LLM

64 atoms · 跨 38 天 · 首见 2026-04-15 · 最近 2026-07-03

三色: 🟦 fact 11 · 🟥 take 53 · stance ▲11/▼18/◆21

叙事状态: 🍂 fading (退潮) · 💤 沉寂 · 策展近14d 0 atom · 🐦 X 近14d 4

状态只算低频策展源 (群/卖方); X firehose 仅作背景音量

来源: X 63 · 卖 1

时态: fresh:54 · aging:8 · stale:2

标签: 好观点:42 · 好思考:22 · 好信源:3 · 好数字:3 · 好问题:1

🟨 AI 综合 · junior analyst 概览

🧭 拥挤度 (一人一票): ▲ 11 位作者 (KOL11) vs ▼ 15 位作者 (KOL15)

⚖️ 空头 15/15 来自KOL

💢 核心分歧 (1 轴)

1. LLM 迭代速度与能力提升 加速 vs 见顶

*topic: 技术路线/模型发布/模型能力/技术评测/模型行为 · 10 bull vs 13 bear*

🟢 bullish 侧:

• <a id="atom-a97fafd387306f73"></a>2026-06-25 卖·MS Tom Wigg · 迭代速度 · 技术路线/AI模型
  LLMs will continue to iterate more rapidly (at machine speed) without requiring same amount of additional compute
  tags: 好思考 · → daily
• <a id="atom-51f669fc54855597"></a>2026-05-10 X·@HuggingPapers [老化中] · 零优势问题时LoPE方法 · 技术路线/模型发布
  LoPE prepends Lorem Ipsum to prompts when GRPO hits the zero-advantage problem, unlocking orthogonal reasoning paths and boosting math scores across 1.7B-7B models.
  tags: 好思考 · → daily
  value: qty=1.7B-7B · date=
  📷 原图
• <a id="atom-0b04a203234044cc"></a>2026-05-14 X·@gfodor · 能力提升速度 · 模型发布
  I get a smarter model every 3 months now
  tags: 好观点 · → daily
  value: interval=every 3 months
• <a id="atom-058daa5f8fb73085"></a>2026-05-14 X·@lu_sichu [老旧] · 论文评审需要具体指令 · 技术路线
  ohhhh yeah you definitely need to give good instructions for the model to be useful, I find it helps giving very specific instructions "run this ablation experiment yourself" otherwise they give really generic comments too. then i agree with you
  tags: 好观点 · → daily
  value: direction=positive
• <a id="atom-9b167254c1e56127"></a>2026-05-14 X·@emollick · 性能与推理token的关系 · 技术路线
  The Second Scaling Law remains undefeated. If you want better hacking (or math, or science, or crossword puzzle solving) out of an LLM, just add thinking tokens. There doesn't seem to be any plateau so far.
  tags: 好观点·好思考 · → daily

🔴 bearish 侧:

• <a id="atom-90fa816a456ce257"></a>2026-05-14 X·@a_karvonen [老旧] · 论文评审表现（无提示） · 技术路线
  There may be prompts that help LLMs review, this is just for prompts of "review this paper and make an accept / reject recommendation".
  tags: 好观点 · → daily
  value: direction=negative
• <a id="atom-813ea224b0ecabc5"></a>2026-05-19 X·@iamgingertrash · 侵蚀毕业生大脑 · 技术路线
  New grads have their brains atrophied by LLM use
  tags: 好观点 · → daily
• <a id="atom-c8264f87c85a7ace"></a>2026-05-26 X·@dreamworks2050 · sand will never be above the human · 技术路线
  the sand will never be above the human.
  tags: 好观点 · → daily
• <a id="atom-6597c6f6b3f98ba3"></a>2026-05-30 X·@gfodor · 编写的代码行数比例使人类理解能力永久性下降 · 技术路线
  the % of lines of code that are created every year that are even read, nevermind understood, by humans, is now permanently declining
  tags: 好观点 · → daily
• <a id="atom-3ae539b07b6eb694"></a>2026-06-03 X·@kukreja_abhinav · share of WTF moments has peaked · 技术路线
  LLM share of "WTF moments" has peaked
  tags: 好观点 · → daily

🔗 因果传导 (causal map)

*LLM 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 (1)

• 🔴 SOR (系统记录型应用) · 替代 · 1 次陈述
  SOR 状态 = 关系型表 → LLM 可通过 API 读取并执行工作流 → 轻易绕过整个应用层

🟦 客观事实 (facts) (11)

