Max Ng (Hululu) (@maxwindiff) header
Max Ng (Hululu) (@maxwindiff) profile picture

Now I have the full picture. @BedrockRobotics

Santa Clara, CA
Joined October 2007
Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)Photo by Max Ng (Hululu) (@maxwindiff)
Max Ng (Hululu) retweeted
Some want such authentication to be allowed to use frontier models. This is a bad idea. Requiring people's real identity to use Internet is always a bad idea, whether it's access to porn, protecting children, or anything else.
2
26
1
89
5,356
Max Ng (Hululu) retweeted
3
3
73
2,890
Max Ng (Hululu) retweeted
Construction is wild and fascinating. Here's an example from one of our sites where the excavators are building "mats". They use logs from the site to float on top of the muddy ground, inch-worming a makeshift road as they go. Excavation is manipulation!
2
3
51
52,864
Max Ng (Hululu) retweeted
A few final thoughts on the events of the past few days: First, by far the worst result of the DoD contract saga has been the supply chain risk designation for Anthropic. In the worst-case scenario (Hegseth really did mean "no one who does business with the U.S. military may conduct ANY COMMERCIAL ACTIVITY with Anthropic"; the courts somehow uphold this or refuse to stay Anthropic's supply chain risk designation while they consider the issue, potentially for many months), this could spell death for the company. The DoD may have just killed the goose that lays the golden eggs, to the delight of our geopolitical adversaries. But it's worse than that. As of Friday, every single investor in data center infrastructure suddenly lives in a world where the U.S. government feels that it would be acceptable, in one fell swoop, to significantly undermine the entire business model of a particular frontier lab. Congratulations, data centers have now become more expensive to build - for everyone (no, not just Anthropic). This also percolates to the business of every U.S. tech company generally (not limited to AI). Even those who dislike Anthropic need to understand that U.S. businesses, and investors in U.S. businesses, require a climate of business certainty. The U.S. government, by singling out a particular U.S. company in this fashion over a mere contractual disagreement, has *already* done great harm to the U.S. business environment. Especially if the courts don't swiftly strike this down, we're in for a really bad time. Second, turning now to OpenAI's contract with the DoD, I invite everyone to be more pragmatic. The alternative to OpenAI providing services to the DoD was never going to be "DoD has no access to powerful AI and can't use it in automated AI weapons or domestic mass surveillance". The DoD's alternative to ChatGPT was MechaHitler. I leave it up to the reader to decide which of these two options is preferable. Third, it is critical to understand that the U.S. military *needs* access to powerful AI. Our geopolitical adversaries will have access to powerful AI. Our geopolitical adversaries will use it on the battlefield. Terrorist organizations could also conceivably obtain access to powerful AI. The United States must, at the very minimum, be able to play defense on this field. But there is no way to do so without access to *truly* frontier models - and that means either Claude, ChatGPT or Gemini. Fourth, and as an extension of the third point, our geopolitical adversaries WILL use fully automated AI weapons. They WILL use them without a human in the loop. They WILL do this before it is reasonably safe to do so. The cost of a human life in countries like Russia approaches zero. This means that we will very soon live in a world in which there shall exist fully automated AI weapons. We get to choose only whether they are wielded solely by our adversaries or also by the U.S. military. I personally hate the idea of fully automated AI weapons without a human in the loop, and I would strongly prefer it if no country developed them, ever. But they *are* coming, and it is useless to hide one's head in the sand and pretend that they are not. Given this, I personally fully support Anthropic's stance on helping the DoD develop fully autonomous weapons, but only when the models become sufficiently powerful to be used in this fashion in a *very* safe manner.
33
39
19
365
93,427
Max Ng (Hululu) retweeted
Experiments + Open weights from our Image-Video VAE work.
