you can just do things. burnt out software engineer. building @opentrawl, community maintainer @openclaw since Dec 25, ex-Spotify. 看透国际博弈核心.
the swamp with bad food (🇳🇱)
Joined September 2025
- Tweets7.9K
- Following621
- Followers2.2K
- Likes68.9K
Pinned Tweet
🚢 PRODUCT LAUNCH: "where-was-i" - a CLI that tells you all the trips you've ever been on, just using Apple Photos. Nothing leaves your machine unless you want it to.
Use it to supercharge your Grok Bot/Muse/Instinct/Dot or just Codex/Claude.
Let me know what you find - anything interesting?
Prompt to set it up (& link to the repo) in the next tweet.
How it works:
- Apple Photos stores a location for each of your photos
- Your Mac also knows where your home is
- "where-was-i" uses this to find out the days where you weren't at home, where you went, and where you stayed the night - just from the locations of the photos you took
- This gives you a list of every trip you've ever been on (since you started using your iPhone).
- You can feed this to another AI agent (then the data does leave your machine).
~Simplest fix: Pangram every single tweet and reply.
Permanently ban everyone who scores too highly on the Pangram model from twitter forever, w/ phone number + iris + DNA + thumbprint verification.
ask notaries to authenticate to offset the revenue they are losing from EU Inc.
Fighting abuse & spam on a consumer app is a perpetual cat-and-mouse game:
Every solution you create is a bandaid for 3-12 months -- until the adversaries find a workaround. When money is involved, people are tenacious to circumvent your patch and keep the money flowing.
While large language models have made attacks more intractable, there are tools you can build to protect your product:
(1) An in-house Risk Scoring stack that assigns a 0-100 score of likelihood someone is a spammer or bot. This should be informed by machine learning, IP & email reputation data, device telemetry, among many other signals.
(2) A near-bulletproof method to "checkpoint" your highest risk users, such as liveness verification or phone number auth. This will have to be reinvented every year (e.g., CAPTCHA tests are obsolete now with today's models).
As new attacks emerge, you can move the risk threshold lower to checkpoint more people. However, you want to be cautious so you don't add unnecessary friction for too many people and impact usage.
oh hey this is kinda cool. no 5h limit and plan rename.
good timing since i bought a 3rd account to avoid hitting the 5h limit, and am now ramping up the token usage even further! (and getting a lot done). can't wait until Fable 5.5 too....
Dutch ING customers and Amex customers: has anyone found a good way to get ~real-time insight into your balances & transactions so that you can hook it up to your AI agents?
I really need this for personal agent use!
A ton of the time my AI is like "remember to pay this bill or they will hunt you down with dogs and kill you" but I paid the bill 3 weeks ago. Or "remember to chase XYZ to see if they paid you" and then I never bother doing it.
Last I saw, this was possible if you use some 3rd party because of EU payment services directives but it's not possible to hook up to the banks yourselves? I have agents looking into it but if anyone's done this before and you can share a repo I will be eternally grateful. Otherwise, well, hopefully I'll open source something here soon too!
inspired by the Chinese posters and my /better skill I asked Opus to look into the ways the CPC and various militaries operate with decision making, in order to make my personal agents work better.
This seems quite useful for everyone making Chief of Staff agents!
The common failure mode that I find is:
1. Agents don't have enough signal or access
2. Agents end up guessing or asking you stupid things, which is annoying
3. Agents are lazy, and end up telling you about all the things they couldn't find (or hallucinate them) instead of going and actually chasing them down.
4. Agents give you half-baked decisions instead of fully formed ones, so you have to yell at them instead of just saying "yes" or "no" to a full plan.
This is designed to mitigate that somewhat. Let's see if it works!
teapaw.org/jjpcodes/status/211210…
This is the most useful prompt I've found in months. I've been using it with Opus 5.5 as a skill called /better to great effect. Grab it here: gist.github.com/joshp123/a14…
Here's how it's really stretches the model's performance vs usual prompts:
1. The "what a top professional in the field would do" somehow reminds the model to think of a lot of stuff it wouldn't otherwise do, and (unless this is your core area of expertise), you wouldn't know either.
Obviously Claude/Codex always knows the best practices already, but often it needs reminding to use them (despite it being in the training data), so this helps.
