Maybe it's a spectrum that spans from maximum utility to maximum efficiency. Humanoid robots can do a lot of tasks but in a slightly less efficient regime, while hyper specialized robots that will do the same very efficiently in industrial setups.

According to this paper, the Unitree G1 humanoid robot secretly and continuously sends sensor and system data to servers in China without the owner's knowledge or consent. arxiv.org/abs/2509.14139

It's also another reminder for all IoT manufacturing equipment that uses the wifi

Future needs not "leaders" like this Almost all sama and dario's interviews and tweets: - we're amazing, everyone should get ready to die; - disseminate fear through dystopian futuristic visions where they win and all die - ask for regulators to help kill all competition - share

The human brain works more like a diffusion model than an a next token predictor. We start with a pure and unfiltered thought. Then, as we speak or write, we iteratively diffuse in it caveats, disclaimers, and tweaks to make it crowd-pleasing, normative and defendable.

The world is about to find about UFOs, alien materials and tech. Meanwhile 7B people still have fun watching movies with guns, betrayal, revenge and ... swooords?

Sonnet 4.5 rate limits are absolutely horrendous

Model merging becomes a thing. Get multiple same-arch models trained separately on specialized datasets aggregated into a bundled model. The merged model retains most of the predictive power. Expert merging is already here: arxiv.org/abs/2509.25712 Can't wait for distillation

It looks like a non dogmatic approach to operational efficiency. I'm sure nobody cries over missed opportunities to work the docks.

small language models are already great at delivering value in offline/ on-prem/ air-gapped environments. Instead of solving the world in one prompt you need to break it down into smaller work and distribute it. But it works. We've written about it here:

Another example of small portable models, highly specialized that can outperform multi-modal large models. In this case OCR. huggingface.co/PaddlePaddle/P…

AI browsers are fundamentally broken by design. This UX form factor ingtmi

The doers are the thinkers

2025 Tsinghua uni admission stats: 49% of students dodged the Gaokao and got in: 24% - basic discipline strength plan ie. Score in top 1% in Stem - exceptional talent discovered before Gaokao. University/ national special plan - 15% ie. Underprivileged students Early

A great in-depth article about nano-imprint lithography NIL, a competing Canon alternative to ASML's EUV. newsletter.semianalysis.com/p/nanoimprint-…

I used GrokCodeFast1 to refactor a transpiled JavaScript file. It COMPLAINED A LOT about the poor naming choices and dumb obfuscation, but then it mustered its nerve and eventually got the job done. Some of these models complain like humans. Their effort is actually constant

Substrate, SV's semiconductor manufacturer announces X-ray lithography. SemiAnalysis does an excellent job describing it in-depth newsletter.semianalysis.com/p/how-to-kill-…

When it comes to investing SEC and FINRA got it wrong. The listed companies have a small risk of going bust. Everyone should be allowed to invest. Unsupervised kids should be allowed to invest after they can read and write. Most importantly, there should be mandatory tests of

This is a cool product launch. Now I'm excited.