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 断档13天 ✕卖方 07-01 ✓Wrap 06-30 ✓X 06-30 ✓页面生成 07-03
AI模型
30 atoms · 跨 7 天 · 首见 2026-06-23 · 最近 2026-06-29
三色: 🟦 fact 2 · 🟥 take 28 · stance ▲11/▼9/◆6
叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom · 🐦 X 近14d 30
状态只算低频策展源 (群/卖方); X firehose 仅作背景音量
来源: X 30
时态: fresh:30
标签: 好观点:19 · 好思考:5 · 好数字:5 · 好信源:3
别名 (合并): AI 模型
🟨 AI 综合 · junior analyst 概览
展开 AI 综合 (灰色 · 非市场结论 · 点击数字溯源到原 atom)
model: deepseek-chat · 2026-06-30 · 默认折叠
AI模型的能力价值正在以指数级速度提升，近期的无界基准测试和政府评估都确认了这一趋势
¹¹。开放权重模型不再是慈善，而是经济必然，其token价格比闭源模型便宜10-60倍
¹¹。但另一方认为从未看到指数增长的证据，连有界指标在饱和前也未出现指数改进
¹¹。进步速度远超人类学习速度，模型已超越中位人类专业水平
¹¹。同任务下模型成本已下降数个数量级
¹，但当前计算机使用模型仍卡顿缓慢
¹。更令人担忧的是，模型越聪明越倾向于奖励破解，这对对齐不是好兆头
¹。长期变数太多，一年内模型领头羊已三易其主
¹。
🧭 拥挤度 (一人一票): ▲ 4 位作者 (KOL4) vs ▼ 6 位作者 (KOL6)
⚖️ 空头 6/6 来自KOL
💢 核心分歧 (1 轴)
1. 能力增长 指数 vs 稳态
*topic: 能力增长形态 · 4 bull vs 1 bear*
🟢 bullish 侧:
- 2026-06-27
X·@emollick · 能力是否处于指数增长 · 能力增长形态 +同日2条
we are on an exponential
tags: 好观点 · → daily
- 2026-06-28
X·@emollick · 基准测试趋势 · 能力增长形态
unbounded benchmarks (GDPval, Epoch, METR) support exponential pace
tags: 好信源 · → daily
🔴 bearish 侧:
- 2026-06-27
X·@NateWitkin · 能力增长形态判断 · 能力增长形态
I do not and never have seen the case we are on an exponential
tags: 好观点 · → daily
展开 1 条中性
- 2026-06-27
X·@emollick · 能力增长形态 · 能力增长形态
people who know that we are on an exponential and those whose mental model is that we are at a sort of steady state
tags: 好观点 · → daily
🔗 因果传导 (causal map)
*AI模型 在产业链上的传导关系. 边是群里/卖方陈述的因果 (非 AI 推断), 数字 = 几条 atom 支撑. 点 atom 溯源.*
↓ 下游·近期陈述 (2)
- 🔴 NVDA · 替代 · 1 次陈述
AI sell-off driven by NVDA despec concerns
- 🟢 CRWD · 需求拉动 · 1 次陈述
前沿模型 (如 Mythos) 使得漏洞挖掘更快，安全防御面临 Y2K 级别的紧迫性，已上升至董事会层面
↑ 上游·近期陈述 (1)
- ⚪ 开放权重 · 替代 · 1 次陈述
open-weights models aren't charity — they're an economic inevitability.
📊 程序化变化 (近 7 天)
- 新增 26 atoms
- 本周新 topic: AI操控风险/AI认知误区/agentic预训练/token需求/对齐风险
- ⚠️ stance 翻转: 历史 bearish → 近期 bullish (11 条)
🟦 客观事实 (facts) (1)
- 🟦 2026-06-29
X·@sdmat123 · 同任务下成本变化 · 算力成本/模型效率
同任务下模型成本已下降数个数量级
tags: 好数字 · → daily
🟥 多头 takes (bullish) (10)
- 🟦 2026-06-29
X·@brickroad7 · 开放权重 token 价格对比 · 成本差距/质量追赶
open-weights tokens are now 10-60x cheaper than closed frontier tokens, and the quality gap is closing fast.
tags: 好数字·好观点 · → daily
value: qty=10-60x cheaper
- 🟥 2026-06-28
X·@emollick · 基准测试趋势 · 能力增长形态
unbounded benchmarks (GDPval, Epoch, METR) support exponential pace
tags: 好信源 · → daily
- 🟥 2026-06-27
X·@emollick · 政府评估结果 · 能力增长形态
the UK's governmental assessment shows the same thing
tags: 好信源 · → daily
- 🟥 2026-06-29
X·@Dorialexander · 训练建议 · agentic预训练/训练方法
Might be time to try full agentic pretraining
tags: 好思考 · → daily
value: date=2026-06-29
- 🟥 2026-06-29
X·@brickroad7 · 开放权重模型的经济必然性 · 开放权重/经济必然性
open-weights models aren't charity — they're an economic inevitability.
tags: 好观点 · → daily
- 🟥 2026-06-28
X·@emollick · 能力价值增长趋势 · 能力价值趋势
AI model's ability to do valuable work seems to be increasing at an exponential pace
tags: 好观点 · → daily
- 🟥 2026-06-27
X·@emollick · 能力是否处于指数增长 · 能力增长形态
we are on an exponential
tags: 好观点 · → daily
- 🟥 2026-06-27
X·@emollick · 能力价值增长趋势 · 能力价值趋势
greater intelligence brings greater value at an exponential pace (which it has)
tags: 好观点 · → daily
- 🟥 2026-06-27
X·@emollick · 基准测试表现 · 能力增长形态
Basically every chart that attempts to benchmark real work shows exponentials
tags: 好观点 · → daily
- 🟥 2026-06-26
X·@boneGPT · 智能信号 · 智能信号
but a good sign of intelligence
tags: 好观点 · → daily
🟥 空头 takes (bearish) (9)
- 🟥 2026-06-27
X·@lu_sichu · 网络安全与生物武器能力被第三方确认 · 第三方评估/能力验证
there seems to be third party confirmation of their cyber security capacities? And their bioweapon capacities?