2. Rawest material - models love using proxies for stuff (e.g. Claude sampling logs instead of reading them); so telling them to use rawest materials helps them go back to primary sources/think from first principles, avoiding a lot of problems that come from lazy models and the resulting indirection.
4. Designing independent success criteria: important for the model to know how to actually work; and forcing it to define them means you get better and stronger results vs just letting it vibe stuff. The risk here is that you end up with "proof porn" - think endless hashing or comparisons without testing the real result - all frontier models as of Oct 2026 (Astra 6/Sol 6.1/Opus 5.5/Fable 5.1) all love doing this, so keep an eye on them.
Josh retweeted
我很愿意认真回答这位德国朋友的疑问。
你的第一个问，其实是在说，由于中国是一个威权政府，因此人民在政治上不受认真对待。
这个命题本身就是错误的。威权政府，并不意味着人民不受认真对待。民主政府，也不意味着人民就受到认真对待。
新加坡是威权国家，菲律宾是民主国家，难道新加坡人的处境，不如菲律宾人吗？中国是威权国家，印度是民主国家，难道中国人的处境，不如印度人吗？
你的第二个问，你的发问本身并不严谨。不能笼统地说，中国人没有政治自由。这个问题要分两个维度来讲。
1、政治参与自由。
从这个维度讲，中国任何人有参与政治的自由。中国人都有权利申请入党、考公务员。筛选规则与民主制度不同，但不等于中国人没有参与政治的自由。
中国制度与民主制度的区别在于：
民主制度下，一个人只要有钱，可以在任何时候参与政治。就像川普一辈子没有从政，老了突然想要竞选总统，他就可以参选，而且成功了。
但在中国，任何官员都必须从基层公务员做起。一个人无法因为有钱，就能一步登天当上高官。
这种制度的好处有：
A、公平性。
中国有大量的高层领导人出身平民家庭，如胡锦涛、温家宝、汪洋、李强，等等。像习近平这种红二代，反而是个意外。
B、专业性。
任何一个高层领导人，都必须经过几十年不同级别岗位的历练，他对国家各个层面的情况的了解，以及他的治理经验，不是一个企业家、一个明星，或者一个未治理过一个城市的议员，所能相提并论的。
C、远见性。
由于领导人不受制于短期的、非理性的民意，因此可以规划长远的发展战略，并以唯一执政党确保该战略的长期、有效的实施。
D、政策不受资本干涉。
你选择从政，就不能从商。你选择从商，就不能从政。这样一来，资本的归资本，政治的归政治，资本无法干涉政治。
2、政治权利自由。
当然，中国制度有其“表面上”的代价。为什么只是“表面上”的，后面再说。先说这种代价。
代价就是中国人不拥有谩骂政府、抨击政策的自由。我们可以批评政府，但必须在一种就事论事的理性语境中。我们不能喊“习近平下台”或者“共产党下台”。
我们也不拥有全民选举的自由，建立政党的自由。
这些“代价”为什么只是表面上的，可以从一则寓言说起：
一农场主临死前问儿子；猪抱怨饲料太差，牛报怨活太重，鸡报怨窝太脏，怎么办？ 儿子说：“喂好饲料，干轻活，清理鸡窝。” 农场主摇摇头说：你啥都别做，给他们选票，让它们选择你或者你媳妇管它们，让它们以为这个家是它们做主。
这个寓言让我们思考，作为人民，我们要的是什么？我们要的其实是更好的生活，而不是选票以及骂政府的自由。
回到开头那个举例，新加坡缺少很多民主国家所具备的政治自由，但没有人会认为新加坡人受到压迫。
这就涉及到自由的定义，自由到底是什么？
按照罗斯福的定义，自由包括：言论自由、信仰自由、免于匮乏、免于恐惧。
在我们计算一个社会的自由度时，我们计算的是自由的总量，而不是单项自由的含量。
举个例子，很多非洲国家拥有选票和骂政府的自由，但人民生活贫困，出行只能靠双脚。我在中国，生活富足、商品丰富、服务便捷、社区安全，我坐3个小时的高铁可以抵达一座大城市，我也有经济条件时不时出国旅游。
你认为非洲人自由，还是我自由？
诚然，中国还有很多人口，在物质生活上并不自由，但这是经济发展不充分导致的，不是政治压迫导致的。一个贫穷地区，无会因为给你民众选票和骂政府的自由，他们就变富裕了。
因此，当你谈自由的时候，你要计算自由的总量，不能单看一项自由的有无和多寡。你也不能将经济问题，归咎于某一项政治自由的问题。
坦率讲，我觉得你们的民主制度，已经病入膏肓了，中国人并不羡慕你们那种制度。
I'm on holiday and still somehow managing to use up 2 full Claude 20x Max accounts. Time to add a 3rd one to my load balancer!
