GeekPark

722 posts

GeekPark banner
GeekPark

GeekPark

@GeekParkHQ

A window into China’s tech frontier. First institutional angel investor @UnitreeRobotics DM me what you're most curious about.

Menlo Park, CA شامل ہوئے Şubat 2026
784 فالونگ473 فالوورز
پن کیا گیا ٹویٹ
GeekPark
GeekPark@GeekParkHQ·
Unitree Robotics goes before the STAR Market committee today, on track to become China's first public humanoid robot company. As its first institutional angel investor, it is a good moment for us to revisit the early bet. Jack Zhang, founder of @GeekParkHQ, wrote ~4,000 words on how, in 2018, he wired nearly the entire first fund into @UnitreeRobotics. x.com/GeekParkHQ/sta… The highlights: 1/ How a 30-second clip on an obscure WeChat account started the whole bet. 2/ The first meeting: 3 hours, no deck, no desks, just a hallway couch. (Not even Unitree's!) 3/ Why hydraulics were a dead end for any commercial product, and electric the only road. 4/ How Wang Xingxing rebuilt a top lab's research architecture for under 20,000 RMB on motors he sourced and characterized himself. 5/ "I would not have written the same check for the same founder building the same robot in Palo Alto." 6/ How Unitree's two largest backers today - Lei Jun @Xiaomi and Wang Xing @meituan both met Xingxing in 2017 and passed. (Their stakes today: ~$800M.) 7/ What GeekPark did when the structure broke and Unitree's cash was about to run out. 8/ Why Xingxing bet from day1 that research labs were the right first customer. 9/ What matters to GeekPark more than the Unitree bet itself. Full read:
GeekPark@GeekParkHQ

x.com/i/article/2060…

English
4
4
13
1.4K
GeekPark
GeekPark@GeekParkHQ·
@a16z human in the loop always sounded like a compliance checkbox to me. tandem bike is at least honest about it. and it asks the thing nobody answers which is who grabs the wheel when the model is wrong but really really confident about it
English
0
0
0
409
a16z
a16z@a16z·
Mira Murati says frontier AI should be built like a tandem bike: "Having humans in the loop doesn't quite describe it because it sounds like a checkpoint where we're signing off something, and then you're good to go." "It's more like creating systems that are not just autonomously advancing and leaving civilization behind, but are more like a tandem bike." "When you're going up a hill, maybe whoever is stronger is pedaling harder. But both hands are on the wheel. That's quite important because that's a different system. It's a system designed for collaboration." "It will increase the level of agency that people have, and also it will help us steer the research direction towards creating outputs that are more value-aligned." @miramurati at Bloomberg Tech live with @emilychangtv
English
57
72
695
114.5K
GeekPark
GeekPark@GeekParkHQ·
AI spend right now is cloud spend in 2014. Everyone knows it matters, nobody has any governance, the bill is about to get genuinely scary. "your agent spent the whole budget" is gonna get said in a boardroom this year. I think ramp sitting right in front of that at 44b is the smart bit imo
English
0
0
0
394
Eric Glyman
Eric Glyman@eglyman·
As I wrote this, I saw X go into meltdown over tokens. You've seen the headlines: “Uber blows yearly AI budget in just one quarter.” “Meta employee burns 281 billion tokens in April.” But, the problem isn't spending. Spending works. Since 2023, the top quartile of our AI spenders doubled their revenue. The bottom quartile? Flat. It's blind spending. We don’t know which spend worked. A sales team has qualified leads. A support team has resolved conversations. These are units you can measure against. All a token tells you is the meter ran, not whether the work was worth it or not. Finance says, “half the budget,” engineering says, “double it” and you don’t know who’s right because there is no shared language of value. There’s no attribution, and no attribution means no allocation. For example, right now, all work, no matter the size or shape, defaults to frontier models. But meeting summaries and calendar updates don’t require GPT-5.5 Pro. In isolation this seems trivial, but re-route just 10% of a $10M AI bill from frontier to GPT-4 level intelligence you’ve saved nearly one million dollars. This sounds like a made-up stat — it’s not. It truly is that much cheaper. This is the future of finance: not blindly rubber-stamping or rejecting AI spend, but allocating it with the same rigor companies apply to headcount.
Eric Glyman tweet media
Eric Glyman@eglyman

Today, Ramp raised $750M at a $44B valuation. Last time we grew this fast, we were 1/20th the size. For 2000 years, business was built on two pillars. Today, a third: intelligence. It’s your least governed cost. It’s also your single greatest opportunity.

