Jason X
1.2K posts
今天晚上各个群都在传一份《国内大模型蒸馏风波的来龙去脉》。
我大概看了一下，很可惜，这更像是一份由外行根据各种流言拼凑出来的 AI slop。
里面的内容有真有假，但几个核心判断基本都经不起推敲。
比如里面声称，智谱 @Zai_org 在今年 4 月份就已经破解并开始蒸馏 Fable。
这个时间线真的很离谱。
因为 4 月份的时候，Fable （那个时候还是Mythos）甚至还没有正式发布，能够使用的基本都是 Anthropic 的内部测试用户，通过Project Glasswing进行访问。
那智谱是怎么拿到的访问权限呢？
难道是 @AnthropicAI 的邀请或者是五角大楼反代吗🤣
还有一个比较典型的问题，是作者对于传闻中DeepSeek 将部分请求路由到 Fable 表示不理解，认为这样算不过来经济账。
实际上，但凡是个从业者或者稍微了解模型公司的研发流程，就知道这件事情不能简单按照 API 成本计算。
如果一个 frontier model 可以帮助你生成高质量训练数据、评测数据，或者提升模型迭代速度，那么获取这些能力本身就是一种研发投入。
用单次调用价格去判断整个策略是否划算，本身就是把模型公司的研发逻辑想简单了。
但最离谱的部分，是 PDF 里面声称 @Kimi_Moonshot 在 K3 发布前，把整个 RL 团队全部解雇了❓❓
这个说法目前没有任何可靠依据。
实际上情况恰恰相反，根据我的确认Kimi RL 相关团队目前仍然存在，并没有所谓“整个 RL 组被裁撤”的情况。
甚至 PDF 里面还进一步延伸出一套完整叙事：
K3 = 蒸馏 + 刷榜 + 裁撤 RL。
我一开始还在认真的区分哪些是事实，哪些是推测，哪些只是作者脑补出来的故事。
直到我看到 PDF 后半部分开始大量加入意识形态化表达，把技术讨论上升到所谓“国模黑暗时代”“技术偷取”等叙事，我反而释然了。
因为这已经不是在分析 AI 行业，而是在借 AI 叙事讲另一个故事。
PS：我感觉这个PDF是黑KIMI来的，因为KIMI的篇幅最大，但里面的大部分内容都是假的，其他家就有真有假了。
中文
@hosseeb @deanwball Advanced open-source models benefit semiconductor manufacturers and AI application layers; the only negative factor is closed-source models. Don't use international competition as an excuse.
English
This argument by @deanwball is being badly misunderstood. It's OK to disagree with it, but first you have to actually understand what he's saying.
He's saying: releasing the weights for a frontier-level model is effectively dumping.
Dumping is when you sell a product at significantly below cost in order to corner market share. It's illegal. The reason: dumping results in short-term consumer surplus, but long-term it prevents the formation of a competitive market and discourages capex outside of the dumper. Standard Oil famously did this in order to consolidate the oil market before it was broken up.
So why is he claiming releasing the weights of a frontier level model is basically dumping? Isn't he just describing open source?
His argument: it's not financially sustainable to train a frontier model and release the weights. In the long run, you will not be able to internalize enough of the gains given the cost of training a frontier model, because neoclouds and other inference providers will be able to outcompete you at actually serving the model. It costs an astronomical amount of money to train frontier models, and if everyone else can serve them, you don't capture enough of the surplus to pay for the training and R&D.
It's not like normal open source when you build some software and then release it and sell services on top of it. The amount of capex required for frontier-level models is an order of magnitude higher than normal software, which is why doing this at frontier level is so economically irrational.
Right now the Hong Kong stock market is ebullient enough that Chinese AI companies are not getting punished for the fact that they're all deeply, deeply unprofitable. Releasing model weights is great marketing, intellectually appealing, and strikes fear into the hearts of their opponents. We can assume the status quo continues for a while because of the AI supercycle. But eventually the AI market will correct, the Hong Kong market will dump, and suddenly these Chinese labs won't be able to afford to training super expensive models without internalizing more of the gains.
But what if China, seeing that this strategy is successfully kneecapping the US lead (by discouraging further capex and lowering valuations), says no--don't stop. And so the Chinese government starts buying up the shares of these companies and demanding that they continue releasing frontier-level weights, profitable or not.
In that case, it becomes a genuine space race. For-profit companies cannot continue to compete on either side. US labs valuations fall, and the White House realizes that to keep their advantage in the AI race, they cannot rely on the free market to maintain their lead. They nationalize the labs and fund them off government subsidies.
