{"id":51634,"topic":"ai","source":"Sohu","title":"80亿参数记性不如U盘：他算出了大模型记忆天花板 - Sohu","url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","url_hash":"869fcf655bc8910a5bcd47fa6824eac1a898b59b","author":"","summary":"<a href=\"https://news.google.com/rss/articles/CBMiiAFBVV95cUxObW9LQlBHTHdwNzBmYktfWHpfR3ByLUF5YWwycnhwb3FNcEh4YWthVDRxcEx0My1DeFVxX2lnZHdaMV9WRnh1d3FMRUgxb1NhUW9CWHJfMW1VUkZTX1lFS0p1OVg5MHRpWmdad0RSZ21RUm1IU2hlaEZ3QkJBVndlTFlLaFZMRHNq?oc=5\" target=\"_blank\">80亿参数记性不如U盘：他算出了大模型记忆天花板</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">Sohu</font>","content":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","image_url":"https://q7.itc.cn/q_70/images03/20260802/83a0f0a37af34af88df17e048b68d5d8.png","lang":"zh","published_at":"2026-08-02T08:41:00+00:00","fetched_at":"2026-08-02T12:15:08+00:00","status":"read","starred":0,"extract_state":"ok","summary_auto":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","cluster_id":null,"extract_retries":0,"extract_error":null,"contract_version":"news_item.v1","format_contract_version":"news_item_formats.v1","dedup_url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 362 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":362,"summary_length":362,"usable_text_length":362,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":362,"summary_length":362}},"news_item":{"id":51634,"canonical_url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","source_url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","title":"80亿参数记性不如U盘：他算出了大模型记忆天花板 - Sohu","source_name":"Sohu","author":null,"published_at":"2026-08-02T08:41:00+00:00","locale":"zh","topic":"ai","tags":[],"rss_summary":"<a href=\"https://news.google.com/rss/articles/CBMiiAFBVV95cUxObW9LQlBHTHdwNzBmYktfWHpfR3ByLUF5YWwycnhwb3FNcEh4YWthVDRxcEx0My1DeFVxX2lnZHdaMV9WRnh1d3FMRUgxb1NhUW9CWHJfMW1VUkZTX1lFS0p1OVg5MHRpWmdad0RSZ21RUm1IU2hlaEZ3QkJBVndlTFlLaFZMRHNq?oc=5\" target=\"_blank\">80亿参数记性不如U盘：他算出了大模型记忆天花板</a>&nbsp;&nbsp;<font color=\"#6f6f6f\">Sohu</font>","full_text":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","excerpt":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 362 characters.","diagnostics_url":"/api/diagnose?url=https%3A//m.sohu.com/a/1057873441_473283%3Fscm%3D10001.325_13-325_13.0.0-0-0-0-0.5_1334","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 362 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":362,"summary_length":362,"usable_text_length":362,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":362,"summary_length":362}}},"display_formats":["compact","card","full","digest_section","json"]},"daily_stack_record":{"title":"80亿参数记性不如U盘：他算出了大模型记忆天花板 - Sohu","url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","summary":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","source":"Sohu","date":"2026-08-02T08:41:00+00:00","content":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//m.sohu.com/a/1057873441_473283%3Fscm%3D10001.325_13-325_13.0.0-0-0-0-0.5_1334","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 362 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 362 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":362,"summary_length":362,"usable_text_length":362,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":362,"summary_length":362}},"tags":[]},"fallback_formats":["markdown","json","html"],"actions":{"read":"/item/51634","export_markdown":"/api/items/51634/export?format=markdown","export_json":"/api/items/51634/export?format=json","diagnose":"/api/diagnose?url=https%3A//m.sohu.com/a/1057873441_473283%3Fscm%3D10001.325_13-325_13.0.0-0-0-0-0.5_1334"},"formats":{"full":{"id":51634,"title":"80亿参数记性不如U盘：他算出了大模型记忆天花板 - Sohu","url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","source":"Sohu","author":null,"published_at":"2026-08-02T08:41:00+00:00","locale":"zh","topic":"ai","tags":[],"excerpt":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","full_text":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","reading_time_min":1,"extraction":{"state":"ok","confidence":0.9,"error":null,"explanation":"High