以下部分引用 AI 总结，现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
Mythos Preview
20 atoms · 跨 8 天 · 首见 2026-04-30 · 最近 2026-06-26
三色: 🟦 fact 13 · 🟥 take 7 · stance ▲12/▼0/◆2
来源: X 20
时态: fresh:19 · aging:1
标签: 好数字:10 · 好观点:7 · 好思考:5 · 好信源:2
叙事 (narrative) (1)
- 2026-05-05
[老化中] · step-11 gate
The change after step 11 could be a compounding effect. It could also be that step 11 acts as a gate: once a model gets past it, later progress becomes easier or at least more reachable within the benchmark setup.
tags: 好思考 · → daily
value: direction=pass
预测 (forecast) (1)
- 2026-06-11 · ahead of trend in exploiting vulnerabilities
Mythos Preview seems clearly ahead of trend in exploiting vulnerabilities
tags: 好观点·好思考 · → daily
事实 (fact) (18)
- 2026-04-30 ⭐ · sycophancy rate reduction vs Opus 4.7 on relationship guidance
Mythos Preview cut that in half again.
tags: 好数字·好思考 · → daily
- 2026-05-02 ⭐ · 性能表现达到10M tokens
GPT-5.5 and Mythos Preview are only marginally more intelligent than previous models and have pretty much the same performance up to 10M tokens
tags: 好观点·好思考 · → daily
value: qty=10M · date=2026-05-02
- 2026-05-02 ⭐ · 超过10M tokens后性能激增
after that they go absolutely ballistic
tags: 好观点·好思考 · → daily
value: qty=10M · date=2026-05-02
- 2026-06-26 · ExploitBench Cap percent
Mythos Preview Cap percent: 约75%
tags: 好数字 · → daily
value: qty=75%
- 2026-06-09 · 训练加速比
Mythos Preview重测加速比: ~61x
tags: 好数字 · → daily
value: qty=~61x · date=2026-06-09
- 2026-06-04 · 模型加速训练代码性能
Claude Opus 4 averaged a ~3x speedup. This April, Mythos Preview achieved ~52x.
tags: 好数字 · → daily
value: qty=~52x speedup · date=April 2026
- 2026-06-04 · AI研究决策能力与人类对比
Mythos Preview improved on humans 64% of the time—up from 22% in 2024.
tags: 好数字 · → daily
value: qty=64% of the time · date=April 2026
- 2026-05-13 · UK AISI 网络安全挑战解决状态
The UK AISI found Mythos Preview is the first model to solve both their cyber ranges end-to-end. No model had ever solved the AISI's "Cooling Tower" cyber range before.
tags: 好信源 · → daily
value: direction=成为首个
- 2026-05-05 · progress rate after step 11
Weaker models tend to flatten out, while Mythos Preview and GPT-5.5 continue progressing at a faster pace, causing the gap at later milestones to widen.
tags: 好数字 · → daily
- 2026-04-30 · 平均通过率
Mythos Preview 68.6% (±8.7%)
tags: 好数字 · → daily
value: qty=68.6% · comparison=±8.7%
- 2026-06-11 · cyber benchmark lead time vs trend
Mythos Preview is ~7 months ahead of trend on cyber benchmarks
tags: 好数字 · → daily
value: qty=~7 months · date=April 2026
- 2026-06-11 · big jump in exploiting vulnerabilities
it represents a big jump in exploiting them
tags: 好观点 · → daily
- 2026-06-09 · 定价
Mythos Preview was priced at $25/$125 per million input/output tokens
tags: 好数字 · → daily
value: qty=$25/$125 · date=per million tokens
- 2026-05-13 · cyber capabilities
a newer Mythos Preview checkpoint completed both our cyber ranges, solving the range “The Last Ones” in 6 of 10 attempts and the previously unsolved “Cooling Tower” in 3 of 10 attempts.
tags: 好数字 · → daily
- 2026-05-13 · 完成 AISI cyber ranges 结果
Mythos Preview is the first AI model to complete both of AISI's cyber ranges, which measure models' cyberattack capabilities; GPT-5.5 solved only one of them
tags: 好信源 · → daily
value: qty=both · date=2026-05-13
- 2026-05-05 · token efficiency before step 11
One thing we can say is that before the step-11 gate, stronger models seem more token-efficient.
tags: 好观点 · → daily
- 2026-06-11 · capability to compromise affected systems
Mythos Preview is often able to use public information about real-world vulnerabilities to construct exploits to fully compromise affected systems
tags: 好观点 · → daily
- 2026-06-11 · low false positive rate in vulnerability discovery
several companies in Project Glasswing highlighted Mythos Preview’s low false positive rate when finding vulnerabilities
tags: 好观点 · → daily
时间轴 (近 20)
- 2026-06-26 ·
fact · ExploitBench Cap percent · → daily
- 2026-06-11 ·
fact · cyber benchmark lead time vs trend · → daily
- 2026-06-11 ·
fact · capability to compromise affected systems · → daily
- 2026-06-11 ·
fact · low false positive rate in vulnerability discovery · → daily
- 2026-06-11 ·
forecast · ahead of trend in exploiting vulnerabilities · → daily
- 2026-06-11 ·
fact · big jump in exploiting vulnerabilities · → daily
- 2026-06-09 ·
fact · 定价 · → daily
- 2026-06-09 ·
fact · 训练加速比 · → daily
- 2026-06-04 ·
fact · 模型加速训练代码性能 · → daily
- 2026-06-04 ·
fact · AI研究决策能力与人类对比 · → daily
- 2026-05-13 ·
fact · cyber capabilities · → daily
- 2026-05-13 ·
fact · UK AISI 网络安全挑战解决状态 · → daily
- 2026-05-13 ·
fact · 完成 AISI cyber ranges 结果 · → daily
- 2026-05-05 ·
narrative · step-11 gate · → daily
- 2026-05-05 ·
fact · token efficiency before step 11 · → daily
- 2026-05-05 ·
fact · progress rate after step 11 · → daily
- 2026-05-02 ·
fact · 性能表现达到10M tokens · → daily
- 2026-05-02 ·
fact · 超过10M tokens后性能激增 · → daily
- 2026-04-30 ·
fact · 平均通过率 · → daily
- 2026-04-30 ·
fact · sycophancy rate reduction vs Opus 4.7 on relationship guidan · → daily
← 实体目录 · 系统日志