Turns out better VAE reconstruction quality can make your diffusion model worse. We learned this the hard way over 4 months. Writeup + open-source Image-Video VAE model. linum.ai/field-notes/vae-rec…
1
20
2,415
Max Ng (Hululu) retweeted
Replying to @OpenAI @Ginkgo
3
75
1,384
Today Thinking Machines Lab is launching our research blog, Connectionism. Our first blog post is “Defeating Nondeterminism in LLM Inference” We believe that science is better when shared. Connectionism will cover topics as varied as our research is: from kernel numerics to prompt engineering. Here we share what we are working on and connect with the research community frequently and openly. The name Connectionism is a throwback to an earlier era of AI; it was the name of the subfield in the 1980s that studied neural networks and their similarity to biological brains. thinkingmachines.ai/blog/def…
230
1,233
449
7,560
3,551,877
Max Ng (Hululu) retweeted
Replying to @francoisfleuret
China is completely out of the race, it’s going to take them decades to catch up with the quality of our *current* AI regulations and by then we will be 3/4 generations of regulation *ahead*
4
8
538
Max Ng (Hululu) retweeted
Replying to @Molson_Hart
It does sometimes feel like much of America has its head in the sand. They think China is Temu. Feels like they aren't paying attention. Feels like the US is at a crossroads.
1
1
7
453
Max Ng (Hululu) retweeted
I have written up some thoughts on AI doom and x-risk, with a focus on why I find a fast-takeoff scenario highly implausible, and why I think that any world model derived from written language is insufficient for advances in real-world technology: addxorrol.blogspot.com/2024/…
21
55
14
211
43,222
Max Ng (Hululu) retweeted
Ok, shaking my Twitterverse network. I'm organising the Performance track for QCon London next year, and I want to make sure I don't bias the track construction too much. If you do performance work, please reply to this tweet or DM me :-) - would love to hear about you!
10
48
69
22,268
Max Ng (Hululu) retweeted
🚨 Last chance alert! 🎉 A handful of #CUBECON full passes are up for grabs! To secure yours, email us at [email protected] .🎟️ These passes aren't currently available for purchase through GHC, but we'll reserve one just for you! Act fast: CubeCon in T-minus 13 days! 🌟
5
5
12
9,393
2029 — The Compute Enforcement Administration (CEA) announces a record seizure of 1750 teraflops of unregistered compute. "Compute like this isn't for streaming cat videos—it's a weapon," said Agent Yudkowsky. "We're fighting a war against digital fentanyl, and today, we won."
24
136
20
1,072
189,105
Max Ng (Hululu) retweeted
lambda lambda fold fold *map* *map* *map* *map* mmm Haskell so good yes yes yes
15
87
6
1,025
67,055
Max Ng (Hululu) retweeted
the AI doomers can't even coordinate on releasing open letters, but sure, we can get the PRC to work with us to stop AI development anywhere in the world
2
9
265
Max Ng (Hululu) retweeted
No Relaxed Friday today. Can't relax knowing that Rob, founder of @17Lands had to pay a serious sum to run 17L from his own pocket. He puts enough work into already, to have to pay on top of that. So pls retweet and consider becoming patreon.com/17lands Every $ counts!
10
52
8
117
61,604
Max Ng (Hululu) retweeted
After taking a year break, I am back #MagicthGathering Hall of Famer looking for sponsor Committed to coming four Japan RC and two Europe/APAC Pro Tour Retweet appreciated Will follow back all account retweeted this tweet 🙏 Welcome to DM for offer🫰
6
93
10
278
49,494
Max Ng (Hululu) retweeted
We're working on a new project as a follow-up to Manifold Garden. We are hiring for a number of positions, all of them remote: • Programmer • Graphics Programmer • Environment Artist • Graphic Designer • Illustrator More info here:workwithindies.com/work-with… #gamedev
16
143
4
368
57,488
Max Ng (Hululu) retweeted
📢 ANNOUNCEMENT 📢 For the month of December we're doing something we've never done before. For every listen we get from Dec 1 - Dec 31, we'll donate $0.25 to @DirectRelief! Share this with your friends and family be part of this with us!
1
7
2
16
Max Ng (Hululu) retweeted
Believe the science
8
24
7
172