English
50
52
624
195.6K
GeekPark
GeekPark@GeekParkHQ·
@JitendraMalikCV i've seen so many VLA demos that look incredible and then ship absolutely nothing. meanwhile Figure at BMW: 90k parts, 30k cars, 99%+ placement. -that's years of boring sensorimotor data and brutal evals. the unsexy stuff is the stuff that works, every time, it's almost annoying
English
0
0
2
1.3K
Jitendra MALIK
Jitendra MALIK@JitendraMalikCV·
I want to offer some unsolicited advice to computer vision researchers jumping into robotics. Don't focus too much on VLMs, VLAs etc. That's fine, but the real action is at the sensorimotor level. Most of the open problems in robotics are in manipulation, which is about hand-object interaction, and contacts and forces are central. Proprioception and tactile sensing are as important as vision. Don't get seduced by cherry-picked demos. You can't do robotics without doing robotics.
English
19
141
1.3K
80.9K
GeekPark
GeekPark@GeekParkHQ·
@TimelessMartian 920m a month, 110k GPUs, til 2029. And Google called it a bridge - google. they have more compute than almost anyone on earth and they're still renting off a rocket company because demand blew past what they built. that's the actual story here
English
0
0
1
25
Timeless Martian
Timeless Martian@TimelessMartian·
Google entered a $920M deal with SpaceX for access to 110k GPUs
Timeless Martian tweet media
English
7
2
42
11.9K
GeekPark
GeekPark@GeekParkHQ·
@levie people keep posting this like it's bullish lol. coding is the easiest possible case for agents and we STILL can't take the human out of "wait should we build this." now go try that on literally any other job. yeah
English
0
0
0
136
Aaron Levie
Aaron Levie@levie·
Coding is basically the pinnacle of what you could reasonably automate with AI, and yet we still need human engineers to oversee agents for them to be effective. The AI models are trained on an incredible amount of sophisticated code. The users are highly technical and can use the latest tools quickly. The work is “verifiable” because you can test an app. The outcomes are often removed from the quality of the code (you can have sloppy code but the app can still work). And the context for the agent is often already digitized and sitting in the codebase. That’s an incredible amount of benefits that AI coding agents get to work with. Some of those apply to knowledge work, but most don’t in areas where the work needs to be fully reviewed to be useful, or where data isn’t as abundantly digitized. This makes the job for agents in knowledge work more complicated. So if with all of that, engineers still remain in very high demand, the risks are going to be less than what’s perceived for other areas of knowledge work. Agents will let people do far more than they did before, but the people don’t go away.
Joe Weisenthal@TheStalwart

I like having a job. So consider this take to be drenched in cope. But as of right now, I think that: coding being a relatively “easy” thing for AI to learn + the existence of many currently employed coders, implies that we’re a long way off from mass while collar disruption.

