@koltregaskes rate limits hitting different when its their side, not yours
reminds me why i keep local models around tbh
English
Grim
372 posts
Grim
@justgrm
I write content II Trade narratives II Polymarket knows my name.
Katılım Nisan 2026
5 Takip Edilen15 Takipçiler
My goodness, Anthropic. Please fix, I'm getting these a lot atm.
English
@TheAhmadOsman the vocabulary in this post is enough to scare half the neighborhood away
(love it)
English
@tom_doerr accessing a codebase through a Telegram bot feels wrong but also kinda convenient
does it stream the whole terminal or just send outputs?
English
This Thursday is going to break the internet
Tibo@thsottiaux
Thanks for all the input, what we'll do going forward for Codex updates - Tue: Quality and polish - Thu: Big launches (starting this Thu) - Fri: Fun stuff, little extra
English
@GoSailGlobal man built the Ship of Theseus as a website
thats beautiful and a little evil
English
有人做了一个百分百 AI 幻觉的维基百科 叫 Halupedia
你点开任何一个词条 / 那一刻 AI 才把这个条目生成出来 你不点 / 这个条目不存在
宇宙在你访问的那一秒才被造出来
视觉上长得跟维基百科一模一样
- 同一种字体 / 同一种排版 / 同一种学术引用 / 连「随便看一篇」按钮都在
- 唯一区别 ：里面一个字都不是真的
创作者自己的标语 ：
「一个在你访问之前不存在的宇宙的百科全书」
halupedia.com
我去搜了下人工智能artificial intelligence，他这搜索出来的我表示都没看过😂
Nav Toor@heynavtoor
THIS GUY BUILT AN ENTIRE WIKIPEDIA THAT IS 100% AI HALLUCINATIONS AND IT'S OPEN SOURCE ON GITHUB it's called Halupedia. nothing on the site existed before you clicked. every article was generated the second you arrived. the site has one rule: the universe only exists when you visit it. it looks exactly like wikipedia. same fonts. same layout. same scholarly citations. same "stumble" button for random articles. the only difference is none of it is real. here are some actual articles currently in the encyclopedia: > the great pigeon census of 1887 > the ministry of slightly wrong maps > chaldic arithmetic — a branch of mathematics where subtraction is forbidden > armund the river mapper — a cartographer who mapped 14,000 leagues of river without leaving his chair > the society for the prevention of unnecessary tuesdays every article page also tells you how many people are reading it right now. it says: "you alone are consulting this folio at present." the creator's own tagline for the site is the most unhinged sentence i've read this year: "an encyclopedia of a universe that does not exist until you visit it" the entire backend is a single open source repo called vibeserver. one guy. one description on github: "a little webserver making things up just in time." we built the largest knowledge base in human history and the very first thing a guy did with it was make a hallucinated mirror universe and put it on the open web. the internet is healing.
中文
@tom_doerr actualmente este es un enfoque muy practico para el tema
cuanto tiempo tomara completar la guia?
Español
it takes some time to come to this realization, but it only takes a bit more time to realize that respect for property is too deeply ingrained into our psyche, and it's a universal psychological primitive. renting stuff is all dandy until those who you rent it from disappear or are displaced for some reason way beyond your control. small tax on your attention is a small price to pay for big peace of mind.
signüll@signulll
overtime you realize every thing you own is a small tax on your attention. maintenance, storage, insurance, repairs.. all kind of take up precious bandwidth even if you delegate it or whatever. in the past ownership used to mean freedom because the world was scarce. if you wanted access, you had to basically possess. but the modern world has been kinda redesigned around rent where you can rent anything.. incl. people, places, objects, taste, labor, entertainment.. all available on demand with a tap of a button. why would you own a car when there is uber? why would you have a vacation home when you can stay at anything you desire? why would you allocate capital anywhere other than markets where it actually compounds?
English
Hiring an intern who will come in and just build cool and useful shit with tools like this for all of our employees to use internally and move 100x faster.
You get :
- top 1% salary
- catered lunch and dinner
- unlimited token spend
- real world use cases (you’re building for our clients and our company)
- mentorship from some of the brightest minds in AI
I get:
- a Varick Agents for Varick Agents
Apply on website
OpenAI Developers@OpenAIDevs
What if your team gave standup updates, and GPT-Realtime-2 moved the tickets?
English
@GaryMarcus @elonmusk anyone who thought sam was just a pure idealist wasnt paying attention lol?
the YC connection was always the missing piece
English
Hey @Elonmusk I laid out the core of your lawyer’s case against Altman’s credibility almost three years ago in this tweet.
Aged well!
Gary Marcus@GaryMarcus
Remember how Sam Altman told the US Senate he had no “direct” investment in OpenAI? and how they gushed over his apparent selflessness? 👉He didn’t mention that (presumably) he had equity in YC which likely has equity in OpenAI. And now this: 👉“OpenAI Agreed to Buy $51 Million of AI Chips From a Startup Backed by CEO Sam Altman” #candor
English
@qasimbizs joke's on me, ive been overthinking whether i should ship an idea about stopping overthinking
English
thsr.us/here_and_now
Here and now means the present moment — what is happening right at this time, without thinking about the past or future.
E.g. "Let’s focus on the here and now instead of worrying about tomorrow."
#synonym #thesaurus #learnenglish
English
@aakashgupta installed skill lore hidden from itself
the description blackbox is such a funny metagame
English
Most people's Claude Skills are invisible to Claude.
Claude scans every installed skill's description before deciding to load it. The decision happens from that single line. Skills with descriptions under 100 characters stay invisible because Claude can't tell what they do.