Now you have government-controlled and distributed models on both sides. That's what Dean is calling the "dystopian hellscape."
The best analogy is drug development: if China were to sell American drugs back to us really cheaply, that would result in a large short-term consumer surplus. Cheap Viagra and Ozempic is obviously great. But in the long run, this would discourage investment in developing new drugs. That's the sense that Dean is saying it's long-term "decel."
Now, I happen to disagree with Dean. I think the consumer surplus of having frontier-level open weight models is huge, even at the current capabilities. I also think China is going to defect from this strategy soon (there's been reporting along these lines, that Beijing will stop allowing large models to be open-weight; I think there are other reasons for this aside from competition). I also suspect that nationalization of labs is inevitable as they take on more geopolitical and cyber capabilities.
But he's not wrong--releasing frontier-level weight models is weird.
The question of how long this market will remain profit-driven is a very coherent question to ask.
martin_casado@martin_casado
"Open-weight models are inherently decelerationist" .... this is a grossly incorrect statement with no supporting arguments or logic that is counter to the long arc of learnings of the industry over the last 50 years. What a stupid thing to say.
English
@RYANHINGSHING 企业端肯定会更愿意用开源模型，企业的核心数据怎么会给闭源公司，Alex carp前几天不是炮轰了某些闭源公司吗？
中文
在幾個月前我說過中國模型在程序能力追近美國模型的趨勢，所以今天kimi在程序能力追近fable對我來說不令人意外
不過我完全不看好開源模型（特別是老中的）在企業端的應用，因有能力以外的因素🤡
至於成本方面，老美模型已有規模優勢，他們會找尋合理定價來進一步限制開源模型的空間
Ryan Leung 梁興盛, CFA@RYANHINGSHING
更正确的说法： 前沿闭源大模型和中国开源模型整体能力之间的差距变大了，但是程序能力差距变小了 IMO, 因为中国1. 能利用Github资源 2. 更容易做到代码任务所需要的可验证强化学习 3. 蒸馏
中文
screw it. i want to give it away.
Here's the exact Claude Code setup we run on every single client MVP at DreamLaunch.
50+ reusable prompts. 12 build workflows. the security and payment checklists keeping 50+ live apps from breaking in production. the client handoff doc we hand over every time.
this is not a "how to use AI" post. this is the literal internal system. same one we used on our last 8 launches.
Like + comment "STACK" and i'll send you the full system. 100% free.
(must be following for the DM to send)
English
I built a Claude skill that makes Vox-style ads
> writes the voice-over
> breaks the ad into clips with timestamps.
> Recreates the video with Arcads MCP in Claude
Comment "VOX" and I'll send you the skill
English
@Brad_Setser Don't pin too many hopes on Macron; he’s spouted so many slogans that I’m absolutely fed up with him. He is mostly just grandstanding to assert France's leadership role in Europe. Macron lacks real backbone.
English
and on a more political note, I do think China's leaders misunderestimated Europe, and made an error in judgment when they didn't respond more constructively to Macron's call for a serious discussion of imbalances. That helped galvinize European opinion
English
I am not thrilled by tariffs. I support very open trade in North America and across the North Atlantic (tho with less tax competition)
But China's model is not one of pure openness. And facing up to the challenge starts by moving beyond old talking points
Michael Pettis@michaelxpettis
This should really be an obvious point.
English
Jason X retweetledi
@aleabitoreddit 机器人赛道短期1-2年都不太行，AI物理大模型还没有达到gpt 1.0水平，机器人离真正进入日常生活、生产环境还有很大的距离
中文
专门写给我的中文读者：
绿的谐波（LeaderDrive，688017，577.3亿人民币）是我在布局人形机器人赛道时最青睐的中国上市标的。
他们的业务涵盖：
谐波减速器（据称占有超过60%的国内市场份额，以及1800多家全球客户）
人形机器人旋转关节减速器
直线执行器
电机/关节，以及许多其他核心零部件。
并且他们正在进军行星滚柱丝杠（planetary roller screws）领域。
目前占据着绝对的主导地位：超过60%的国内市场份额以及1800多家全球客户。
主要客户包括优傲机器人（Universal Robots）、优必选（UBTech）、智元机器人（Agibot），而特斯拉（Tesla）、Figure 以及绝大多数其他人形机器人研发商都很可能是他们的潜在客户。
如果你看一下这些零部件的 BOM（物料清单），粗略估计它们占到了每台研发出的人形机器人的 4-15%，随着他们进军更多细分领域，这个比例可能还会更高。
我认为没有人能对这类公司进行精确的财务建模，其前景更多取决于未来几年内会有多少人形机器人实现量产。
这其实是对一家能够在每台生产出的人形机器人中占据可观份额的公司，进行的一种方向性做多（directional long）。
在可扩展的规模化量产方面，中国显然是该领域的领导者，许多西方玩家无法将成本降到绿的谐波那样的水平。
随着我们向实体人工智能（physical AI）的规模化迈进，我极其看好机器人这个赛道。
中文
Jason X retweetledi
GPT 5.5 Pro 调研生成了一份关于 Codex 的Goal指令如何用的文档。
仔细阅读学到了两个技巧：
1. 觉得写不好goal时，先用plan模式，让AI反问自己一些问题，让AI帮收敛写Goal指令。
提示词模板：
/plan Help me turn this vague task into a strong Codex goal.
