Bend 2 ships a language that makes AI's bugs fail to compile
Taelin's proof-checked language runs C-fast and 100x on GPUs; plus a ternary packing win for local models and Claude now leads 26% of Anthropic's R&D.
Copy markdown
Your invariants become laws the AI can't break
Bend 2 lets you declare rules in a LAWS.bend file and the compiler proves no code can violate them — merging a bug becomes "mathematically impossible: it is a theorem." In the repo's game demo, an AI agent that tried to add board-wraparound was blocked and forced to retry until its change kept "winning is impossible" true. For you: pin the invariants an AI coding agent literally cannot merge past.
C on the CPU, ~100x on the GPU, proofs in a second
Bend targets C-class single-core speed and fans out across thousands of GPU cores with no hand-written kernels, and it checks proofs in about a second instead of the minutes classic provers take. It's Apache-2.0 and installs with one curl line on Linux/macOS (WSL on Windows). Caveat: it's young — expect bugs and thin support for strings, floats, and libraries.
Elsewhere: a ternary trick shrinks local models below 1.5 bits
A new paper, BITCOS, beats standard five-trit packing on 26 of 29 models — as low as 1.485 bits per weight — by exploiting that up to half of ternary weights are zero. It reports up to 1.28x faster matrix-vector math and 1.18x/1.27x end-to-end throughput on CPU/GPU with AVX-512, AVX2, and Intel Xe2 kernels. No code yet, but a technique to watch for squeezing BitNet-style models onto hardware you own.
Elsewhere: Anthropic says Claude now leads 26% of its R&D
In a Sept 17 transparency report, Anthropic said Claude leads about 26% of its model R&D — up from zero in February — with roughly 90% of R&D done alongside Claude and about 30,000 internal agents running. It's urging every lab to publish the same metric as a public gauge of "recursive self-improvement." No new API, but a marker of how fast the tools you build on are compounding.