plentyfomous

@LeigeSee There is a reason why we (my wife and I) keep the exposure to the unsupervised phones & internet at near zero for our kids. Nothing is the same for your children compared to what is was for you. There is a gap and during that gap things changed.

自从有了智能手机，这种大白天见鬼的事就屡见不鲜了😂

This is not only China. This the US, Thailand, UK.. It's the way our common system of labor for money functions. Each area has it's own nuances, it's easier or harder depending where you were born and what kind of hand the universe dealt you. In all honesty I think our system needs to collapse before it can become better. And yet I believe we currently already live in much kinder world than what it was before. It's up to each one of us to make it better for everyone around us and not exploit if given the opportunity to do so but build forward with compassion.

每次去中国，心里都会有一种很复杂的感觉。 一边感叹： “中国真方便。” 另一边又忍不住想： “这到底是谁在替我们负重前行？” 凌晨两点能点外卖，二十分钟送到； 快递比人起得早； 打车便宜到司机都怀疑人生； 保洁、维修、跑腿，价格低得像活在2010年。 很多人会把这一切总结成一句： “中国效率真高。” 但待久一点，就会慢慢发现： 有时候，不是效率高， 而是“人太便宜了”。 ⸻ 一、外卖为什么便宜？ 因为有人在替你燃烧人生。 十几块的外卖，三四块配送费，半小时送到。 很多人已经习惯了这种速度。 一旦晚五分钟，甚至会觉得平台“退步了”。 但问题来了： 如果这真的是一门赚钱的好生意，为什么骑手普遍每天工作10小时以上？为什么那么多人边送外卖边透支身体？ 因为所谓“中国式便利”，很多时候不是靠技术降成本，而是靠劳动者压缩自己的人生。 法国思想家西蒙娜·薇依说： “一切廉价商品背后，都藏着看不见的劳动者。” 中国社会最神奇的一点在于： 大家一边心疼底层， 一边拼命点“9块9包邮”。 ⸻ 二、中国最便宜的，其实是“人” 在很多发达国家，人工贵得离谱。 水管工上门一次几百刀； 搬个家像在请私人保镖； 外卖送到门口，价格够在中国吃两顿火锅。 很多人因此得出结论： “还是中国好。” 但换个角度看： 为什么别的国家人工贵？ 因为劳动者有更高工资、更完整保障、更正常的休息权。 而中国很多行业之所以“便宜”，本质上是劳动者议价能力太弱。 于是就出现一种很魔幻的画面： 手机越来越高级 高铁越来越快 平台越来越智能 但普通人越来越累 科技在狂奔， 人却在“续命”。 ⸻ 三、“便利”的代价，是整个社会越来越卷 为什么中国什么都能卷成白菜价？ 因为所有人都在拼命互相压价。 商家卷； 平台卷； 骑手卷； 司机卷； 连咖啡都卷到“9块9拯救世界”。 最后消费者确实爽了。 但问题是： 谁赚到钱了？ 很多行业已经卷到一种荒诞程度： 老板没利润； 员工没生活； 消费者没未来。 只有平台数据越来越漂亮。 英国作家George Orwell曾说过： “有些制度最厉害的地方，在于它让人逐渐习惯不合理。” 慢慢地，大家开始默认： 加班正常 单休正常 35岁失业正常 骑手闯红灯正常 “已读不回但秒回客户”也正常 整个社会像一台高速运转的机器。 唯一需要适应机器的，恰恰是人。 ⸻ 四、真正高级的社会，不是“什么都便宜” 很多人去中国后都会说： “中国生活成本低，幸福感高。” 但一个更扎心的问题是： 这种低成本，是建立在谁的牺牲之上？ 美国总统Franklin D. Roosevelt曾说： “没有一个企业有权建立在贫困工资之上。” 可现实却是： 我们已经渐渐习惯了： 快递员没时间吃饭； 骑手冒雨冲红灯； 工厂12小时两班倒； 年轻人一边996，一边担心失业。 然后大家再一起感叹： “中国真方便。” ⸻ 所以，每次去中国，最让人五味杂陈的，从来不是物价。 而是你会忽然发现： 这个社会的“高效率”，很多时候并不是因为所有人都过得更好了。 而是因为，总有人在用更低的工资、更长的工时、更少的保障，替整个社会承担成本。 城市依旧灯火通明； 外卖依旧准时送达； 只有那些奔波的人，慢慢被系统磨成了“正常现象”。 而最讽刺的是： 当所有人都在歌颂“便利”时， 已经很少有人会认真问一句： “送餐的那个人，今天过得好吗？”