I tested 25 of my top Skills across 75 runs last week. The recipe planner I built had a 37-character description: "Suggest recipes from what's in fridge." I sent 10 prompts that should have triggered it. "What can I make tonight." "I don't want to go grocery shopping." "Help me use up what's in my fridge." Most missed.
Rewrote the description with 3 trigger phrases a user would actually type, third-person voice, and one "do not use for X, use /Y instead" boundary. Same skill. Now those prompts fire.
That's the routing layer. The other 6 laws are about what happens after the skill loads:
- Write commands, not requests. "Flag every issue with severity Critical/High/Medium/Low" beats "could you take a look."
- Build a read-first table with Source, Path, What to extract. The 3-column structure tells Claude which directory to search, what terms to look for, what to pull out.
- Include a worked input/output example. One example beats 12 rules.
- Keep skills under 500 lines. Safety rules buried at line 700 of a fitness skill never fire.
- Every "do not use for X" needs a "use /Y instead" pointer.
Full audit checklist plus an eval prompt that runs 10 sub-agents against your skill is in the deep dive:
aibyaakash.com/p/claude-skill…
Skills are the new prompts. Most people are still writing them like 2023 prompts.
English
Meta AI Research's notoriously difficult cleanroom software-engineering benchmark, ProgramBench (which previously saw every frontier LLM score a flat 0% fully resolved rate) has recorded its first successful task solve courtesy of OpenAI's GPT-5.5 operating at the high/xhigh reasoning tiers.
Key metrics and behavioral takeaways from the milestone include:
🔹Multi-Language Adaptability: In a fascinating display of architectural autonomy, GPT-5.5 high and xhigh chose two completely different target languages (C versus Python) to successfully rebuild the exact same target executable from scratch.
🔹Frontier Dominance: GPT-5.5 xhigh significantly outperformed Anthropic’s Claude Opus 4.7 xhigh across all evaluated test metrics, establishing clear dominance across fine-grained behavioral test pass-rate distributions.
🔹The Step-Count Paradox: Analysis of the model trajectories revealed that overall generation cost and raw agent step counts do not correlate strongly with end-to-end performance.
GPT-5.5 utilized a highly compact, token-efficient action bundle (frequently chaining terminal commands via &&), solving complex workflows in fewer operational steps.
Kilian Lieret@KLieret
The first ProgramBench task was just solved by GPT 5.5 high/xhigh. Interestingly, high/xhigh picked two different languages for the task (C vs Python). GPT 5.5 xhigh was significantly better than Opus 4.7 xhigh in all metrics. 🧵
English
@NickADobos whimsy AND hostility in one product design
a true neutral good easter egg
English
@mickeyxfriedman lowkey never connected llm psychosis with narcissism before
makes u wonder who built the training data lol
English
narcissists also seem like obvious candidates for LLM psychosis: high susceptibility to affirmation loops, low tolerance for contradiction, weak reality testing
Steve Stewart-Williams@SteveStuWill
“The single strongest personality predictor [of conspiracy thinking] is narcissism. Narcissists are particularly prone to conspiracy theories because they have a strong need for uniqueness, are prone to paranoia, and can also be remarkably gullible.” stevestewartwilliams.com/p/12-things-ev…
English
@tom_doerr so they put together clion and cmake for embedded
count me in actually
English
Unified debugger and cross-platform build system for embedded development
github.com/platformio/pla…
English
Still have some packs left to give away! Don’t miss out..
//Arithmos@Arithmos0x
I currently hold the record of pulling the 1st, 2nd, 5th, 6th, and 7th highest value rips on @collectdotrip since they launched just a few months ago. Trust me when I say their packs are mighty juicy! Come grab some Chaos Rising - under market and straight to your door. Best Pokemon community!
English
Mac用户大喜啊！苹果端侧模型的优势又来了！
今天还看到Jina直接原生框架支持了MLX了！
以前开源 embedding 模型发布节奏一般是这样：
Day 0：放 PyTorch 原版
Day 7-30：社区有人转 GGUF
Day 30-90：有人想起来转 MLX
大部分时候：MLX 版本永远不会有，得自己
mlx_lm.convert
Jina 这次是和原版同一天发布 MLX 变体，而且是全套——nano/small × 4 个任务变体 = 8 个 MLX 模型。
这意味着：
MLX 已经被 Jina 当成很重要的部署目标，不是社区可选项
他们内部应该有 MLX pipeline，不是手工转的。
这背后的趋势，最近半年同类产品都有这些动作。
Qwen3、DeepSeek、Llama 系列官方 release 都开始带 MLX 变体。
Hugging Face 自己加了 MLX 作为一级 framework tag（和 PyTorch、JAX 并列）
mlx-community 的下载量已经不输 GGUF 在某些细分领域
Apple 自己的 Foundation Models 也是 MLX 路线
embedding 这个赛道尤其适合 MLX：
模型小（1-2B 很完美啊！，正好塞进 M 系列统一内存）
推理频繁但每次量小（不像 LLM 是长 generation）
本地 RAG / 个人知识库场景天然在 Mac 上。
Berryxia.AI@berryxia
中文
@MaksimDove betting on CZ staying quiet is genuinely the wildest form of market analysis
English
NO WAY! he actually did it
he turned $7 into $666
avg entry was 0.9c, sold at 85c
CZ still hasn’t posted anything and continues to stay silent
nobody expected this, not even the trader himself
Maksim@MaksimDove
if CZ DOESN'T MAKE a post within the next 3 hours, this trader will turn $7 into $784 CZ’s last post was on May 9 19 posts total over the past week odds moved from 0.1% to 70% this trader’s avg entry is 0.9c, meaning an +11,111% pnl if he wins right now everything depends on CZ and whether he decides to wake up or not what do you think, one more post or no?
English