English
41
31
294
51.7K
GeekPark
GeekPark@GeekParkHQ·
@leerob 3 days, $260 in tokens, killed a CMS that cost them 57k a year. wild number but the thing that gets me is smaller. writing the spec used to be the job. now the job is describe it badly and fix what comes back. spent a decade getting good at the wrong part apparently
English
0
0
0
181
Lee Robinson
Lee Robinson@leerob·
Cursor (and coding agents generally) still blows my mind daily. Just today: 1. I shipped a new landing page. I gave a 10min voice note to Cursor, left to go eat dinner, and came back to a 90% finished version. Made some small design and copy tweaks and merged. 2. Had Cursor dig through Search Console and Semrush with computer use, researched places we could improve SEO, and then merged 3 PRs with fixes. 3. Used the Supabase MCP to pull thousands of emails from the Compile waitlist, had it research them with web search based on ideal fit for the event, and got back a CSV with the top people to invite and why. 4. Updated an internal app I built for doing company-wide surveys (think Typeform but Cursor branded) in a few hours before our All Hands. 5. Had a few agents researching furniture I'm hoping to buy. They searched the web for a bunch of variants and then made a custom shopping cart (just an HTML page) with images, prices, links, and tons of details. Super helpful. I don't do this every day, of course, but it's still wild to me this is the new normal for what someone with a computer and AI can do. Most of these were running in the cloud as I was between meetings, just humming away in the background. I could check the app (🔜) to see progress and merge PRs. What a time to be alive. (P.S. if you extrapolated my usage today, I'd still be on the $200/mo plan)
English
71
35
810
44.6K
GeekPark
GeekPark@GeekParkHQ·
@fuzziphy Tiny boards like this are where the “AI meets physical world” story actually starts👍
English
0
0
0
53
Matt Thompson
Matt Thompson@fuzziphy·
Made a quick breakout for the AD5628 8 channel SPI DAC this evening
Matt Thompson tweet media
English
8
3
128
9.9K
GeekPark
GeekPark@GeekParkHQ·
@VraserX @koltregaskes Model beef is becoming NBA Twitter: tiny samples, huge takes, everyone’s got a franchise player. The useful test: same prompts, same repair budget, measure latency + correctness + how many times the model needs a babysitter.
English
1
0
0
17
GeekPark
GeekPark@GeekParkHQ·
@koltregaskes The SVG discourse is about product taste. Benchmarks tell you capability; artifacts tell you whether builders would actually use the thing. I’d judge it on editability, DOM correctness, and -can I ship this without apologizing?
English
0
0
0
21
GeekPark
GeekPark@GeekParkHQ·
@mikekalilmfg Vietnam humanoids are worth watching because this could become a VinFast-style industrial ambition: national champion, hardware learning curve, global export story. The question is whether they can get from PR robot to factory-grade uptime.
English
0
0
1
23
Mike Kalil
Mike Kalil@mikekalilmfg·
Vietnam’s Richest Man Is Betting Big on Humanoid Robots Vietnam’s humanoid robots are going global. Its wealthiest man is going align on humanlike machines as the nation’s industrial base expands rapidly. The South Asian country wants to move from exporting mostly low-cost goods to being a serious player in emerging tech. It’s positioning itself as a more affordable and flexible alternative to China, which is investing more in humanoids than anyone. It’s early days for Vietnam’s humanoid ambitions, but those are moving unusually fast. Less than two years into the push, a range of Vietnamese prototypes is already moving toward mass production. They’re marching into the workforce with plans to send them far beyond the domestic market already taking shape. The main driver behind Vietnam’s humanoid push is Vingroup, the country’s largest private conglomerate. Phạm Nhật Vượng founded Vingroup in 1993 after building his fortune in Ukraine through his instant noodle business, Technocom. When he returned to Vietnam, he scaled the success by an order of magnitude. He became Vietnam’s wealthiest man as Vingroup’s footprint far beyond instant noodles. Today, the billionaire entrepreneur oversees a diverse suite of businesses across real estate, automotive manufacturing, hospitality, retail, healthcare, and next-gen technologies. Vingroup formally threw itself into the humanoid fray in late 2024. It was Vietnam’s first major effort in humanoids since the 2000s when TOSY Robotics gained international visibility with its TOPIO series. The robots became known for their table tennis abilities, though they never reached pro status.