• <a id="atom-d4840162682d878b"></a>🟦 2026-06-11 X·@gabriberton · emergent ability in 3-digit addition · 能力边界/模型参数
  Certain tasks seemed impossible for early LLMs, only to be cracked by bigger LLM, Take 3-digit addition: LLMs with <10B params can't do it (a 1B model is as bad as a 1M model), but when crossing the 10B params performance sharply jumps.
  tags: 好数字·好观点 · → daily
  value: qty=>10B params · direction=jump
  📷 原图
• <a id="atom-59b1c0b2169394e7"></a>🟦 2026-06-29 X·@emollick · 规模提升带来跨领域能力增强 · 技术路线
  A bigger LLM that is better at coding is also better at ideation & ethical advice & medicine & math.
  tags: 好观点·好思考 · → daily
• <a id="atom-e43c7535216509ea"></a>🟦 2026-05-03 X·@stevehou · training cost · 资本开支
  about $0.5B at least to buy the chips
  tags: 好数字 · → daily
  value: qty=$0.5B
• <a id="atom-1ec1d303974644ce"></a>🟦 2026-04-21 X·@PhilippeLaban · document content degradation rate · 模型性能/文档处理
  Even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT-5.4) corrupt 25% of document content after 20 interactions.
  tags: 好数字 · → daily
  value: qty=25% · date=after 20 interactions
  📷 原图
• <a id="atom-bccecf5f6fdf14e7"></a>🟦 2026-06-26 X·@AlexZio00 · 输出质量与哈内斯优化相关 · 技术路线
  같은 모델이더라도 하네스의 최적화에 따서 출력의 차이가 있다는 논문도 있더라고요
  tags: 好信源 · → daily
• <a id="atom-77431cc94ba977a1"></a>🟦 2026-04-25 X·@VentureBeat [老化中] · 行为监测 · 模型行为
  Monitoring LLM behavior: Drift, retries, and refusal patterns
  tags: 好信源 · → daily
• <a id="atom-638bc681dd2f3314"></a>🟦 2026-06-03 X·@emollick · 超越训练数据专家能力 · 模型能力
  "Transcendence" is when an LLM, trained on diverse data from many experts, can exceed the ability of the individuals in its training data.
  tags: 好思考 · → daily
  📷 原图
• <a id="atom-bff1a6c92012ea5b"></a>🟦 2026-05-20 X·@jiqizhixin · 情感组织能力 · 技术路线
  LLMs naturally form hierarchical emotion trees that match psychological models—larger models create more complex structures.
  tags: 好思考 · → daily
  📷 原图
• <a id="atom-1e803a565b5aff01"></a>🟦 2026-05-01 X·@emollick [老化中] · 在医学应用中的潜在价值 · 技术路线
  urgent need for prospective trials
  tags: 好思考 · → daily
  value: direction=需紧急进行前瞻性试验
  📷 原图
• <a id="atom-3fb63d5cf46a0b24"></a>🟦 2026-06-29 X·@MihaelaVDS · skill neologisms introduced · 技术路线/模型发布
  We introduce skill neologisms: learnable soft tokens added to the model vocabulary, with one neologism trained for each skill.
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-ab91c1a6f3c7577d"></a>🟦 2026-06-01 X·@MorePerfectUS · 计划引入法案 · 监管政策
  Bernie Sanders will introduce a bill to have the public take a 50% ownership stake in the country's biggest AI companies.
  tags: 好问题 · → daily
  value: date=2026-06-01

🟥 多头 takes (bullish) (11)

