Erik

@realNyarime Phil Schiller? What an idiot. On stage talking about USB ports.

面对Neo的接口 库克是否会想起2012年那场遥远的WWDC

@alreadydawn Absolutely terrible post. 1. Nobody ever wanted to go to Jinan. Of course it's empty. 2. Wrong. You have to wait 10, sometimes 40 minutes to get a Didi during commuting hours. 3. Coco Park has never been a hot mall

I'm seeing more and more instances of demand not keeping up with supply in China: * Hotels are abundant, even at random cities like Jinan, where rates are amazingly low. You can get a decent place at below 20 usd per night. The ones I've stayed at so far seem to lack customers in general. The hotel I stayed in Yangzhou has a coffee/drinks bar that has been unmanned for quite some time and just collecting dust. * Lines of taxis waiting for riders outside of HSR stations. Some pointed out that this has been the case for years, but the number of potential riders does not match the sheer number of cars. Even at 3rd tier cities at night where few people are out and about the station, the taxi lines stretch on forever like the Great Wall. (I get that taxis around the world have been severely impacted by ride-hailing softwares like Uber and Didi, but these great wall-like taxi lines are nuts to witness.) * Malls outside of the hot ones in Tier 1 cities (ie Coco Park in Shenzhen) are quiet. The one I'm next to here in Qingdao by their Olympics sailing center is nearly deserted, with most storefronts closed. I'm sure there are many reasons for this supply/demand problem (as do most complex issues), but a big factor has to be involution (內卷), where the Chinese try to outcompete each other at the exact same thing, driving supply way up and prices down. This all reminds me of a story shared by @spandrell4 that goes: "When a Jewish businessman opens a gas station at a new town, other Jews come to the city to open various businesses - bakery, accounting firm, law firm, restaurant. When a Chinese businessman opens a gas station at a new town, other Chinese come to the city to open gas stations."

@kakashiii111 Lmao. Nvidia knew everything. Random companies buying hundreds of millions worth of GPUs, as if Nvidia didn't know what was going on.

It is interesting that all these smugglers' companies have billing locations in Singapore and headquarters in the US. Not sure if Nvidia knew, but they made accounting changes twice in less than a year that may not be a coincidence, and could be related to these sales. Maybe?

@AsiaStock The reason is you try to launch in the Chinese market while you sit in the US and make strategic decisons based on zero knowledge of the market. Chinese firms fail at going international for the same reasons.

为什么西方公司难以打入中国市场？Uber 创始人 Travis Kalanick 将与中国企业竞争分为两个阶段。 第一阶段：中国竞争对手会以极高效率复制 Uber。一个新功能上线一周，就已经被完整克隆。当你把“复制”做到极致时，最终会发现没有更多东西可以复制。 第二阶段：中国公司开始转向本土创新，推出更符合本地市场和文化的功能。Kalanick 举了美团的例子，比如外卖取餐柜，这种模式并非简单复制，而是在中国市场中非常高效。 在AI竞赛中，这又会如何演变？

@lauriewired What do you mean, seems likely to be aimed at the server market? Their official release and documentation says it's a server product. You just have to look at the specs to understand it's a server product.

open questions (feel free to reply if you know): 1. I don’t understand how this can possibly have the RVA23.1 spec already. It’s only been out for ~6 months, yet alibaba seems to be claiming tapeout (“5nm verified for production"). It’s not a paper launch; *someone* has some real silicon otherwise they wouldn’t be able to claim SPECint2017 numbers…(unless it’s only simmed?) 2. TDP is up in the air atm. My guess is the core is likely to compete in the server-market 3. Undisclosed fab…but who else could it possibly be besides TSMC? We know it’s 5nm… 4. It seems to use their own additional “Matrix Multiply Extension” for AI inference. Alibaba is claiming the C950 has “naive support for Qwen3 and DeepSeek V3”…but no figures or toks/sec. I’m not sure how much they’ve improved past this, but you can find an open source version of their MME on github. But, it hasn’t been updated since december of 2024… github.com/XUANTIE-RV/ris… (also attaching the giant chart from David Huang that shows a ton of SPECint2017/GHz scores)

Wow, RISC-V is really gaining traction again. Alibaba just announced the Xuantie C950…which is basically claiming Apple M1 (ish) levels of performance. I don’t see a lot of people talking about it! (2.6/Ghz SPECint2017, Apple M1 P-core also around ~2.6)

@TheStalwart China makes the best titanium bikes today

When I see quotes like this, the main thought I have is that there is a deep continuity in Chinese thinking about industry, this obsession with building out capacity and establishing self-reliance, and also that the gap between Mao- and post-Mao-era is overstated.

@adcock_brett We don't need new interfaces. Voice is a nice to have. We need new backends.