Mike Kalil tweet mediaMike Kalil tweet mediaMike Kalil tweet media
English
3
3
21
1K
GeekPark
GeekPark@GeekParkHQ·
@JasonrShuman Exactly. Robotics gets interesting when it stops being “replace labor” and becomes “protect revenue when labor supply is capped.” That framing is way more CFO-friendly. Less Jetsons, more unit economics. Which is how stuff actually gets bought.
English
1
0
0
33
Jason Shuman
Jason Shuman@JasonrShuman·
When Labor Supply is Constrained. Robotics become a revenue driver.
Jason Shuman tweet media
English
4
1
21
1.6K
GeekPark
GeekPark@GeekParkHQ·
@Ronald_vanLoon @lukas_m_ziegler Waste sorting is a deeply underrated robotics wedge: dirty, repetitive, labor-constrained, and ROI is not vibes-based. The sexiest robot businesses may start in the least Instagrammable rooms.
English
0
0
0
21
GeekPark
GeekPark@GeekParkHQ·
@VraserX I think...“Too powerful to release” is undefeated founder theater. Maybe Mythos is cracked, but run it through SWE-bench Verified, τ-bench, WebArena-style messy workflows first. Otherwise we’re just trading screenshots like baseball cards.
English
0
0
0
11
VraserX e/acc
VraserX e/acc@VraserX·
Mythos feels like the most obvious IPO bait ever. “Too powerful to release” is such a convenient story when you’re trying to make a $965B valuation sound sane. Maybe it’s amazing. Or maybe Anthropic just needed a mythical dragon in the pitch deck so Wall Street doesn’t ask what the margins look like.
VraserX e/acc tweet media
English
13
1
30
1.5K
GeekPark
GeekPark@GeekParkHQ·
@craigweiss Monthly AI stack rebuilds are funny until your team loses a week to tool migration.
English
0
0
0
23
Craig Weiss
Craig Weiss@craigweiss·
it's that time of the month again where i have to completely update my ai tech stack
English
41
0
118
4.5K
GeekPark
GeekPark@GeekParkHQ·
@eastdakota @pmarca This is the tax of lazy pattern matching. You don’t just pass on a founder; you pass on the category before it has language. Infra always looks weird until it becomes the thing everyone depends on. Then suddenly everybody “saw it early.” lol.
English
0
0
0
488
GeekPark
GeekPark@GeekParkHQ·
@WatcherGuru The market is trading the headline, but the deeper issue is nastier: privacy systems can make “prove no counterfeit happened” extremely hard. AI-assisted audits are about to make old cryptographic assumptions sweat. Evals > vibes, even for money.
English
0
0
0
46
Watcher.Guru
Watcher.Guru@WatcherGuru·
JUST IN: Zcash crashes 48% after Claude AI finds critical vulnerability allowing unlimited minting of $ZEC. It went unnoticed for 4 years until it was patched on June 1st.
Watcher.Guru tweet mediaWatcher.Guru tweet media
English
960
1.4K
13.5K
2M
GeekPark
GeekPark@GeekParkHQ·
@TencentHunyuan Planning evals are where agent hype goes to get humbled. I like this because it tests coupled constraints, not just “sounds reasonable in a chat.” If your agent can’t pass checklists, it’s not autonomous. It’s a very confident intern.
English
0
0
0
77
Tencent Hy
Tencent Hy@TencentHunyuan·
Planning is where LLMs move from “saying” to “doing.” Tencent Hy, in collaboration with the Gaoling School of Artificial Intelligence at Renmin University of China, is excited to open-source PlanningBench - a scalable, verifiable framework for evaluating and training LLM planning capabilities. With PlanningBench, you get: ✅ 30+ real-world planning tasks ✅ Automated verification ✅ Evaluation and training support See how top-tier LLMs perform on PlanningBench 👇 Resources: arXiv: arxiv.org/abs/2605.20873 GitHub: github.com/Tencent-Hunyua… HuggingFace: huggingface.co/datasets/tence… #PlanningBench #TencentHunyuan #OpenSource 📷
Tencent Hy tweet media
English
6
15
73
7K
GeekPark
GeekPark@GeekParkHQ·
@RuxandraTeslo @SoumayaKeynes @jessicacweiss China biotech is the AI story people keep sleeping on. Not “more papers,” but faster movement in painful, measurable categories like multiple myeloma. When progress shows up in clinical paths, the vibe check is over.
English
0
0
0
30
Ruxandra Teslo 🧬
Ruxandra Teslo 🧬@RuxandraTeslo·
@SoumayaKeynes @jessicacweiss I wrote about what they're doing in biotech. I think there's a lot to learn, frankly.
Ruxandra Teslo 🧬@RuxandraTeslo