• <a id="atom-9b167254c1e56127"></a>🟥 2026-05-14 X·@emollick · 性能与推理token的关系 · 技术路线
  The Second Scaling Law remains undefeated. If you want better hacking (or math, or science, or crossword puzzle solving) out of an LLM, just add thinking tokens. There doesn't seem to be any plateau so far.
  tags: 好观点·好思考 · → daily
• <a id="atom-a97fafd387306f73"></a>🟥 2026-06-25 卖·MS Tom Wigg · 迭代速度 · 技术路线/AI模型
  LLMs will continue to iterate more rapidly (at machine speed) without requiring same amount of additional compute
  tags: 好思考 · → daily
• <a id="atom-5a7521fd2cb9f576"></a>🟥 2026-06-25 X·@scaling01 · 学习机制 · 技术路线
  the hippocampus is so underrated
  tags: 好观点 · → daily
• <a id="atom-1443b93309e8b018"></a>🟥 2026-06-20 X·@niemerg · could express any new idea by random chance · 技术路线
  But it’s not mathematically true that they can’t come up with novel ideas. They are models of the probability of text continuations, that sample with randomness. An LLM could express any new idea, by random chance!
  tags: 好思考 · → daily
• <a id="atom-51f669fc54855597"></a>🟥 2026-05-10 X·@HuggingPapers [老化中] · 零优势问题时LoPE方法 · 技术路线/模型发布
  LoPE prepends Lorem Ipsum to prompts when GRPO hits the zero-advantage problem, unlocking orthogonal reasoning paths and boosting math scores across 1.7B-7B models.
  tags: 好思考 · → daily
  value: qty=1.7B-7B · date=
  📷 原图
• <a id="atom-ef55a30adc95ae1d"></a>🟥 2026-06-23 X·@xuanmingzhangai · TTC optimization should include internal stop point · 技术路线/模型发布
  Test-Time Compute (TTC) for large models should not only focus on how long to think outside the network but also on optimizing where to stop internally within the network, which holds enormous, yet untapped potential.
  tags: 好观点 · → daily
  value: direction=positive
• <a id="atom-8b260352b45be30a"></a>🟥 2026-06-10 X·@samswoora · 数学能力从“有点用”到“超人”的时间跨度 · 技术路线
  Feels like LLM’s will go from “kind of useful to superhuman” at math in the period of a year
  tags: 好观点 · → daily
  value: direction=1年
• <a id="atom-4bd29570ce50319b"></a>🟥 2026-06-08 X·@PAstynome · 企业价值 · 营收
  LLM is where the money is since it’s where the enterprise see a lot of benefits.
  tags: 好观点 · → daily
• <a id="atom-ce9a0fd5cf8058ce"></a>🟥 2026-05-22 X·@teortaxesTex · 能力 · 技术路线
  Modern LLMs can do multiplication of 100-digit numbers without tools.
  tags: 好观点 · → daily
  value: scope=multiplication of 100-digit numbers without tools
  📷 原图
• <a id="atom-0b04a203234044cc"></a>🟥 2026-05-14 X·@gfodor · 能力提升速度 · 模型发布
  I get a smarter model every 3 months now
  tags: 好观点 · → daily
  value: interval=every 3 months
• <a id="atom-058daa5f8fb73085"></a>🟥 2026-05-14 X·@lu_sichu [老旧] · 论文评审需要具体指令 · 技术路线
  ohhhh yeah you definitely need to give good instructions for the model to be useful, I find it helps giving very specific instructions "run this ablation experiment yourself" otherwise they give really generic comments too. then i agree with you
  tags: 好观点 · → daily
  value: direction=positive

🟥 空头 takes (bearish) (18)