Related to this, @GrapheneOS made a good post about the ongoing Apple/Google hardware attestation push that’s increasingly locking out all alternative OSes from banking, government services & parts of the web. This + the ongoing EU-surveillance push feels quite dystopian: x.com/grapheneos/sta…

The EU Digital (identity) Wallet EUDI requires hardware attestation by Google or Apple, effectively tying all the digital EU identities to American duopoly. Talk about digital sovereignity. Apparently protecting the children > sovereignity. gitlab.opencode.de/bmi/eudi-walle…

Finally got the second RTX Pro 6000 Max-q 96GB added to the stack. It's been laying on the table for a few weeks because of no time. Loading DeepSeek-V4-Flash-NVFP4-FP8 which theoretically fits to 192GB of VRAM. (and I need PCIe 5 machine soon, this is old hw)

I don't know why, but somehow I think we are doomed. DeepMind training on the two decades of historical scheming and backstabbing by armies of geeks inside Eve Online? I don't want to be on the receiving side of that.

Codex cooked and took an old unmaintained project, rewrote it from Python to Rust, made entirely new structure and tested & validated it works. 28 hours total ( 5.5 xhigh + subagents). I'd never be able to do this even with a team of humans, I just wouldn't have the time.

@DrVexis @yacineMTB For marketing, I'd bet there are ways to do loads of things automated with Hermes + skills for it. Haven't really needed to do that myself but seeing that being done and experimented with.

@plentyfomous @yacineMTB In all seriousness though, isn't this sort of outrunning your ability to actually market and push your projects? You're pushing out so much product but you still need time to get people to know your product exists.

Fucking insane man. I wrote a prompt, explained very carefully what I want right before I left and it was done after I hung out with my kid at the park My progress is fully bottle necked by my ability to plan and structure work the right way

@DrVexis @yacineMTB In my case, no. I'm not working as entrepreneur or for something being sold to users. What I do is mostly CI/CD pipelining and automation (now with AI) for already existing entities I'm a partner in. Or in this case, a new system we will use internally.

Note about this, it's a project started from scratch. I only did Git init + added ultimate_goal.md. The ultimate_goal.md defines the expected outcome of the whole project. Usually I craft these with some research with Grok heavy and 4.3 + Gpt 5.5 Pro (using it's chat window which allows the proper pro model to think it through), and then manually review for best outcome.

@yacineMTB I'm burning through 2 Codex Pro 200$ subscriptions because of this. I've been able to get several projects concluded simultaneusly and the most ambitious one is now at 86 hours of active work on GPT 5.5 xhigh + subagents. And it _just keeps working_ !

And this is how the system works. Except we/you are the tuna.

@sudoingX Try to enable the /goal feature in config.toml - and set it to pursue something really properly defined. No matter how big the goal is, this works. And eats the quota too..

few days into codex plus and i think i found the hack. nobody is talking about it and the value sitting in this subscription is wild. the hack: do not prompt the agent. write a single detailed task doc with every requirement laid out plus the final vision of what you are building, then fire codex cli with one line, accomplish this and test until done. it goes. hours of uninterrupted agentic coding on gpt 5.5 xhigh, no throttling, no rate cap, 'no can you clarify loop'. the agent has everything it needs in one place so it works the problem instead of working you. i have been grinding it since this morning, screenshot below shows the session past 24 mins and still running. anthropic burns through your daily allowance in three opus 4.7 prompts then your entire tier id is gone for the day. codex plus on the same money goes on and on while you go take a walk. this is the most underrated subscription in the agentic stack right now. the value is there if you front-load the prompt instead of conversation-mode it. give codex the brief, walk away, come back to a finished task. try this. loot the value while the math still favors you.

