Hyeon

Updated Robotics Radar today with a new public brief and a few site improvements. The new note is titled “Humanoid Robotics Is Moving From Demo Quality to Factory KPIs.” The core idea is that humanoid robotics is starting to move past the first visual phase of the market (walking videos, manipulation demos, and viral) clips into a more operational proof standard: uptime, cycle time, integration time, intervention rate, repeatability, safety, and whether the robot can keep doing useful work inside a real factory after the camera leaves. Today’s report connects three signals. BMW is expanding its work with Figure 03 at Plant Spartanburg, moving into logistics sequencing rather than just isolated demos. AGIBOT is reporting production line metrics from its Longcheer deployment, including throughput, success rate, downtime, and integration time. Apptronik opened Robot Park in Austin, which looks less like a simple facility announcement and more like real world data infrastructure for training and improving Apollo robots. The common thread is that the humanoid category is beginning to be measured less by “can it do the task once on video?” and more by “can it improve a workflow repeatedly, safely, and economically?” That is the transition Robotics Radar is trying to track: demo → pilot → fleet evidence. I also updated the site structure around this lens. Robotics Radar now has a cleaner Robot Index for humanoid OEMs, specs, deployment status, and supplier exposure; a Market Signals section for quick tape checks around robotics and physical AI proxies; and a clearer split between Radar notes, Company Files, Frameworks, and value chain research. The broader market read through is that public robotics proxies can move on narrative scarcity before deployment economics are proven, so I want the site to track both sides separately: category formation on one side, and security-level structure, liquidity, and overhang risk on the other. robotics-radar-site.vercel.app/radar/2026-07-…

@0x_snatch @QFEX @trondao KR hunter

@hyeon__dev @QFEX @trondao thanks!

$USDT on @trondao network is now live on @QFEX Deposit $USDT trade global markets 24/7 Need an invite code? join our discord 👇

@Ru7Longcrypto Are you ok, ru?!

行情真的太干巴了，不知道分享点什么好 在小韩打了NAD+，感觉随着年纪越来越大，身体明显出现滑坡，要注重保养了 最近也会开始吃NMN、PQQ，认真佩戴WHOOP手环 以前真的不会觉得特别累，现在每天都会觉得睡不醒🫪 还有什么抗衰老神器可以在评论区分享

@0xWenMoon instagram

Any creators who also post on YouTube?

@YashasEdu omg! bull point

@hyeon__dev Yes ser

@YashasEdu Have you tried it?

@hyeon__dev The product is crazy good ngl

@EbenezerDYOR It's just beginning!

@hyeon__dev What an amazing thread

@get_truenorth gm TrueNorth

@hyeon__dev good piece man

@032felix Thx for reading!

@hyeon__dev Solid work bro🙌

cc. @get_truenorth @moonshot6666 @corgil @WillyChuang @EbenezerDYOR

@adamnemecek1 @stanfordnlp Everyday, I talk with AI

@hyeon__dev @stanfordnlp AI is more like a 90-year-old scientist who won the Nobel Prize 60 years ago. There are undeniable glimpses of genius but you might have to periodically remind him who you are and why you are talking to him.

Currently, AI is like a brilliant intern who shows up as a brand-new hire every single day. True AGI, as Percy Liang notes, is more like an employee who learns your company's workflow for two months and returns every day carrying that accumulated experience. Continual learning goes beyond long context (feeding past data) or long term memory (retrieving past data). It fundamentally evolves how the AI makes decisions and executes tasks over time. If CoT extends reasoning time per problem, continual learning extends the lifespan of intelligence across multiple tasks and time. At this stage, AI transitions from software you subscribe to into an operational asset whose value compounds over time. Capabilities won't be measured by base model benchmarks alone, but by: Base Model + Memory + Tools + Permissions + Behavioral History & Feedback. Even with identical base models, AI will diverge based on where and with whom it worked. A company’s real moat won't be raw API access, but the proprietary organizational experience built within its system. This is also why DeepSeek’s cost efficiency strategy matters. Internally, lower compute costs accelerate experiments toward AGI. Externally, cheap APIs and open source models expand the feedback loop. It's not just a price war; it's a strategy to push the ceiling of intelligence faster with limited resources.

@realkimchiboy 또 인간은 적응의 동물 아니겠습니까 적응해봐야지요

@hyeon__dev 멋지지만 슬프네요. 인간의 강점이 사라져 가는게 보이니

@dorong_x huge insight!

