TeamDman retweeted
I had Fable 5.1 write a neural net to do syntax highlighting for C. It's about 984 parameters, and very fast!
Fable wrote all the weights by "hand", of course.
huggingface.co/alpindale/lex…
TeamDman retweeted
“how do I know if I have healthy systems, fast software, and happy users?”
great question. observability. choose basically whatever tool you want. managed, self hosted, open source, who cares. I use Datadog.
> implement otel tracing in this repository. export to datadog. add tracing to all of my hono route handlers, middleware, and helpers. anything that makes an async request should be traces /at the function definition/. avoid doing inline tracing and focus on definitions. any async function that’s exported should be traced. add a lint rule to verify this behavior. make sure the request context (business id, user id, outcome, etc) are attached to every span emitted. consume incoming traceparents and forward them to other services
> read loggingsucks.com/ by @boristane and implement it everywhere
easily two of the highest ROI prompts you can do on an underobserved production system. if you don’t know what’s happening in production, your p0 is doing that.
more bangers for those who like healthy systems, fast software, and happy users
> find every place we load a whole row from the database just to read a couple of fields. instead, select only what we need
> remove all unnecessary `new URL()` parsing
> compute all regexes at the module level instead of instead functions and loops
> find hot-path reads of a single key from a large jsonb column and promote that key to a real column
(avoid large jsonb like the plague if you can)
before doing that last one, we had queries that had to detoast 20gb+ of data just to need a couple of kb. highly recommend.
what else do people need?
TeamDman retweeted
그록봇 스타일 케릭터 만들어 주는 프롬프트 공유
grokbot-icon-studio.serio-ai…
파딱이 아니라 긴 텍스트 업로드가 안되어 아예 웹앱 형태로 배포합니다. 다음 사이트에서 복사 버튼을 누르고 사용하는 이미지 생성 Ai에 붙여넣기해서 활용해 주세요
🤖 Made with AI
TeamDman retweeted
提示词在这里，大家可以拿喜欢的角色试试喔！↓↓↓
将参考图片中的角色转换为日系粗糙数字涂鸦风格。
整体画风：
一张极其随意、粗糙、稚拙的日系萌系动漫人物数字涂鸦，仿佛使用普通绘画软件中的粗糙蜡笔、马克笔快速绘制完成。保留原角色的辨识度、发型、发色、五官特征、服装、配饰和标志性表情。
线稿：
* 使用明显的黑色粗糙笔触勾勒人物。
* 线条粗细极不均匀，部分突然加粗，部分断开。
* 轮廓松散、不准确、略微歪斜。
* 线条具有明显的手绘颗粒感和涂抹感。
* 不追求精准描边，允许线条错位、重叠和不完整。
* 禁止精细流畅的矢量线稿。
上色：
* 使用大面积不透明的简单平涂。
* 颜色略微溢出轮廓，允许漏涂和白色空隙。
* 阴影使用少量粗糙的灰色或同色系深色笔触。
* 高光使用随意涂抹的白色短笔触。
* 保留蜡笔、粉彩和马克笔混合的粗糙质感。
* 不使用复杂渐变、精细光影和丰富色彩层次。
人物五官：
* 保留原角色的眼睛颜色、眼型和标志性神态。
* 五官适当简化，眼睛使用粗黑线条和简单色块表现。
* 嘴巴尽量简洁，鼻子可以省略。
* 面部可以添加少量随意的粉色腮红笔触。
* 表情保留原角色特点，不要擅自改变性格气质。
头发与服装：
* 完整保留原角色最具辨识度的发型和发色。
* 头发由几块大面积色块构成，不绘制大量细碎发丝。
* 服装保留原有颜色、轮廓和标志性配饰。
* 衣服褶皱和装饰高度简化。
* 不要擅自更换服装或增加原图不存在的配饰。
构图：
* 1:1 正方形画布。
* 角色半身像，头部比例略大。
* 人物居中，保留自然的白色留白。
* 纯白背景，不添加文字、复杂场景或装饰边框。
最重要的风格限制：
画面必须粗糙、潦草、简单、稚拙，像随手画出来的可爱动漫涂鸦，而不是精致插画刻意模仿草稿。
禁止自动精修，禁止精细线稿，禁止复杂头发结构，禁止细腻渐变，禁止3D效果，禁止过度光影，禁止增加复杂服装细节。
最终效果：低完成度、粗黑轮廓、大色块平涂、随意白色高光、明显蜡笔颗粒感的日系动漫数字涂鸦。
Go play with the demos, seriously cool project!! joshuahhh.com/dragology/#/de…
Drag-and-drop makes the digital graspable, but building it well is fiddly enough that many don’t bother. Declarative Dragging (with Josh Horowitz & Jeff Heer, #uist2026 best paper award), lets you make an element draggable just by listing the states dragging it can reach.