Interview me for missing success criteria, verification commands, constraints, boundaries, iteration policy, and blocked stop conditions.
Then draft a final `/goal ...` command.
2. 写好Goal的六要素：结果、验证、约束、边界、迭代和阻塞条件
官方标准模板如下：
/goal [Outcome].
Verification: [commands/artifacts/evidence].
Constraints: [what must not change].
Boundaries: [allowed writes / forbidden paths].
Iteration policy: [one focused change, rerun checks, log progress].
Stop when: [evidence proves completion].
Pause if: [blocked conditions / human decisions / budget cap].
详细调研报告见评论区，有不少模板可直接用。
中文
Jason X retweetledi
5 useful questions if you want to learn LLM application development this year:
1. What should I build first?
2. How do I know if my RAG system works?
3. When do I need agents?
4. What should I evaluate?
5. How do I move from a notebook to a working application?
We cover these kinds of questions on LLM Zoomcamp.
LLM Zoomcamp is a free, hands-on course where we build practical LLM applications with RAG, agents, vector search, evaluation, monitoring, and more.
The live cohort starts on June 8, 2026.
Join here: github.com/DataTalksClub/…
English
Jason X retweetledi
Jason X retweetledi
Jason X retweetledi
Jason X retweetledi
Palantir knows everything.
The founder of Palantir is fleeing the country.
Leading Report@LeadingReport
Peter Thiel has temporarily relocated his family to Argentina, enrolled his children in school there, and bought property, partly due to Thiel’s concerns about the United States’ future, per NYT.
English
Jason X retweetledi
Jason X retweetledi
最近最火的Codex优化网络速度Use Case，写了个提示词，亲测效果不错：
1. 在Codex中输入 “/goal” ，如果中文版输入 “/目标”，如果不用，直接发提示词也行。
2. 提示词如下：
优化当前电脑的网络速度和稳定性。
请按“先诊断、再最小可逆修改、最后复测”的方式执行，不要直接破坏性重置网络。
诊断要求：
1. 先跑 before 基准：networkQuality、DNS 查询耗时、到路由器的 ping、到公网 DNS 的 ping。
2. 区分真实公网链路和本机代理/VPN/TUN：检查 scutil --nwi、route get default、scutil --dns、scutil --proxy。
3. 检查 Wi‑Fi 质量：频段、信道、带宽、RSSI、噪声、Tx Rate、周边干扰。
4. 检查 MTU、丢包、mDNS/DNS 缓存、网络服务顺序。
5. 找出高流量或会接管路由的后台进程，如 VPN、Tailscale、Shadowrocket、Stash、iCloud、Dropbox、网盘、下载器。
优化要求：
1. 只做安全、可逆、低风险修改。
2. 把真实使用的 Wi‑Fi/以太网排到网络服务第一位。
3. 禁用明显无用的伪网络服务或旧网络服务，但不要删除配置。
4. 根据实测 DNS 延迟设置更快的 DNS。
5. 刷新 DNS 和 mDNS 缓存。
6. 停止或提示我关闭明显占用带宽的后台程序。
7. 如果需要 sudo 或会影响 VPN/远程连接，先说明风险，不要强行执行。
复测要求：
1. 再跑 after：networkQuality、DNS 查询耗时、路由器 ping、公网 ping。
2. 对比 before/after：下行、上行、空闲延迟、加载延迟、丢包、DNS 耗时。
3. 总结发现的 3 个主要问题、已修复项、未修复但建议手动处理项。
中文
Jason X retweetledi
即梦Seedream 4.5跑了500个艺术家风格。
提示词用通用名词，更能看出不同艺术家AI生图特点：
“一个女子坐在窗边读书，一只猫趴在她腿上，窗外是花园，用{artist}的风格绘制。”
网站地址：jm-style.qiaomu.ai
好多个性极为鲜明，比如一些经常在潮流T恤看到的风格
中文