DeepSeek followed its permanent price cut with Reasonix, a native coding agent built around high caching and low cost — the open-weight frontier is now competing on agent tooling, not just model price. The compute-economics story sharpened: new data shows memory has grown to nearly two-thirds of AI chip component costs, reframing the supply bottleneck from logic to DRAM and HBM. Agent reliability stayed under scrutiny, with a study on 'constraint decay' showing how LLM agents degrade on backend code generation as requirements accumulate. Google acknowledged it is navigating AI security in real time, and consumer AI kept getting stranger — an Amazon Bee wearable and a fleet of meal-making robots in San Francisco's Tenderloin. On the research side, agent-skill work is maturing from hand-crafted prompts toward a systematic discipline.
DeepSeek shipped Reasonix, a native coding agent optimized for caching and cost, extending its price-led pressure into agent tooling. In parallel, new data shows memory has grown to nearly two-thirds of AI chip component costs — the hardware bottleneck is shifting from logic to memory, which reshapes where the next capacity constraint bites.
DeepSeek released Reasonix, a native coding agent built around aggressive caching and low cost (445 HN points).
Epoch data shows memory now accounts for nearly two-thirds of AI chip component costs — the bottleneck is shifting from logic to DRAM/HBM (299 HN points).
A widely-shared study on 'constraint decay' shows LLM agents degrade on backend code generation as requirements pile up, and two papers push agent-skill construction from one-shot prompts toward a systematic, deep-learning-like discipline. The thread of the whole stretch: agents work in demos, strain under real accumulated constraints.
Study showing LLM agents degrade on backend code generation as constraints accumulate across a task (179 HN points).
Google acknowledged it is figuring out AI security in real time, an Amazon Bee wearable drew the now-standard intrigue-and-unease reaction, and a fleet of robots began making meals for a San Francisco nonprofit. The consumer and physical edge of AI keeps advancing faster than the norms around it.
Reporting on how even the largest platforms are improvising AI security responses as novel abuse patterns emerge.
Hands-on with Amazon's Bee always-listening AI wearable — useful but unsettling, the recurring verdict on ambient AI hardware.
A robotics deployment is preparing meals for a San Francisco nonprofit — embodied AI landing in social-service operations.
A widely-shared study on 'constraint decay' shows LLM agents degrade on backend code generation as requirements pile up, and two papers push agent-skill construction from one-shot prompts toward a systematic, deep-learning-like discipline. The thread of the whole stretch: agents work in demos, strain under real accumulated constraints.
Systematic study of how agents distill, store, and consume model-generated skills from past experience.
Moves agent-skill evolution from loosely-controlled self-revision toward a managed optimization process.
Lens shows a 3.8B text-to-image model matching or beating much larger systems on a tighter training budget, PiD speeds high-resolution latent decoding, and a Shannon-channel view of LLMs tries to explain the non-monotonic scaling behavior that simple power laws miss.
A 3.8B text-to-image model competitive with or surpassing state-of-the-art systems at a much lower training budget.
Recasts LLM scaling through a noisy-channel lens to explain non-monotonic phenomena that monotonic power laws fail to capture.
Speeds high-resolution latent decoding for text-to-image systems using pixel diffusion.