Industry · Software · Tech Policy · Events
How software gets built: developer tools, engineering practice, platforms and the open-source economy, from The Silicon Ledger.
Claude Fable 5.1 wins a new coding benchmark with a 38.8% success rate, but that means it fails nearly two-thirds of the time when tested on proprietary code from real companies.
Amazon Web Services, a founding member of the OpenSSF, is helping maintainers overcome the computational burden of securing open source projects through its promotional credit program, which covers testing, artifact distribution, and enterprise-grade security tools.
Perplexity's Portable Computer now runs on Windows PCs with compatible NVIDIA GPUs, bundling model orchestration, tool integration and sandboxing into a single application. The catch: users need at least 24GB of VRAM.
GitClear's analysis of 623 million code changes reveals that AI-assisted development boosts output modestly while triggering a dangerous shift toward copy-paste coding and away from refactoring, threatening long-term maintainability.
At the LEAP East conference, the open-source operating system project showcased new clustering technology for distributed AI workloads and an agent-based desktop assistant, signaling a shift toward building regional alternatives to Western AI infrastructure.
Large language model inference charges accumulate quickly on repeated queries. A layered caching approach—exact matching, semantic similarity, and hybrid methods—can slash costs by more than half while reducing latency.
As Chip Huyen returns to P99 CONF, her framework for reducing AI inference costs without hardware upgrades remains relevant—though the rise of agents and reasoning models has shifted how practitioners think about the problem.
Cohere unveiled North Small Translate, a mixture-of-experts model that outperforms competitors like DeepL and Google Translate on standard benchmarks while addressing data sovereignty concerns for regulated industries.
