Stu

We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: rikiyat.github.io/blog/weigh…

Verizon's CEO said today that SpaceX will need at least 4-6 years and $300B-$500B to fully build out its @Starlink Mobile U.S. carrier network: "Their service just can't compete with a terrestrial network. In suburban and urban areas, where 98% of our revenues are, we're somewhere between 100x–1,000x more efficient than they are. We know exactly what they're going to do. We know who they're talking to on towers. There's no magic wand. There's no secret sauce. There's no shortcut. They're going to look at 65,000–80,000 macro sites. They're going to need backhaul, RAN, and a tremendous amount of spectrum. It's going to take a minimum of four to six years to fully build that out. It's going to cost hundreds and hundreds of billions of dollars, I assume somewhere between $300 billion and $500 billion. Then they need to build retail, customer service, back office systems, et cetera. All of that gets them basically the same network and same capabilities and same cost structure as what we have today. Then sometime over the next four to six years, they have the possibility of becoming the sixth, not the fourth, the sixth competitor in the market. You can be sure over these next three to five years or six years, we are not going to be standing still. Four to five years from now, we will be two to three generations better than where we are today. We will have high-speed broadband gig-plus coverage across the country with our millimeter wave. We are ahead of most people and already experimenting with what 6G applications will look like." (via Verizon's Q3 earnings call earlier this morning)

increasingly concerned that ai is revealing the mathematical world as negatively curved. instead of everyone wandering about in idea-space and clustering around the same famous conjectures (local positive curvature), ai is pushing out the boundaries rapidly and revealing that the space of mathematical ideas is ~hyperbolic (negative curvature). though this means there's lots left to discover, it also means that small separations between people/agents will lead to wandering alone. we may be emerging into an idea-space where there's so much to see and do but everyone is alone in a private heaven, unable to share their subtle joys with others...

We've received overwhelming requests from the security community for an abliterated DeepSeek V4.1. So we abliterated DeepSeek V4.1 Flash (1027 checkpoint) ourselves. Now live on OrcaRouter's Cyber Tier, alongside OrcaCyber Zero and OrcaCyber Ultra. - 1M context - Reduced refusals for security research - Built for advanced cybersecurity workflows The best cyber defense is continuously scanning your own production infra with the best cyber models. Different models catch different vulnerabilities. Use them all. Available to verified security researchers. orcarouter.ai/models/orca/or…

turns out it classifies what audio it hears, before reporting back to samsung. sensor preferences, dog-barking and baby-crying event counters, and camera streaming duration are built into telemetry paths.

Introducing the people search API 1 billion people, one API > decision makers to sell to > founders to build with > engineers to hire Let your agents find them.

Notice how none of them are C++ services. C++ remains the safest language. The undefeated king still sits on the throne.