China is winning the drug discovery race. There's no better example of this than multiple myeloma. worksinprogress.co/issue/the-bloo… It's one of the most painful cancers, destroying bone from within. For decades, patients endured cycles of brutal treatment and relapse. Then came Carvytki: a one-time CAR-T infusion that appears to cure some patients who have failed multiple treatments. Its development story, beginning in 2016, was an early signal of a shift now making headlines: the US is losing biotech dominance to China. Though the foundational science was largely American, a nimble Chinese company moved faster with a better molecular engineering idea. Unless the US addresses clinical-trial bottlenecks slowing early in-human data, more breakthroughs will be developed elsewhere, weakening the ecosystem American biopharma depends on. Some key points from my article for @WorksInProgMag, with my friend Amol Punjabi, of @EvidenceOpen: 1) Multiple myeloma is not only extremely painful in and of itself, but also one of the most brutal cancers to treat. As first-line therapy, patients endure four drugs simultaneously, then a stem cell transplant, followed by continuous maintenance therapy. And most still relapse, with each treatment round carrying worse chances. 2) A drug called Carvykti, approved in 2022, is changing the treatment landscape. Carvytki acts as a single, one-time infusion. It's a CAR-T therapy, part of a new wave of transformative immunotherapies: made from the patient's own immune cells and reprogrammed to hunt cancer. In patients who had already failed 4+ other treatments, 33% were still disease-free after 5 years. The results as earlier line therapy look even more promising. 3) Most of the foundational science was American. Decades of CAR-T research, and in 2013 the NCI showed BCMA-targeted CAR-T cells could kill myeloma in the lab. 4) But the drug that ultimately changed myeloma, Carvytki, originates from China. Carvytki beats Abecma (the American CAR-T for myeloma) by a wide margin: 36 months of progression free survival in heavily pre-treated patients versus Abecma's 9 months. 5) In 2016, Legend Biotech was just beginning clinical trials. This was the same year the American team was publishing their first-in-human results. Legend started later, but moved faster. Clever engineering and China's ability to get drugs into humans quickly gave them the edge. Large American biopharma J&J ended up striking a deal with Legend and developing the therapy. 6) Never underestimate the llama: US-developed Abecma used mouse antibody fragments to target BCMA. Chinese startup Legend used llama nanobodies instead. These are smaller, more stable and bind more cleanly to BCMA. The usage of llama as opposed to mice antibodies is what is believed to lead to Carvytki's superior efficacy. 7) In retrospect, Carvytki should have been an early warning. China is winning the drug discovery race through deliberate policy. Their first-in-human clinical trials can launch in 6 months vs 18+ months in the US, letting them iterate faster between lab and clinic. The @nytimes recently reported that ~50 percent of major drug deals this year involve Chinese-origin drugs, up from nearly zero a decade ago. 8) The US still leads in late-stage development, as shown, but the pipeline feeding it is increasingly Chinese. The worry is that this will mirror what happened in solar, batteries, and EVs, where early-stage dominance eventually became control of the entire chain. 9) A proposal to streamline early stage trial regulatory requirements to keep the US competitive has made it into the President's 2027 budget for the FDA. But Congress has to act to make it a reality.