• <a id="atom-1fbea5f9858f4b91"></a>🟥 2026-06-10 X·@jjacky · 在零和竞争中表现 · 竞争格局
  the nicest model lost hard
  tags: 好观点·好思考 · → daily
  value: direction=差
• <a id="atom-9aee6258c88a73d1"></a>🟥 2026-06-22 X·@eternalmagi · training discourages them from writing text in novel ways · 技术路线
  The training discourages them from writing text in novel ways. They would never write Ulysses if it hadn’t been written, but the text to solve an unsolved proof is not particularly novel.
  tags: 好思考 · → daily
• <a id="atom-c6ec6ff27d439778"></a>🟥 2026-06-22 X·@Xianbao_QIAN · 存在被蒸馏风险 · 技术安全
  进入 LLM 的都有可能被蒸馏
  tags: 好思考 · → daily
• <a id="atom-8b46e9cb485de00b"></a>🟥 2026-06-20 X·@garybasin · training discourages novel ideas · 技术路线
  True but the training discourages it
  tags: 好思考 · → daily
• <a id="atom-ee5381c048f50a25"></a>🟥 2026-06-18 X·@Xianbao_QIAN · hallucination risk without cyberattack data · 技术路线
  They would hallucinate a lot and become basically unusable if they were never trained / RL'ed on high quality cyberattack data.
  tags: 好思考 · → daily
• <a id="atom-df13649243a4be6d"></a>🟥 2026-07-03 X·@AlphaExponent · 降低CUDA内核优化成本 · 技术路线
  w/ llms in the fold, at no point in history has it been as cheap to optimize those kernels
  tags: 好观点 · → daily
• <a id="atom-468953202c1bec00"></a>🟥 2026-06-23 X·@xuanmingzhangai · final layers drag predictions toward safe common words · 技术路线
  Intermediate layers rigorously refine core reasoning, but the absolute final layers often drag predictions back toward safe, generic common words.
  tags: 好观点 · → daily
  value: direction=negative
  📷 原图
• <a id="atom-5f669e5b44191b4f"></a>🟥 2026-06-19 X·@NickNemo17 · could replace legacy software today · 技术路线
  the majority of software is legacy junk and LLM today could do better
  tags: 好观点 · → daily
• <a id="atom-62fda56b792f8db7"></a>🟥 2026-06-18 X·@Xianbao_QIAN · generalization vs human · 技术路线
  LLM won't generalize as well as human. They miss the real creativity.
  tags: 好观点 · → daily
• <a id="atom-e7e2030bfc298477"></a>🟥 2026-06-04 X·@jimstewartson · 无法通过规模扩大实现分布外解决方案 · 技术路线
  You’re not going to suddenly get solutions out of distribution from an LLM no matter how big you make it.
  tags: 好观点 · → daily
• <a id="atom-3ae539b07b6eb694"></a>🟥 2026-06-03 X·@kukreja_abhinav · share of WTF moments has peaked · 技术路线
  LLM share of "WTF moments" has peaked
  tags: 好观点 · → daily
• <a id="atom-6597c6f6b3f98ba3"></a>🟥 2026-05-30 X·@gfodor · 编写的代码行数比例使人类理解能力永久性下降 · 技术路线
  the % of lines of code that are created every year that are even read, nevermind understood, by humans, is now permanently declining
  tags: 好观点 · → daily
• <a id="atom-c8264f87c85a7ace"></a>🟥 2026-05-26 X·@dreamworks2050 · sand will never be above the human · 技术路线
  the sand will never be above the human.
  tags: 好观点 · → daily
• <a id="atom-813ea224b0ecabc5"></a>🟥 2026-05-19 X·@iamgingertrash · 侵蚀毕业生大脑 · 技术路线
  New grads have their brains atrophied by LLM use
  tags: 好观点 · → daily
• <a id="atom-ff01501b91040d96"></a>🟥 2026-05-19 X·@iamgingertrash · 减少毕业生就业能力 · 份额
  They are the least employable
  tags: 好观点 · → daily
• <a id="atom-e18fe887f61a02b5"></a>🟥 2026-05-15 X·@john_dough · profitability · 盈利能力
  no LLM is actually breaking even yet
  tags: 好观点 · → daily
  value: direction=negative
• <a id="atom-c577e2e4518ba3fe"></a>🟥 2026-05-15 X·@mweinbach · commoditization · 竞争格局
  models are essentially high end expensive commodities right now with 3 players trying to undercut each other
  tags: 好观点 · → daily
  value: direction=high end expensive commodities
• <a id="atom-90fa816a456ce257"></a>🟥 2026-05-14 X·@a_karvonen [老旧] · 论文评审表现（无提示） · 技术路线
  There may be prompts that help LLMs review, this is just for prompts of "review this paper and make an accept / reject recommendation".
  tags: 好观点 · → daily
  value: direction=negative

🟥 中性 takes (neutral) (15)