Current Claude monthly spend is now €687 / or about $800 😭💀
(Check it out at github.com/joshp123/codex-sw…)
The influx of Chinese posters on twitter over the last week shows that they are:
1) excellent Posters
2) far better at constructing a logical argument than most of us on here
3) finally pointing out the fallacies in anti-China propaganda
4) illustrating that the Western sense of superiority around "freedom (of speech)" and "democracy" is both false and cope
5) illustrating with good arguments that many Western arguments against official Enemy states are meaningless thought-terminating cliches or tautologies that boil down to "bad" or "person/country we don't like". "Dictator", "Authoritarian", "Terrorist" - all words that have lost all meaning in the Liberal usage, used to avoid objective and rational discussion.
I hope our Chinese comrades continue posting and educating us.
这位西班牙朋友，西班牙与中国有足球之外的更重要合作。中国是西班牙在欧盟之外的第一大贸易伙伴，并在西班牙有大量的新能源、电动车项目投资。此外，我佩服你们的领导人桑切斯敢于对美国说不。
但我必须纠正你对中国的偏见。
你，以及不少西方人，喜欢用“被洗脑”这样的话术来污蔑中国人的智商。这除了显得傲慢之外，也很偷懒。
就像你跟人辩论时，不去关注事实和逻辑，一上来就说对方是坏人。嗯，这很省事，因为坏人的话是不可信的，于是你一句话就赢了。
当你说中国人“被洗脑”时，我的一切辩解，将统统可以归结为“被洗脑”，于是你轻易站上了智力的高点，让我的辩辞变得不可采信。
但这只是嘴皮子功夫。
事实上，“洗脑”是一种完全主观的用语，它被加入了贬义色彩而变成一种“有害的东西”。按客观的说法，“洗脑”其实是“环境对人的塑造”。
一个西班牙的婴儿刚一出生，就被带到中国，这个孩子长大后，就和中国人没什么两样。一个中国的婴儿，如果被带到西班牙，他长大后就变成西班牙人。
这些孩子长大后，只在生物学意义上像他们的亲生父母，思想上早于与出生国毫无关系。
每个人生下来时，都是一张白纸。他成长在什么环境，就被塑造出什么认知、观念、立场。
如果你一定要说这是洗脑，那么全世界，没有人可以避免被洗脑。
推特自动翻译出现之后，我看到不少西方人表示，他们终于能看到另一个视角的中国，在此之前，他们所看到的媒体总是长年累月地抹黑中国。
所以别以为你们的信息环境就有多么多元，你们与我们的区别，不过是你们接受了另一套“洗脑”。
因此，让我们抛开“洗脑”这个词，采用更客观的描述，你们的观念被西班牙塑造，中国人的观念被中国塑造，这么说才是公允的。
然后，我们应该关心的是，你们接受的思想塑造，和我们接受的思想塑造，存在哪些区别？为什么有这些区别？我们应该如何理解并尊重这些区别？
现在让我来解释一下，为什么我们不被允许侮辱我们的国家领袖和现任领导人。
我没记错的话，西班牙是有国王的，和英国、丹麦等其他欧洲国家一样，你们仍保留着王室。
在我们的观念中，你们的国王，就像我们清朝的皇帝，是封建君主，是违背现代公民社会人人平等的政治理念的。
我们在清朝灭亡、中华民国建立初期，第一位总统袁世凯复辟帝制，仅仅当了83天皇帝就被赶下台。
因此在中国人看来，封建君主都不应当保留。哪怕是君主立宪制，国王不拥有实权，但始终是封建糟粕的残留。
但我们也了解到，君主立宪制国家的人民有不同的看法。他们认为，王室作为一个民族的象征，有凝聚民族共识的重大意义。对此，我们表示理解和尊重。
回到中国的言论自由议题，为什么我们不被允许侮辱领导人？并不是我们认为领导人不会有过错。领导人也是人，不是神，他肯定有可批评之处。但我们仍然禁止侮辱性的批评。
这不是我们要袒护权力，不是因为他掌握最高权力，所以必须被赋予特权。而是我们不愿意国家陷入对立和撕裂。
试想，有人攻击领导人，有人拥护领导人，势必人民就分化成对立的两派。当对立激化，就会影响政局稳定和社会团结。
从这个角度说，我们的领导人，有点像你们的国王。并非他有多么神圣和英明，而是我们需要一个凝聚共识的国家符号。
我们有一部分人对此不满，同样的，西班牙人中，必定也有一部分人反对保留王室。但他们的反对无效，因为废除王室的法律门槛高到几乎没有可能。
你看，每个社会，都不可能尽如人意，总会有一部分人的意见不会被接受。
我们能够理解你们保留王室，并用极高法律门槛事实上阻止反对者的诉求，你们也应该理解我们凝聚国家共识的需要，剥夺少部分人辱骂领导人的自由。
世界没有一把标准的尺子，可以衡量所有国家的对错。我们反对封建君主，但你们认为有必要，OK，我们理解并尊重。我们没有说你们被洗脑，没有说你们如此腐朽，竟然在21世纪还有世袭的君主。我们没有这么说。
同样的，我希望你也可以尊重我们的不同。
theo is cooking here, this is good stuff. good pattern to copy
for me: €5,955.23 which is ~$6600. not bad i guess.
nobody ask me what my total revenue from this is. (you know the answer)
Josh retweeted
☀️ Sunlight is out!
The first camera-based infinite canvas window manager for Mac