Today I'm excited to introduce Hark, a new artificial intelligence lab building the most advanced, personal intelligence in the world We've been in stealth for 8 months, assembling one of the greatest AI and hardware teams on the planet I want to explain why I started Hark and what we're focused on I've spent the last 3 years working on the hardest AI challenge imaginable: giving AI a humanoid body. On the digital side, I've been using all the existing LLM chatbots - and I have to say, they feel incredibly dumb to me AGI, in the limit, should feel like a sci-fi movie. It should be able to listen and talk. It should have persistent memory and be highly personalized. It should see and touch the world. But we're far from this today We are crafting a new interface to AGI. Intelligence that lets you offload your mental workload into a system that begins to think like you and sometimes ahead of you We started Hark with one goal: build the world's most advanced personal intelligence - paired with next-generation hardware designed to serve as a universal interface between humans and machines hark.com

@DeeperThrill Your post-planning workflow can be improved significantly if you use a skill like rlm-workflow. You shouldn't need to ask for reviews manually. github.com/doubleuuser/rl…

My coding workflow is currently: 1. Spend about ten minutes writing a plan for a new feature in a markdown file as bullet points. 2. Spend about an hour with a bunch of back forths between Opus 4.6 and Codex 5.4 xhigh, asking each one to improve the markdown plan file. 3. Read the final markdown plan, which ends up as a bunch of bullet points, includes implementation details, and is usually between 200-500 lines long. 4. Ask either Claude Code or Codex to implement it. 5. Ask the other one to review the implementation. 6. Run my /deslop command on the implementation (see my pinned tweet). 7. Deploy, test, and ask an agent to fix any bugs.

@chatsidhartha @CloudflareDev Cloudflare is new AWS and AWS is dead?

Durable Objects are the greatest invention of our time and I guarantee you, @CloudflareDev will be the default compute provider to run agents everywhere very soon Nothing beats the speed, scale, isolation and intuitiveness of this platform

@fayazara wicked

Here's me spinning up multiple dynamic workers on the fly in milliseconds. they're all isolated btw

@teortaxesTex GLM 5 performs similarly to GPT 5.2 (No thinking) so it's actually not such a bad showing. Kimi climbing too.

what's worse, the best Chinese model is quite far below Gork 4.20 (best Western models are at 0.79 now btw) WeirdML and CritPt are some of the best proxies of progress towards actual recursive self-improvement, building automated research scientists.

@whyyoutouzhele China's technological advantage and speed of development does not come from 996, but from the last 4 decades work on building various types of infrastructure. If China stopped 996 culture, it would be more innovative, more efficient and could launch better products, faster.

@whyyoutouzhele I have more than a decade's experience working in China tech. 95% of 996 is performative; fake work. And nobody starts at 9, they start at 11. 6-7:30 people have dinner and slack off. It creates a very inefficient work environment, where there is no planning and lots of rework.

“中国没有生活，只有工作” 3月22日，网友转发一位外国企业家评价中国“996”：早9晚9、一周6天，“只有工作没有生活”。他称无意批评，但强调应平衡工作与生活，希望企业让员工真正发展。 视频发出后迅速引发网友们的共鸣，评论区甚至有网友表示“996”算好的了，我们“777”早7到晚7，有事只能请假。还有的网友表示，“996”算好的了，之前在广州进厂都是“897"(早上八点到晚上九点，一周上七天)。

@thdxr Inference will indeed explode when there are agentic loops inside every piece of software. On the other hand we will probably see 3x-10x improvements in every part of the stack annually, from model architecture, to chip architecture, to the actual software we run.

the infra we have right now is barely enough for today's ai inference demand the options are demand goes up, stays flat, or goes down everyone seems to be certain that building out more capacity is "bubble fomo behavior" they're that certain demand is going down?

@caterpillarous in that case they need to switch to Codex

其实从去年下半年开始就这样了，google deepmind和labs的人大多数写代码用的都是opus 😅

@gormankind Probably this weekend, as long as I have tokens left. Will drop you a message if I do.

@try__working Glad you like it!! If you try Rivet open to any feedback

for the Claude-pilled designers who don't want to think about MCP

@peterom You shouldn't sense check, that's drift. You should audit against locked artifacts from previous steps.

@try__working Nah, that's all included as part of a sense-check step

Sharing Megaplan, a harness for Claude + Codex to make very robust plans! It uses their respective strengths to achieve a level of quality that's far beyond what either alone can manage. Once installed, just say 'Megaplan this for me'! Link: github.com/peteromallet/m…

@peterom After you generate the plan and start implementing it, you then need to tell it to run tests, add logging, review the code and implementation against the plan, create a manual qa plan for you etc. Lots of work, and you're bound to forget some steps.

@try__working What are you referring to here?

@peterom no, its way more cumbersome to have to steer the model during implementation of the plan

@try__working That seems very cumbersome