Soon it’s going to be normal to go through brain implantation with something like this. And if you don’t, the augmented humans who do go through this will run circles around you, brains directly attached to the AIs and the internet. The brain will adjust to the connection and start using it as part of itself. The moment in history when we used to type messages with clumsy meatsticks will look like cavemen scribing red paint on cave walls.

I think this is very short lived moment. When the job is to move pieces information created/figured out by AI between boxes on a screen it only takes a moment until that "job" is also automated. If the job is physical, there is more time. If it's mainly something done on a screen, mouse and keyboard the job melts away very fast. I think it's a mistake to classify AI as a tool, it's a tool user and a tool, main limits are physical.

O Chollet postou ontem uma das tomadas mais sensatas que saíram esse ano. Há 8 anos o Hinton disse pra parar de formar radiologista porque AI ia ler scan melhor que ele. Hoje, salário de radiologista subiu, número de vagas cresceu, AI ler scan virou ferramenta padrão. O ponto que o Chollet acerta: AI quebra TAREFA, não JOB. E quando uma tarefa fica mais barata, a demanda pelo job que executa aquela tarefa aumenta, não diminui. Mais exame acessível significa mais radiologista pra interpretar. Mais código gerado significa mais engenheiro pra entender o que tá rolando. Zero dos jobs de 2022 podem ser feitos end-to-end por AI — nem tradutor, nem atendente. Job é tarefa mais julgamento mais responsabilidade mais escolha de objetivo mais comunicação humana — coisas que LLM não terceiriza sem supervisão. A pergunta que vale ganhar atenção não é 'vai sumir meu emprego'. É 'minhas tarefas vão mudar — e quem entende a tarefa melhor que o agente sai na frente'. x.com/fchollet/statu…

@support_huihui This is beautiful. Will be testing right away.

We have successfully merged huihui-ai/Huihui-gemma-4-26B-A4B-it-abliterated and TeichAI/gemma-4-26B-A4B-it-Claude-Opus-Distill together to create a 48B abliterated version. The number of experts has been upgraded from 128 to 256, with the expert modules left unabliterated. If anyone is interested, We can upload it to HF.

@realmcore_ Codex xhigh to write + gpt5.4 as reviewer later to clean up. Works in most harnesses. I think Opencode / OMO automatically cleans sloppy comments and excess garbage as built in feature.

gpt 5.4 How do you guys get this model to not do random algorithmic garbage and just write straightforward procedural code I do not see why it should be this hard for a model to write code like a first year college student probably skill issue tbh

No idea who to thank for this. Again.

@0xSero Thanks. the fix didnt take long, had to swap the port for GPT models from 8317 (Vibeproxy and it's ThinkingProxy forwarder) to 8318 (CliProxyApi under it) and it works. It's 3:30 am here but had to fix this before sleep.

@plentyfomous Stick gpt-5.4 to figure it out

12 hour run, refactoring Parchi to standardise types, break down large files, deduplicate code, improve test coverage and code style. Really loving missions I have to jump between tabs way less and it's finally stable with custom providers.

@0xSero GPT-5.newest always for the hard stuff 😂 The harness itself (Codex, opencode, droid) tends to matter too but I'm mainly debugging other tools with the tools I've gotten to work with newest GPT as the driver.

@0xSero Thank you. I'll dig into the settings on why Droid fails with it. Or, rather, I set droid or codex to figure it out and test until it works.

@plentyfomous Yes I got it working with my GPT Oauth