Involution retweeted
had this interaction with claude yesterday
me: what's the cutoff date for your training data?
claude: jan 2025
me: google us news events post jan 2025 so you're up to date
claude: *searching*
me:
claude: *searching*
me:
claude: jesus christ
it literally said "jesus christ"
Involution retweeted
If you thought the Dot-Com crash of 2000 was bad, or the financial collapse of 2008, you won’t want to miss the upcoming AI crash. @futurism again FTW on AI reporting.
Involution retweeted
Here's a question I know many are wondering about: why did China wait until now to use rare earths as leverage against the US? Why not in the first Trump administration when the US started the trade hostilities? Or when the Biden administration unleashed the chips export controls 3 years ago?
I just watched a fascinating explanation by a Chinese analyst and, unexpectedly, a big part of the explanation is... helium.
I had no idea but as he explains (source here: xiaohongshu.com/discovery/it…), all the way until 2022 China imported 95% of its helium and most of it was controlled by the US. Of the world's ten largest helium producers, four were American companies, and the remaining six all used American technology.
Helium isn't just a party balloons gas: it has plenty of industrial applications for things such as quantum computing, rocket technology, MRI machines, as a coolant for chip lithography equipment, etc.
In a nutshell what he's explaining is that with helium the US had an even stronger card to play if China ever used the rare earths card.
This raised huge alarm bells inside China. In an article published in late 2022 in the journal Frontiers in Environmental Science (frontiersin.org/journals/env…), several researchers from PetroChina’s Beijing-based Research Institute of Petroleum Exploration and Development stressed that China would be greatly affected if the US imposed a “stranglehold” blockade on helium exports.
So over the past few years there were gigantic efforts in China to break the "helium shackles," with seven helium extraction facilities going into production, and China also switching imports from the US in favor of imports from friendly countries like Russia.
China's research ecosystem also went into overdrive to find solutions to the helium dependency issues, with China's Academy of Sciences awarding its annual 2024 "Outstanding Science and Technology Achievement Prize" to a new helium extraction technology project (english.casad.cas.cn/newsroo…) because "these scientific and engineering achievements broke the long-standing monopoly of the US and ensured the security of China's helium resources" (guancha.cn/internation/2024_…)
The result: by the end of 2024 China had cut its helium dependence on the US to less than 5% (scmp.com/news/china/science/…). The "helium shackles" were broken.
That's what most people don't realize: power isn't about intentions or rhetoric - it's about what you can actually do. Many wonder why countries almost never retaliate when the US imposes sanctions or export controls. The answer is simple: they can't. They lack the alternatives, the technology, the supply chains.
China is the first country that systematically worked to eliminate every single pressure point, with humongous efforts. It's not just helium: it's chips, energy, telecommunication, pharmaceuticals, etc.
That's why the rare earth card can finally be played now. Not because China suddenly became aggressive, but because they have developed the capabilities to say "no."
Last word: as a European, this is both depressing and inspiring. Depressing because it highlights the immense magnitude of the task at hand to become genuinely sovereign and develop our own capabilities to say "no." Inspiring because China demonstrated that it can actually be done, and relatively fast if we execute competently. Although with the current crop of folks at the helm in Europe, that last part is admittedly a very, very big "if"...
Involution retweeted
Rude prompts to LLMs consistently lead to better results than polite ones 🤯
The authors found that very polite and polite tones reduced accuracy, while neutral, rude, and very rude tones improved it.
Statistical tests confirmed that the differences were significant, not random, across repeated runs.
The top score reported was 84.8% for very rude prompts and the lowest was 80.8% for very polite.
They compared their results with earlier studies and noted that older models (like GPT-3.5 and Llama-2) behaved differently, but GPT-4-based models like ChatGPT-4o show this clear reversal where harsh tone works better.
----
Paper – arxiv. org/abs/2510.04950
Paper Title: "Mind Your Tone: Investigating How Prompt Politeness Affects LLM Accuracy (short paper)"
Involution retweeted
it’s funny that the main argument for AI is usually not that it’s good or useful, just that it’s inevitable so we might as well learn to live with it. even its advocates talk about it like it’s a plague
Involution retweeted
My theory is that all the data centers being forced on us aren’t for future public demand, they’re for current military/govt demand—that AI is the backbone of the new surveillance state but they can’t say that so they pretend it’s because people love making videos of Pikachu.
Involution retweeted
“AI is the asbestos we are shoveling into the walls of our society and our descendants will be digging it out for generations” - @doctorow, speaking truth and taking no prisoners.
Involution retweeted
Prince actually refused to license the use of holograms in his likeness, calling them “the most demonic thing imaginable … If I was meant to jam with Duke Ellington, we would have lived in the same age.” He would have HATED this.
Involution retweeted
As this person says further down, the false promise of AI is that kids can learn and acquire skills without effort. But that's not how learning has ever worked, and the danger now is that the current generation of students will have no skills and know nothing.
I guess Trent Reznor isn't quite as alive as needed. As Alive As You Need Me To Be is the first song by an artist to truly grasp the AI era.
Involution retweeted
New drone technology startup in series A. They have a go-to-market plan. xitter.cf/Shinsho_ni/status/1974…
Involution retweeted
It's the hot take everyone is too afraid to make. Russia - a society of old people sending their tiny remaining youth to die on the altar of imperial nostalgia - is the kind of thing that will happen everywhere in the future because everyone will face similar demographics.
Involution retweeted
🎨 Super impressive. Researchers built an optical diffusion model that uses light, not heavy computer math, to generate images with quality similar to standard AI diffusion but at near-zero power.
This work shows a light powered image generator that produces images with almost no power, matching digital diffusion quality that usually needs 1,000 steps, by doing inference in 1 shot.
Diffusion image models normally start from random static and remove noise over many steps, so large batches burn compute and energy fast.
The UCLA team keeps the diffusion idea but moves the heavy work into light, so photons handle the operations while electronics mostly sit idle.
A compact digital encoder, trained on public image data, first builds the static pattern for the prompt, which costs a little energy upfront.
A spatial light modulator writes that pattern onto a laser beam, then a second programmed spatial light modulator converts the patterned light directly into the picture.
The key trick is that light propagation and these modulators carry out the required transforms physically, so power draw is mostly the laser and the panel drivers, not big matrix math on GPUs.
Tests covered faces, butterflies, and full color Van Gogh style scenes, and the optical output was comparable to a digital teacher diffusion model that needed 1,000 steps while the optical path needed just 1 pass.
Cutting thousands of digital updates down to a physical pass slashes energy and heat, so the carbon cost of mass image generation can drop a lot.
Speed and tiny power also point to practical uses like real time AR or VR imagery, and even on device generation for phones or lightweight AI glasses.
If it can be scaled, this looks like a clean path to scale generative imaging without chasing ever larger clusters, because it trades GPU cycles for cheap physics.