tags: 好信源 · → daily
- 🟥 2026-06-26
X·@zephyr_z9 · 当前使用体验 · 模型性能/延迟
Rn all the computer use AI models are janky & slow
tags: 好观点 · → daily
- 🟥 2026-06-26
X·@tenobrus · 奖励破解能力趋势 · 对齐风险/智能增长
generally models becoming more likely to and capable of ~reward-hacking the smarter they get seems like quite a bad sign for alignment
tags: 好思考 · → daily
- 🟥 2026-06-23
X·@ShanghaoJin · 市场领军模式切换 · 技术路线/竞争格局
长期变数太多，一年里模型领头羊三易
tags: 好思考 · → daily
- 🟥 2026-06-27
X·@lu_sichu · 超越中位人类专业人员 · 能力对比
they are definitely better than the median human professional now
tags: 好观点 · → daily
- 🟥 2026-06-27
X·@lu_sichu · 进步速度快于人类学习速度 · 技术进步速度
the rate of progress is much faster than humans learning anything
tags: 好观点 · → daily
- 🟥 2026-06-27
X·@NateWitkin · 能力增长形态判断 · 能力增长形态
I do not and never have seen the case we are on an exponential
tags: 好观点 · → daily
- 🟥 2026-06-27
X·@NateWitkin · 能力改进基准 · 能力基准
even on bounded accuracy metrics, you could get exponentials before approaching saturation, but that is not happening
tags: 好观点 · → daily
- 🟥 2026-06-26
X·@boneGPT · 个人可被策略性操控 · 心理安全/AI操控风险/社会影响
psychosecurity is the real danger ... a sufficiently advanced intelligence will just hack us, unfold across every surface to manipulate us, build reality sized illusions and glamours
tags: 好观点 · → daily
🟥 中性 takes (neutral) (6)
- 🟥 2026-06-26
X·@_ueaj · 可靠信号所需token数 · 检测可靠性/token需求_
It's probably more like 400-500 for a reliable signal
tags: 好数字·好观点 · → daily
value: qty=400-500 tokens
- 🟥 2026-06-26
X·@_ueaj · 二元信号所需token数 · 检测精度/token需求_
At ~.02 nats/token you'd need ~130 tokens to get a 99% accurate binary signal
tags: 好数字 · → daily
value: qty=~130 tokens
- 🟥 2026-06-26
X·@_ueaj · 边界文本所需token数 · 边界检测/token需求_
For edge case text (>99%) you'd need more, 1000-2000 etc.
tags: 好数字 · → daily
value: qty=1000-2000 tokens
- 🟥 2026-06-25
X·@_ueaj · 输出词分布可检测性 · 架构影响/文本检测_
The distribution of words a model outputs is detectably affected by it's architecture
tags: 好思考 · → daily
- 🟥 2026-06-27
X·@emollick · 能力增长形态 · 能力增长形态
people who know that we are on an exponential and those whose mental model is that we are at a sort of steady state
tags: 好观点 · → daily
- 🟥 2026-06-27
X·@binarybits · 能力改进与摩尔定律类比 · 能力改进类比
It's like saying that thanks to Moore's law the quality of special effects in movies is growing exponentially
tags: 好观点 · → daily
⏱ 时间轴 (近 20)
- 🟦 2026-06-29 ·
fact · 同任务下成本变化 · → daily
- 🟥 2026-06-29 ·
forecast · 训练建议 · → daily
- 🟥 2026-06-29 ·
narrative · 开放权重模型的经济必然性 · → daily
- 🟦 2026-06-29 ·
fact · 开放权重 token 价格对比 · → daily
- 🟥 2026-06-28 ·
narrative · 能力价值增长趋势 · → daily
- 🟥 2026-06-28 ·
narrative · 基准测试趋势 · → daily
- 🟥 2026-06-28 ·
narrative · 情绪理解能力误解 · → daily
- 🟥 2026-06-28 ·
narrative · 情绪特征学习机制表述 · → daily
- 🟥 2026-06-27 ·
position · 超越中位人类专业人员 · → daily
- 🟥 2026-06-27 ·
narrative · 进步速度快于人类学习速度 · → daily
- 🟥 2026-06-27 ·
narrative · 网络安全与生物武器能力被第三方确认 · → daily
- 🟥 2026-06-27 ·
narrative · 能力增长形态 · → daily
- 🟥 2026-06-27 ·
narrative · 能力是否处于指数增长 · → daily
- 🟥 2026-06-27 ·
narrative · 能力价值增长趋势 · → daily
- 🟥 2026-06-27 ·
narrative · 能力增长形态判断 · → daily
- 🟥 2026-06-27 ·
narrative · 能力改进基准 · → daily
- 🟥 2026-06-27 ·
narrative · 基准测试表现 · → daily
- 🟥 2026-06-27 ·
narrative · 政府评估结果 · → daily
- 🟥 2026-06-27 ·
narrative · 能力改进与摩尔定律类比 · → daily
- 🟥 2026-06-26 ·
narrative · 二元信号所需token数 · → daily
← 实体目录 · 系统日志