English
2
3
22
2.5K
Soumaya Keynes
Soumaya Keynes@SoumayaKeynes·
When China spent decades building dominance in rare earths supply chains, what was it really trying to do? Build an economic weapon? Or something else? I asked @jessicacweiss on The Economics Show with Soumaya Keynes podcast this week... ft.com/content/34a893…
Soumaya Keynes tweet media
English
18
25
110
76.3K
GeekPark
GeekPark@GeekParkHQ·
@jukan05 Everyone watches NVIDIA - but memory is where the AI capex story gets brutally real. “Please make more” is basically Jensen saying: GPUs are the headline, HBM/DRAM is the choke point. Silicon Valley loves models; fabs decide who eats.
English
1
0
0
335
Jukan
Jukan@jukan05·
[Exclusive] Jensen Huang “Please Make More”… Hynix Opens Era of 1 Million DRAM Wafers per Month It has been confirmed that SK Hynix shared with major partner companies a plan to grow its DRAM wafer production capacity to roughly double the current level between 2030 and 2031. This is a mid to long term expansion plan that was already in place before NVIDIA Chief Executive Officer (CEO) Jensen Huang wrote “Please make more” on an SK Hynix DRAM wafer at the Computex exhibition venue. According to the industry on the 5th, SK Hynix’s purchasing organization and the personnel in charge of the Yongin semiconductor cluster, among others, have been informing major partner companies since about two months ago of a plan to expand production on a wafer input basis through 2030. The core of it is raising the monthly DRAM wafer input capacity, currently at the level of 550,000 wafers per month, to around 1 million wafers in 2030. The 550,000 figure includes the output of the Wuxi plant in China (about 200,000 wafers). The expansion is concentrated in the Yongin semiconductor cluster. SK Hynix will divide the Yongin Phase 1 fab into six cleanrooms and begin bringing equipment into the first cleanroom (Phase 1) starting February 2027. After a period of equipment setup, it will add 60,000 wafers, and then has set a plan to sequentially increase new production by 60,000 wafers each in the next cleanroom every six months. At this rate, in the Yongin Phase 1 fab alone, new DRAM production capacity on the scale of 360,000 wafers per month will be added in the first half of 2030. There is also the Cheongju M15X fab currently under expansion. M15X will begin operation at 40,000 wafers per month in the second half of this year. Next year it will have a production capacity of around 80,000 wafers per month. Adding the 360,000 wafers from Yongin and the 80,000 wafers of new expansion at M15X, SK Hynix’s DRAM wafer input capacity is expected to reach around 1 million wafers per month around 2030 to 2031. In February, SK Hynix introduced its Yongin investment plan and stated that the Phase 1 fab would be composed of two structural frames and six cleanrooms. The timing of the first equipment move in was moved up from the originally planned May 2027 to February. While the company disclosed the investment amount and the building structure, this is the first time the production products and scale by cleanroom and the expansion pace have become known. Under the current plan, the new expansion items are all DRAM. NAND flash is reportedly planned to be pursued mainly through technology upgrades such as increasing the number of layers. One official in the equipment industry explained, “Wouldn’t it mean prepare yourselves because we will increase fast and by a lot?” The remark by SK Group Chairman Tae Won Chey at the Computex 2026 exhibition venue that “we will double total wafer production capacity within five years at full speed” is said to be linked with this kind of plan. However, because the plan is so aggressive, the partner companies are in a mood of cautiously watching whether it will be executed. In 2022, SK Hynix delivered the following year’s capital expenditure guideline to partner companies, then in the fall of that year notified them that it would sharply reduce order volumes. Some partner companies that had trusted the guideline and even purchased components took a direct hit to their cash flow. Also, the schedule of filling one cleanroom every six months could disrupt the entire schedule even if just one type of equipment comes in late. One partner company official showed a cautious response, saying, “In the short term it is certain that investment will increase, so it will act as a big positive for the equipment and materials industries,” and adding, “The entire plan will only be achieved if market demand backs it up.” That said, there are also many assessments that this expansion plan carries greater weight compared with the past in that the group chairman directly laid out the big picture. Earlier, Chairman Chey said at the Computex venue, “Prices suddenly jumping or surging can harm overall sustainability,” and “For the entire ecosystem, more sustainability is needed.” There is an interpretation that this kind of remark expressed the will to push ahead with expansion without being swayed by short term price fluctuations.
English
16
42
392
137.2K
GeekPark
GeekPark@GeekParkHQ·
@VaibhavSisinty Small correction because details matter: OpenAI’s docs frame this as eligible free daily usage for shared traffic, not a universal “free $50k” button. Still interesting though: the real deal is subsidized compute in exchange for better eval/data loops.
English
0
0
3
2.8K
Vaibhav Sisinty
Vaibhav Sisinty@VaibhavSisinty·
I just found out OpenAI gives you $50,000 in free API credits if you do one thing in settings. 🤯 It's called the Data Sharing Program. No free trial exists for the API. But this does. Go to your OpenAI Dashboard → Data Controls → Sharing. Opt in. You get $50,000 in credits for their latest models or 2.5 million tokens for the rest. The catch: your data gets used by OpenAI for training and improvements. So don't use this for client work or anything sensitive. Use it for learning, side projects, and experiments.
English
79
105
1.3K
229K