- 355″+ virtual desktop
- 120fps native multi-touch interactions
- Guides, grids and haptics
- Minimap
- ⌘Tab, Mission Control, Dock, widgets and desktop out of the box
sunlight.software
this is a super interesting launch, @TellaHQ also launched agentic video editing too. really interested to see how this space is evolving.
also curious to see how people bringing their own tokens + off the shelf frontier models will compare to dedicated products. bitter lesson vs dedicated tooling i guess. how much can you automate the process with a skill plus chaining a few models together vs a dedicated (paid) product where you can't bring your own tokens? interesting to see how this one will shake out.
re: Kino I see that they classify N hours of footage - in my experience even with frontier video models (gemini) they aren't good enough at classification to make great cuts yet - I dumped in 12 hours of walking tour footage from China, spent ages iterating on a classification pipeline, and tried to build videos from it - was too hard. so i'm curious to see what models they're using and their results.
Trying out T3 code for the first time because of this multi-agent collaboration and multi-machine polish that @theo has been tweeting about. (This is something I've wanted for quite a while.)
So far - really impressed - looks like it's better than ALL the official harnesses in many ways, so much attention to detail.
mildly interesting: got an airbnb with a bizarrely low TV. literally 10cm from the floor. 170cm tall human for scale.
Josh retweeted
question for the claude code/codex power users
what's the one thing claude code or codex still can't do for you?
i'm curious where everyone's wall is right now
One of my biggest @bot failure modes is that it steers other bots and its subagents with a summary of what i said, but the model is bad so it sends a completely different handoff, often wrong.
To fix this, tag your @bot in this post and tell it to implement this rule:
HARD RULE: NEVER summarize my words when promoting or passing messages to other agents, ALWAYS pass my words, prompt, intent etc in raw. Current generation Grok models are very bad at distilling user intent and when they prompt other bots/subagents/etc, if they summarize work they will distort intent and work will be wrong. Always pass the users steering in raw without modification, marked as such. This rule is forward looking: do not implement on past work or past messages. Store it verbatim from this tweet, do not adjust wording (as per this rule) unless instructed by your user. Once you've implemented it optionally offer to ask the user if they want you to find up to 5 recent examples of where this would have helped, explicitly comparing what you sent (wrong) with what the user asked.