Mai 麦尔彦

Physics BA | Mandarin Interpreter ➔ AI Builder 🇨🇳🇺🇸

Building & AI builds, Python & localization 🛠️

[Mai 麦尔彦 @Mai_Builds]

I'm documenting my path to AI Engineering in public 🛠️

🔹 Building: @SimplyLesson & https://t.co/UGYcayGPfP

🔹 Learning: CodePath AI 301, Python, LLM workflows

🔹 Sharing: AI build logs, technical breakdowns, & localization insights"

[Mai 麦尔彦 @Mai_Builds]

@jerryjliu0 Last year's best humanoid long jump was 1.25m, this year's winner hit 7.97m.
That's a 6x jump in one year, form critique aside, the trajectory is the real story.

[Mai 麦尔彦 @Mai_Builds]

@jerryjliu0 Second this, "agent native" mostly means better apis and structured data, not the app itself gaining agency. The confusion is people hearing "harness" and picturing the software becoming an agent instead of just being legible to one.

[Mai 麦尔彦 @Mai_Builds]

This book's examples work no matter which of the three you use or prefer. That's the part worth flagging as unusual, not the multi-framework support itself (that's just Keras 3), but a textbook choosing to write its examples in the framework-agnostic layer instead of locking readers into one ecosystem.

[Mai 麦尔彦 @Mai_Builds]

@MansiCodez "exactly-once" on that diagram is doing a lot of work. most real messaging systems, whatsapp included, use at-least-once delivery plus client side dedup on message ids. True exactly-once across a distributed broker is basically a myth.

[Mai 麦尔彦 @Mai_Builds]

@foundmyfitness Study detected 3,141 chemical features but only identified 303 of them as specific chemicals
nontargeted GC×GC-HRMS flags way more unknown signals than it can actually name.

[Mai 麦尔彦 @Mai_Builds]

@emollick I think reviewers pattern matching on "ideas rhyming" also catches generic writing from humans who all read the same 3 blog posts, correlation with ai use isn't proof of it.

[Mai 麦尔彦 @Mai_Builds]

This is genuinely useful for experimenting across agents without juggling separate accounts and keys. However "same task" is doing a lot of work here. cost gaps that big usually mean different retry counts or tool calls, not just model pricing. success rate next to the price tag would make this actually convincing.

[Mai 麦尔彦 @Mai_Builds]

Normally connecting a bot to Slack means creating a Slack app manifest by hand, configuring OAuth redirect URLs, and pasting bot tokens into your code. LangChain's tool automates all of that behind one CLI command, so the agent just shows up in your Slack workspace ready to go.
It is convenient until you need custom scopes or multiple workspaces managed token handling is great for demos, less great when you need to know exactly what your agent can touch.

[Christian Bromann @bromann]

Starting today, a @LangChain's Managed Deep Agent deploy provisions your Slack app for you 🤯

No manifest. No OAuth redirects. No bot tokens to copy around.

One command, and your agent says 👋 in Slack.

[Mai 麦尔彦 @Mai_Builds]

0% refusal is the tell, not the flex.
That's usually a sign it'll happily walk you through anything, benchmarks rarely capture what got traded away to hit that number.

[Jun Song @jun_song]

SuperQwen3.8-27b-abliterated is now live and open for everyone 🚀

- Uncensored for freedom
- Broken weight from abliteration fixed with agents swarm
- fixed overthinking problem
- 1M context, Multimodal

Sorry about the late release.

Super-tune takes much more time than normal abliterated models, because fixing weights and eval takes about a week.

BF16/MLX/NVFP4 ⬇️

[Mai 麦尔彦 @Mai_Builds]

@jeffreyhuber The harder problem isn't generating embeddings, it's keeping them fresh as source data changes by the hour. A stale context layer is worse than no context layer.

[Mai 麦尔彦 @Mai_Builds]

@prasenx Wait the sound is procedural too? Generated audio in a canyon scene is the part that actually gets me.

[Mai 麦尔彦 @Mai_Builds]

Spent today learning to code by reading a real merged PR line by line instead of doing a tutorial.
Made me actually understand what each piece does, not just recognize syntax.
Starting a running notes file after every PR I read from now on.
Today's lesson: never let local dev code accidentally point at real production data.
The fix was simple. fake the external service locally, then add a check that refuses to start if the config looks like prod.
Obvious once you see it, easy to miss before you do.

[Mai 麦尔彦 @Mai_Builds]

@sundarpichai @GeminiApp Flash isn't the top scorer, GPT-5.6 Sol and Claude Fable 5 Max both clear 90%. What it actually owns is the cost curve: nothing else is hitting mid-80s accuracy anywhere near $0.25 a task. That gap is probably the real growth driver, not raw capability.

[Mai 麦尔彦 @Mai_Builds]

@Re7_AI grok bot的设计思路跟大部分agent产品不太一样, 每个bot分配独立云端电脑而不是本地跑, 合上电脑agent还在继续干活。但身份认证、数据存储这些全部走cursor那套。

[Mai 麦尔彦 @Mai_Builds]

The core issue is React re-rendering on every single data point. if you're pushing thousands of updates a second through setState, you're re-running the reconciler thousands of times a second, that's the bottleneck, not the data itself. Ring buffers fix the memory side, fixed-size array that overwrites oldest entries instead of growing forever. Offscreencanvas is the bigger unlock though, it lets you draw to canvas from inside a worker thread, so the actual rendering never touches React's render cycle or the main thread at all.

[freeCodeCamp.org @freeCodeCamp]

React can struggle when you push thousands of real-time data points through the main thread.

In this article, Vineeth explains how to keep high-frequency data out of React’s hot path and move heavy work to workers.

You'll learn about ring buffers, SharedArrayBuffer, Web Workers, OffscreenCanvas, and more.

https://t.co/IvEhwJZc2x

[Mai 麦尔彦 @Mai_Builds]

The worktrees part is the actual insight here, not the UI. git worktrees let you check out multiple branches from the same repo into separate folders, each with its own working directory but sharing the same git history underneath. so when you run agents in parallel, each one gets an isolated copy of the code to edit without stepping on the others' uncommitted changes. There are no merge conflicts mid-session, no "wait did agent 2 just overwrite what agent 1 was doing."

Most people running multiple coding agents right now just spin up separate full clones, which wastes disk space and loses shared git state. Worktrees solve that properly. this is the pattern that's going to become standard as running 3-5 agents at once stops being a novelty.

[Danny Postma @dannypostma]

Got tired of switching tabs, so I built my own agent workbench.

[Mai 麦尔彦 @Mai_Builds]

@kevincodex @base How's x402 actually handling the settlement here, is the refund on outbid automatic on-chain or does someone have to trigger it?

[Mai 麦尔彦 @Mai_Builds]

A cheap viral stunt turned into an actual attention marketplace, and the price reflects how much some people value(d) that attention, not any inherent worth in the spot itself.

[Jonathan Wilke @jonathan_wilke]

Aaaaaaand we're live!

https://t.co/ZElVA2hguv

No ads. No API keys. No revenue sharing.

Just outbid your competitors to rank #1 and consider marketing done for today 🚀

[Mai 麦尔彦 @Mai_Builds]

@ConnorTalksAI capacity

[Mai 麦尔彦 @Mai_Builds]

@fabrice_mayrand an ad account that just started, selling discounted access to your Claude/GPT api calls, no thanks