Some thoughts after listening to recent interviews with Chinese AI leaders. There is plenty of overlap with how Western AI leaders see the future, but the perspectives coming out of China are interesting in their own right. There is no clear answer on which direction will win: how companies should be structured, how committed they should be to open source, or how they should balance product development against the pursuit of AGI. I mostly treat those debates as different possibilities rather than firm conclusions. As a user and investor, the two ideas that caught my attention were continual learning and embodied AI. At this point, I reserve most of my own brainpower for thinking. I delegate a large part of the execution, information gathering, and repetitive work to agents. From a user's perspective, the existence and usefulness of general-purpose agents already feels astonishing. Coding agents are great, but I suspect the complicated workflows we build around them today will eventually disappear into the lower layers of general-purpose agents. Users will no longer see or think about them. They will simply become part of how the agent works. If general-purpose agents become commercially viable, the productivity gains could be enormous. We are not there yet. Maintenance, management, reliability, and cost are still difficult. Even as an enthusiastic user, however, the biggest missing piece is obvious: these systems do not really feel as if they accumulate experience. I can build elaborate setups with Obsidian, MEMORY.md, Honcho, and other memory systems. They help, but the agent still feels like it is selectively retrieving written records through prompts. It does not feel like the system has genuinely learned from everything we have done together. Context windows have grown dramatically, but they still have limits. The moment a user becomes conscious of those limits, the experience starts to break down. I assumed this would eventually be solved, even though I had no idea how. What surprised me in Liang Wenfeng's interview was his view that nobody in the world has yet presented a convincing solution. A system needs to learn new information without losing old capabilities. It needs to remember useful experiences, incorporate new knowledge, correct what it previously got wrong, and prevent bad information from accumulating over time. These problems are tightly connected, and simple prompting or external memory files do not appear sufficient. I do not have a clever answer here. I just hope someone solves it. The next topic is embodied AI: intelligence that can work in the physical world. As an investor, I still believe the next major opportunity lies somewhere around physical AI and robotics. General-purpose agents may first find commercial success through advertising, shopping, and other digital businesses, but nearly every major AI leader appears to see robotics as an eventual destination. Perhaps general-purpose robotics becomes possible once general-purpose agents solve continual learning. A robot operating in the real world cannot rely forever on a static model and a carefully maintained memory file. It has to accumulate experience, adapt to new environments, correct mistakes, and keep learning without forgetting how to do its existing tasks. And if the goal is to change the real world, physical action eventually matters more than another layer of software. Better computer programs will obviously change society. But cleaning, manufacturing, transportation, and caregiving will not be done by our ancestors. My relatively short but somehow very long journey trading Korean equities is now over, but I remain bullish on the changes AI will bring to the world. I also believe that using these systems consistently, studying them, and gradually expanding what I delegate to them will make a meaningful difference in my life. I will still allocate much more of my portfolio to dependable index funds. Survival comes first. But learning about new technologies and holding a few individual stocks with call-option-like exposure to the ones I believe in adds some excitement to life. They are a bit like lottery tickets, except backed by research and a view of how the world might change. For that strategy to work, steady cash flow matters more than anything else. So I also need to remain diligent as a crypto player.

@ramztd 건망증 치료 부탁해요

@hyeon__dev 너무 능력이 좋지만 수상하게 건망증이 심한 친구

AI may blur job boundaries long before it eliminates jobs. When PMs can prototype code, designers can build working demos, and engineers can test product ideas directly, the first order effect is organizational ambiguity: who decides what is ready? Stone’s point: cheaper prototypes don’t make expertise obsolete. They shift scarcity toward systems thinking, deciding what reaches production, connecting quality, security and data, and owning outcomes. But how will juniors learn? Lenny’s Podcast — Netflix CPTO Elizabeth Stone youtube.com/watch?v=t0GiTy…

@nowlovepan 읽어주셔서 감사합니다~~

@hyeon__dev 좋은글 감사합니다

@nowlovepan 오 인용해주셨군요! 맞아요 결국 개인이나 기업에서 핏에 맞는 AI를 잘 학습시킬수 있느냐가 다음 스텝에서 커다란 차이를 만들것 같아요. 저도 매일 AI와 함께 일도 하고 실생활에도 적용시키고 있는데 요즘 정말 즐겁습니다~~

이 글은 지금 AI가 매일 새로 입사하는 인턴처럼 매번 처음부터 일을 배워야 하는 문제를 얘기함 저도 Codex를 계속 쓰면서 체감하는 중임 처음에는 말투와 작업 방식을 하나씩 알려줘야 했는데 지금은 싫어하는 표현과 원하는 소재까지 어느 정도 알고 있음 이게 더 쌓이면 같은 모델이어도 남이 쓰는 AI와 내가 쓰는 AI는 완전히 달라질 거임 그럼 우리는 지금 뭘 해야 할까? 나만의 프롬프트/작업 규칙 정리하기 좋아하는 방식과 싫어하는 표현을 AI에 계속 피드백하며 맥락을 학습시키기 같은 이런 과정들에 시간을 투자해야 할 때. 결국 내 AI를 '길들이는 과정' 자체가 내 미래 경쟁력이 될 듯

@CoinList hello? @CoinListSupport

At that point, I decided not to use @CoinList anymore. Did you guys just see this post now, or did you get hacked by kimi?

@Ru7Longcrypto 可以帮我问问朋友该买什么吗？

昨天朋友已经离婚了。在民政局门口，前妻问有什么话要最后对她说的吗？朋友说：还是那句话，时间会证明光模块和算力的。