confidence: full text extraction produced 362 characters.","diagnostics_url":"/api/diagnose?url=https%3A//m.sohu.com/a/1057873441_473283%3Fscm%3D10001.325_13-325_13.0.0-0-0-0-0.5_1334","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 362 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":362,"summary_length":362,"usable_text_length":362,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":362,"summary_length":362}}},"quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 362 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":362,"summary_length":362,"usable_text_length":362,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":362,"summary_length":362}},"actions":{"read":"/item/51634","export_markdown":"/api/items/51634/export?format=markdown","export_json":"/api/items/51634/export?format=json","diagnose":"/api/diagnose?url=https%3A//m.sohu.com/a/1057873441_473283%3Fscm%3D10001.325_13-325_13.0.0-0-0-0-0.5_1334"}},"digest":{"id":51634,"title":"80亿参数记性不如U盘：他算出了大模型记忆天花板 - Sohu","url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","source":"Sohu","topic":"ai","published_at":"2026-08-02T08:41:00+00:00","excerpt":"80亿参数记性不如U盘：他算出了大模型记忆天花板 新智元报道 【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。 大模型的记性，比你想象中可能差远了。 一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。 可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。 7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。…","quality_bucket":"high","quality_reason":"High confidence: full text extraction produced 362 characters.","reading_time_min":1,"cluster_id":null},"card":{"display_title":"80亿参数记性不如U盘：他算出了大模型记忆天花板 - Sohu","subtitle":"Sohu · 2026-08-02","summary":"80亿参数记性不如U盘：他算出了大模型记忆天花板 新智元报道 【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。 大模型的记性，比你想象中可能差远了。 一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。 可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8…","badges":["quality:high"],"links":{"read":"/item/51634","original":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","diagnose":"/api/diagnose?url=https%3A//m.sohu.com/a/1057873441_473283%3Fscm%3D10001.325_13-325_13.0.0-0-0-0-0.5_1334"},"quality_warning":null},"export":{"title":"80亿参数记性不如U盘：他算出了大模型记忆天花板 - Sohu","url":"https://m.sohu.com/a/1057873441_473283?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334","summary":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","source":"Sohu","date":"2026-08-02T08:41:00+00:00","content":"80亿参数记性不如U盘：他算出了大模型记忆天花板\n新智元报道\n【新智元导读】参数越多，大模型就能装下越多知识？一篇被NeurIPS拒过的论文把账算死了：每个参数只有3.6 bit。15亿参数模型的记忆，还不如一支U盘。\n大模型的记性，比你想象中可能差远了。\n一个80亿参数的大模型，一口气吞下15万亿token的训练数据，堆到硬盘上差不多7TB。\n可它真正「背」得下来的，少得可怜：每个参数只装得下3.6 bit。一个英文字母8 bit，连半个都填不满。\n7TB数据灌进去，整个模型记得住的大约只有两千分之一。你往里塞再多，它也只装下这么点。\n捅开这一真相的，是7月5日ICML 2026的一篇获奖论文，题目就叫《How much do language models memorize?》（《语言模型能记住多少东西？》）。","confidence":0.9,"diagnostics_url":"/api/diagnose?url=https%3A//m.sohu.com/a/1057873441_473283%3Fscm%3D10001.325_13-325_13.0.0-0-0-0-0.5_1334","quality_bucket":"high","failure_kind":"none","retryable":false,"quality_reason":"High confidence: full text extraction produced 362 characters.","quality_profile":{"profile_version":"extraction_quality.v2","bucket":"high","confidence":0.9,"failure_kind":"none","retryable":false,"retry_after_attempts":0,"reason":"High confidence: full text extraction produced 362 characters.","operator_guidance":{"severity":"ok","recommended_action":"trust_full_text","next_step":"Use the extracted full text as the primary article source.","operator_label":"Ready","can_retry":false,"can_use_summary":false,"diagnostics_required":false},"content_depth":{"contract_version":"content_depth.v1","category":"full_text","label":"Full text","has_full_text":true,"has_summary":true,"content_length":362,"summary_length":362,"usable_text_length":362,"source_field":"content"},"legacy_collapsed":false,"signals":{"extract_state":"ok","extract_error":null,"extract_retries":0,"content_length":362,"summary_length":362}},"tags":[],"format_contract_version":"news_item_formats.v1"}}}