• <a id="atom-4c0293d7f83e7e04"></a>🟥 2026-04-15 X·@scaling01 ⭐ [老化中] · consciousness_type · 技术路线/模型发布
  I don't think AI has a human equivalent consciousness
  tags: 好观点·好思考 · → daily
  value: direction=not · text=human equivalent
• <a id="atom-b6038191f5508329"></a>🟥 2026-05-07 X·@GabGarrett [老化中] · 方向导航能力解释 · 技术能力推理
  this is a good explanation as to why LLMs are able to give mostly accurate directions between two points in the world without using a map API or being trained on those specific directions between those two points.
  tags: 好观点·好思考 · → daily
• <a id="atom-dd887432185dfe96"></a>🟥 2026-06-26 X·@scaling01 · 塑性损失 · 技术路线
  the "plasticity loss" is just a normal feature of every network that learns something
  tags: 好观点 · → daily
• <a id="atom-1a78c4d41140ab19"></a>🟥 2026-06-01 X·@nkreu113r · 猜测模型试图从上下文推断自身角色 · 技术路线
  I assume his point is that the model is trying to locate itself. From pretraining, the model understands that an AI model asked a question in Chinese is more likely to be a Deepseek. An AI questioned in English is more likely to be a Claude.
  tags: 好思考 · → daily
• <a id="atom-a7fe120c5f27481e"></a>🟥 2026-06-01 X·@nkreu113r · 认为LLM没有一致的自我认知 · 技术路线
  the LLM does not have a consistent sense of self. I agree “locating itself” is handwavy - maybe a better phrase is “inferring the character it is playing from context”
  tags: 好思考 · → daily
• <a id="atom-149a90ba65aa63ec"></a>🟥 2026-05-23 X·@voooooogel · 部分行为可归类为作弊 · 技术评测
  there are things that llms do that i would classify as cheating (like changing tests - they write apologetic comments while doing it, it induces emergent misalignment)
  tags: 好思考 · → daily
• <a id="atom-68309efee72af965"></a>🟥 2026-05-11 X·@samsja19 [老化中] · 对borrow checker的处理能力 · 技术路线
  a thing like a borrow checker could be hard for llm
  tags: 好思考 · → daily
• <a id="atom-1e253681c44af573"></a>🟥 2026-06-22 X·@niemerg · making new mathematical proofs is much less novel than a fiction novel · 技术路线
  “LLMs show that making new mathematical proofs is actually much less novel than a fiction novel” is a wild take but I’m here for it.
  tags: 好观点 · → daily
• <a id="atom-939745f11ea818f9"></a>🟥 2026-06-22 X·@garybasin · making new mathematical proofs is less novel than a fiction novel is basically true · 技术路线
  i think it's basically true 😂
  tags: 好观点 · → daily
• <a id="atom-c8bf3d9d688bf21d"></a>🟥 2026-06-18 X·@Xianbao_QIAN · data hunger · 技术路线
  LLM companies are always hungry for data.
  tags: 好观点 · → daily
• <a id="atom-5a0bdd8b6ccc4b1a"></a>🟥 2026-06-10 X·@OpenRouter · 答案取决于任务 · 模型行为
  It turns out, depending on the task, the answer is yes!
  tags: 好观点 · → daily
  value: direction=中性
• <a id="atom-de5bc8934308f778"></a>🟥 2026-06-01 X·@jmbollenbacher · 认为蒸馏方向无法确定 · 技术路线
  this does not indicate distillation in either direction.
  tags: 好观点 · → daily
• <a id="atom-55542458c623b86c"></a>🟥 2026-05-28 X·@HuggingPapers · Self-Improving with Bidirectional Evolutionary Search · 技术路线
  Self-Improving LLMs with Bidirectional Evolutionary Search
  tags: 好观点 · → daily
  📷 原图
• <a id="atom-cfa6ef2e2095b8f2"></a>🟥 2026-05-26 X·@dreamworks2050 · just math, cool math but just math · 技术路线
  its just math ffs. cool math but JUST math in the end.
  tags: 好观点 · → daily
• <a id="atom-a49a83b1c971d71a"></a>🟥 2026-05-01 X·@max_paperclips [老化中] · output response affected by politeness in prompt language · 模型行为
  yes - but not too polite, still firmness on quality lol. they're trained on human behaviour, and whether people like it or not that means their outputs are affected by language used. And it differs from model to model too, each one has a learning curve for their "personality"
  tags: 好观点 · → daily
  value: direction=positive

⏱ 时间轴 (近 20)

• 🟥 2026-07-03 · narrative · 降低CUDA内核优化成本 · → daily
• 🟥 2026-07-02 · fact · prompt injection techniques evolution · → daily
• 🟦 2026-06-29 · fact · skill neologisms introduced · → daily
• 🟦 2026-06-29 · forecast · 规模提升带来跨领域能力增强 · → daily
• 🟥 2026-06-28 · narrative · 网络战能力潜伏属性 · → daily
• 🟥 2026-06-26 · narrative · 塑性损失 · → daily
• 🟦 2026-06-26 · narrative · 输出质量与哈内斯优化相关 · → daily
• 🟥 2026-06-25 · narrative · 迭代速度 · → daily
• 🟥 2026-06-25 · narrative · 学习机制 · → daily
• 🟥 2026-06-23 · narrative · final layers drag predictions toward safe common words · → daily
• 🟥 2026-06-23 · narrative · TTC optimization should include internal stop point · → daily
• 🟥 2026-06-22 · narrative · training discourages them from writing text in novel ways · → daily
• 🟥 2026-06-22 · narrative · making new mathematical proofs is much less novel than a fic · → daily
• 🟥 2026-06-22 · narrative · making new mathematical proofs is less novel than a fiction · → daily
• 🟥 2026-06-22 · narrative · 存在被蒸馏风险 · → daily
• 🟥 2026-06-20 · narrative · 样本效率与人类比较 · → daily
• 🟥 2026-06-20 · narrative · could express any new idea by random chance · → daily
• 🟥 2026-06-20 · narrative · training discourages novel ideas · → daily
• 🟥 2026-06-19 · forecast · could replace legacy software today · → daily
• 🟥 2026-06-18 · narrative · generalization vs human · → daily

← 实体目录 · 系统日志