
張貼留言