СледваСледващо ръководство
Обединяване на модели
технически
Фирми РЪКОВОДСТВО
Sakana AI is a Tokyo-based AI company that explores nature-inspired approaches, including evolutionary search and collective intelligence.
Чуйте това ръководствоAI-разказан · ~2 min
Its Evolutionary Model Merge research uses automated search to find recipes for combining existing pretrained models; this is model composition, not a general guarantee that merged models outperform their sources.
Sakana AI was founded by Llion Jones, a co-author of the Transformer paper Attention Is All You Need, and David Ha. Sakana means fish in Japanese, reflecting the company’s interest in nature-inspired collective intelligence. Its Evolutionary Model Merge work explores automated combinations of pretrained foundation models. The method represents candidate merge recipes, evaluates resulting models on chosen tasks, and uses evolutionary optimization to refine candidates. Sakana’s publication describes parameter-space merging and layer-wise model composition experiments; findings are tied to the tested models and benchmarks. The company also develops other research systems, including The AI Scientist, a multi-agent system for automating parts of scientific research. These efforts are distinct projects. Reusing pretrained weights can avoid a full from-scratch training run, but it still requires compatible inputs, compute for search and evaluation, and careful testing for quality and inherited limitations. Sakana’s separate AI Scientist system is described in its own paper as generating research ideas, code, experiments, visualizations, and papers with a simulated review process; this is distinct from evolutionary model merging.
Paper result: a Japanese-capable model is produced through a reported merge recipe; quality is benchmark-specific.
Paper result: the study evaluates merge recipes on named math tasks and benchmark sets.
Paper result: the method searches parameter and data-flow/layer-arrangement choices.
Illustrative: compare search-and-evaluation resources against a from-scratch baseline for a specific project; no universal cost saving is implied.
Automated model composition may help explore combinations of pretrained models. Its value depends on compatibility, search and evaluation cost, and whether measured results transfer beyond the benchmarks used to select a recipe.
Безплатен бюлетин
Три проверени AI истории всяка делнична сутрин, написани на достъпен език. Безплатно завинаги, без реклами.
Един имейл всеки делничен ден. Отписване с едно кликване. Никога не продаваме и не споделяме вашия имейл адрес.
Тествайте себе си
Незабавна обратна връзка за всеки отговор и сертификат за споделяне с подлежащ на проверка идентификационен номер, след като преминете курс.
Подкрепете безплатното AI образование. AI Understanding е 501(c)(3) с нестопанска цел — без реклами, без платен достъп, никога. Направете дарение
Sakana AI is a Tokyo-based AI company that explores nature-inspired approaches, including evolutionary search and collective intelligence. Its Evolutionary Model Merge research uses automated search to find recipes for combining existing pretrained models; this is model composition, not a general guarantee that merged models outperform their sources.
Той развива рецепти за комбиниране на тежести и слоеве на множество предварително обучени отворени модели, вместо да тренирате наново.
Лион Джоунс, съавтор на „Вниманието е всичко, от което се нуждаете“, съосновава Sakana с Дейвид Ха.
Сакана означава риба; компанията предпочита много малки колективни агенти пред един монолитен модел.
Той търси както да претегля/интерполира параметри, така и кои слоеве да подрежда и в какъв ред.
Reusing pretrained weights can avoid a full from-scratch base-model training run, but the merge search and evaluation still require resources and the total cost depends on the task.
Продължавай да учиш
Още ръководства, избрани за тази тема
СледваСледващо ръководство
Обединяване на модели
технически