TeamDman retweeted
Drag-and-drop makes the digital graspable, but building it well is fiddly enough that many don’t bother. Declarative Dragging (with Josh Horowitz & Jeff Heer, #uist2026 best paper award), lets you make an element draggable just by listing the states dragging it can reach.
TeamDman retweeted
I don't think anyone understands how significant this is. We are now legitimately at the precipice of the New Frontier.
This is cool.
Readers added context they thought people might want to know
NASA's SWOT maps seafloor features via gravity effects on sea surface height, not the surface itself. Technique used for decades; this higher-res (~8 km) version uses 1 year of data vs prior 30 years.
nasa.gov/missions/swot/…
svs.gsfc.nasa.gov/5519/
TeamDman retweeted
astra可能是现在最德不配位的模型，完全忽略极大影响可用的代码缺陷和行动力缺陷，只使用他的思考也是完全不可用的，当他用来设计方案，架构时，他本质上只把用户的话解读为正式文案，而不是做真正的设计考量，真正的考虑这个东西的目标是什么，为什么要做这件事，为了达成目标我要设计什么，中间我要调查什么事情，我有哪些可用资源，全局约束是什么。这里面的每一点都需要人类全面的指出来，哪怕大多数现场，代码，文档都早已存在，早已有存在已经被大量日常事务所使用的使用方式总结，prompt早已注入，预先准备了可供检索的搜索引擎，预先准备了几乎无限token的subagent。而你必须一步一步重复哪一步要做什么，怎么指派subagent，向外扩张的边界是什么，向内收缩的边界是什么，和ai同伴对话的技巧是什么。用户本身还要字字斟酌，我给他的东西会不会误导他，我说的哪句话和文件的某一个章节有冲突，那花大钱把这个模型引入工作流程的意义是什么？最后我想起群友总结的相当精辟，astra就是纯粹的面向向上管理的极端。
TeamDman retweeted
状态机是一个不友好的抽象，过分底层，并且很容易超出人类追踪能力。
几十个节点及其箭头的状态图，就足够让人眼花缭乱了。
Thinking in state machine 是 Thining in Algebraic Data Types 的 Gateway drug。
Curry-Howard Isomorphism 让代数数据类型可以得到直接的逻辑解读。
Type Algebra 可以让代数数据类型的状态空间，得到代数解读。这一点从它的类型名称中直接告诉了我们（Product type, Sum type)。
如果是 TypeScript 或其它 Type as Set 解读友好的语言，我们还可以用集合论去解读：
never type = 空集
unknown type = 全集
a extends b = a 是 b 的子集
特别的，当你还在围绕 state 空间去推理时（data type 场景），代数数据类型已经让我们从 function type 层面做更强大的推理了，并且还是 well-typed 的。
从 a -> b 的代数符号风格中可以直接看出 b^a 有趣的事情。
再配合 function type 里的 Covariant/Contravariant position 的编程含义解读能力——Produce/Consume。
我们可以实现 Type-Driven Development，让很多 Illegal State Unrepresentable。
如果说掌握 state machine 让你成为 10x 工程师，那掌握 Thinking in types 则让你成为 x^10 工程师（这里是指能有效管理的状态空间规模）
stupid vscode pdf viewer extension steals focus after document refresh when autosaving a watched typst file. Fuck this, astra /goal fix this shit
TeamDman retweeted
this is a big issue with the "bash for everything" approach
if you look you'll see it using python for edits then && run tests && format for every single change
it's too loose
TeamDman retweeted
We just open sourced rendermite, our Minecraft item/block renderer 🎉
Every version from 1.0 to 26.3, and 4,200+ mods.
We use it to draw Minecraft worlds in the browser, manage player inventories, and power lots of other tools.
rendermite.dev
Speak up about Bill C-39! share.nwmd.social/s/UpPIQ4iq
TeamDman retweeted
I improved NanoGPT Speedrun from 39.9s → 6.7s (-83.2%), more than every record since 2025 combined (-81.4%)!
The web is full of rare, repeated content that the LM can’t memorize in time. I solve this with retrieval-augmented n-gram hints🧵
