[email protected]

recursive-mode is an installable skill package for coding agents. It gives your agent a file-backed workflow for requirements, planning, implementation, testing, review, closeout, and memory, instead of leaving the whole process scattered in context.

@dashuaishuai888 @ruihanmeimei01 he was buying Nvidia puts

@ruihanmeimei01 问问他为啥那么多外债

32岁外卖员凌晨下班，买了只平时舍不得吃的鸭腿庆祝生日，“三年不敢歇一天，存下29万，终于还清外债”！！你身边有没有这样的朋友？

@mitsuhiko Linking their Twitter profile is a strong enough signal

Alright. Proposal: the first time one submits a PR they need to jump on a video chat with the maintainer to explain their PR. If they fail they are banned from GitHub. x.com/mitsuhiko/stat…

@Z_Rex2017 A horrible fate sealed by liberal arts students

一个月以后，懂王也访问完中国了。东升西降大势已定，未来没有太多变数了。 老中精准的看出了熬就行，老美会自爆的。 从老美熬死苏联，到老中熬到老美自废武功，新的大国竞争范式已经出来了：熬死竞争对手就是最正确的做法。如果后来大国政客不瞎，大概也是沿着这个路径走了。

@larpcapitalwc Moving to China and learning Chinese is for the real billionaires

i feel like 'travelling the world' is now working class coded

@mksglu @YoniBraslaver Recursive-mode

@YoniBraslaver Context Mode

Cloudflare's Code Mode post argued agents are more efficient with code than a menu of MCP tools. We ran the experiment on monday's GraphQL API. SDK: 1 step, 15k tokens. Real MCP server: 4 steps, 158k tokens. 8.4× the cost, same output. yonibraslaver.pages.dev/posts/code-mod…

@aiandcloud 应该是

@trydotworks 哈哈哈 是啊！ 我上次偶然看到有人说是他掺 40% deepseek 🤣

今天最热门可能是这个， 英文圈发现了闲鱼和淘宝，可以以极低价格买到 Claude 和 GPT 🤣 他们这发现的太晚了，不过如果说想要质量稳定可靠，以及在意隐私和安全，还是官方渠道更好一些

It's so annoying seeing these types of posts because it's the wrong way to collect user feedback. Unstructured, out of context, completely subjective and with unknowable environment variables. Run proper user research.

@tierrapartners Puts on Tierra Partners for delivering a stale 36-months old take that's completely wrong

Call centers are so walking dead

@MarkMBissell @collision Accommodations have the highest margins in all of e-commerce.

insane fact from @collision: "I feel like Booking.com is a very underappreciated success story in tech... If you invested a dollar in Booking.com and a dollar in Google 20 years ago, you made much more money as a Booking.com shareholder."

@kcosr hope you brought the fishing gear otherwise very disappoint

touched grass

@Milkmeray Mustard and ketchup. Hotdog sauce does not exist.

i just found out my American friend doesn't know what hotdog sauce is. It literally has an american flag on the bottle of hotdog sauces wdym they don't have it and never heard of it? Germany just made that shit up!???😭😭 HUHHH

@undrvalue people will pay Upworkers and Fiverrs to create skills and manage their agents

What’s the bear case for Upwork at 5x forward PE, growing earnings with a stable topline. Is it not free money here? $UPWK

@dok2001 also the wrangler config is annoying

What are we missing?

@dok2001 billing limits

@muskinpingvirus HSK4S水平完全不能了解

原来HSK考这么深啊……

@OrganicGPT @Kimi_Moonshot DeepSeek is behind on productization. They basically haven't even started doing it. Will get a lot better but we don't know when.

DeepSeek v4 Pro is a huge let down. Maybe V4Flash is better but Pro keeps forgetting things, re-edits files it just edited, gaslights the user, has the most erratic reasoning, skips planning, etc. @Kimi_Moonshot (K2.6) has been better in that regard! Both in Claude Code harness

@chatsidhartha Yeah I really hope there will be a 256gb option

@trydotworks I’ll get a studio M5 ultra when it comes out. Hopefully with more memory.

I just ordered a maxed out MacBook Pro M5 Max with 128 GB of memory Chat, am I going to regret this

@Michaelzsguo you need to put a harness on the model, /goal doesnt work. use recursive-mode

1d 13h 20m, 3,596,831 tokens. Goal achieved? Not quite. It was a hard problem. The agent tried its best and went through 20 full model/eval rounds. In the end, the agent talked itself out of the original contract and declared the goal achieved. I probably would have stopped it anyway, since I could also see from the sidecar that it was struggling. Still, it was a good experiment. My 14" MacBook Pro held up well under a sustained run, with no throttling or heating issue. Qwen3.6 35B A3B OptiQ 4-bit running locally on MLX also held up well. It generated thousands of training data samples, averaging around 50 tps with reasonably good quality. Very impressive. DeepSeek 4 Pro was a good teacher for the training, though there are still areas for improvement. The end result: we LoRAed an expert model, Qwen3-4B-Instruct-2507 + MLX LoRA. We produced a compact 56 MB LoRA adapter on a 4B Qwen base that reaches ~59% three-way decision agreement on the original eval slice, ~91% violation recall, and ~98% valid JSON, but with a high false-positive rate. It is deployable, but probably not quite usable yet. Still, it gives me a clear direction for where to go next. I’ll write more about the whole process later. Stay tuned.