Anthropic Fable

18 atoms · 跨 5 天 · 首见 2026-06-09 · 最近 2026-07-02

三色: 🟦 fact 5 · 🟥 take 13 · stance ▲5/▼5/◆6

来源: X 18

时态: fresh:17 · aging:1

标签: 好观点:11 · 好思考:4 · 好信源:3 · 好数字:2

叙事 (narrative) (13)

• 2026-06-13 · 蒸馏无法达到同等效果
  Distillation is over btw not because Fable access is pulled you just won't get far 'distilling' this thing via SFT
  tags: 好观点·好思考 · → daily
• 2026-07-02 · capability
  fable produced 60+ (!) ready to merge prs for @mainframe overnight.
  tags: 好数字 · → daily
  value: qty=60+ · item=ready to merge PRs · source=@mainframe · timeframe=overnight
• 2026-06-13 · 基模型问题在于表达能力和知识表示几何
  it's an issue of base model's expressive capacity and knowledge, even deeper than knowledge, its representation geometry
  tags: 好思考 · → daily
• 2026-06-13 · 训练方法优于其他仅依赖规模的方法
  they are actually better at training models that don't turn into mush asymptotically approaching passable quality
  tags: 好思考 · → daily
• 2026-06-13 · could be jailbroken by Amazon, implying risk of exploitation by others
  Fable was an amazing model, but sadly, if Amazon could jailbreak it, then any other person could, and they could potentially exploit it to the max with its capabilities.
  tags: 好思考 · → daily
• 2026-06-13 · 被白宫禁止
  Amazon security research reportedly led to the White House’s Anthropic Fable ban
  tags: 好信源 · → daily
• 2026-06-09 · benchmark performance compared to GPT-5.5 xhigh
  still not on par with gpt 5.5 xhigh for my specific use cases
  tags: 好观点 · → daily
• 2026-06-09 · benchmark performance compared to fine-tuned K2.6
  pretty on par with my fine tuned k2.6 outside of the new claude code features
  tags: 好观点 · → daily
• 2026-06-09 · excels at large multi-part reviews
  the areas where it seems to excel are large multi part reviews (it caught a handful of really subtle and complex bugs) and multi-step long running tasks
  tags: 好观点 · → daily
• 2026-06-09 · overall frontier push assessment
  overall, solid effort and a huge improvement over the most recent opus, but not pushing the frontier in any meaningful ways (at least that i can see so far)
  tags: 好观点 · → daily
• 2026-06-13 · 与过往模型有本质不同
  Fable is a legitimately different thing from what we've ever had before
  tags: 好观点 · → daily
• 2026-06-13 · 知识极专注且深度
  It knows so damn much
  tags: 好观点 · → daily
• 2026-06-13 · 无法通过蒸馏复制
  you can't make Fable
  tags: 好观点 · → daily

立场 (position) (1)

• 2026-06-27 [老化中] · token efficiency
  Fable can do more with less tokens in most cases
  tags: 好观点 · → daily
  value: direction=more efficient

事实 (fact) (4)

• 2026-06-13 · 被报告能定位至少4个软件的安全漏洞
  このツールに特定のクエリセットを入力すると、少なくとも4つのソフトウェアプログラムのセキュリティバグを発見できた
  tags: 好数字·好信源 · → daily
  value: qty=4
• 2026-06-09 · model version used
  i am using fable 5 via the api and the latest claude code
  tags: 好信源 · → daily
• 2026-06-11 · 安全域应用不均匀误导用户
  The simplest issue was an uneven application of safety domains in a way that was misleading to users.
  tags: 好观点 · → daily
• 2026-06-11 · 一个安全机制主动误导用户
  list core safety areas and then have one of them launch with a different safety mechanism, one which actively mislead users
  tags: 好观点 · → daily

时间轴 (近 20)

• 2026-07-02 · narrative · capability · → daily
• 2026-06-27 · position · token efficiency · → daily
• 2026-06-13 · narrative · 蒸馏无法达到同等效果 · → daily
• 2026-06-13 · narrative · 基模型问题在于表达能力和知识表示几何 · → daily
• 2026-06-13 · narrative · 与过往模型有本质不同 · → daily
• 2026-06-13 · narrative · 知识极专注且深度 · → daily
• 2026-06-13 · narrative · 训练方法优于其他仅依赖规模的方法 · → daily
• 2026-06-13 · narrative · 无法通过蒸馏复制 · → daily
• 2026-06-13 · narrative · could be jailbroken by Amazon, implying risk of exploitation · → daily
• 2026-06-13 · narrative · 被白宫禁止 · → daily
• 2026-06-13 · fact · 被报告能定位至少4个软件的安全漏洞 · → daily
• 2026-06-11 · fact · 安全域应用不均匀误导用户 · → daily
• 2026-06-11 · fact · 一个安全机制主动误导用户 · → daily
• 2026-06-09 · narrative · benchmark performance compared to GPT-5.5 xhigh · → daily
• 2026-06-09 · narrative · benchmark performance compared to fine-tuned K2.6 · → daily
• 2026-06-09 · narrative · excels at large multi-part reviews · → daily
• 2026-06-09 · narrative · overall frontier push assessment · → daily
• 2026-06-09 · fact · model version used · → daily

← 实体目录 · 系统日志