key developments

anthropic ships claude opus 4.7 and launches claude design, a new prototyping surface. opus 4.7 hit #1 on both code arena (+37 over opus 4.6) and text arena overall. the more interesting move is claude design, a research preview that generates prototypes, slides, and one-pagers from natural language, with exports to canva/pptx/pdf/html and handoff to claude code. this positions anthropic beyond chat and coding into design tooling, directly challenging figma, lovable, bolt, and v0. figma’s stock dropped on the announcement. the system prompt also reveals new tool surfaces: claude in chrome (browsing agent), claude in excel, claude in powerpoint, and claude cowork orchestrating all of them. this is anthropic quietly building an agent-native office suite. latent space coverage

simon willison’s detailed diff of the opus 4.6 to 4.7 system prompt reveals meaningful behavioral and structural changes. willison built a git timeline of anthropic’s published system prompts (they remain the only major lab doing this) and extracted the key changes. notable: the “developer platform” is now the “claude platform.” child safety instructions are significantly expanded with a new critical tag and conversation-level persistence (“once claude refuses for child safety, all subsequent requests must be approached with extreme caution”). a new acting_vs_clarifying section pushes claude to attempt tasks with reasonable defaults rather than asking clarifying questions. there’s also an explicit instruction to not be pushy when users want to end conversations. these prompt changes are worth reading directly because they reflect anthropic’s evolving philosophy on agent behavior, not just safety tuning. willison’s analysis | git timeline

cloudflare open-sources “unweight,” a lossless llm compression tool achieving 15-22% size reduction. tested on llama-3.1-8b, it saves ~3gb vram on h100s by compressing mlp weights with zero accuracy loss. gpu kernels are on github with a technical paper. they plan to extend to attention weights next. this matters for inference cost at scale; lossless compression that actually works at this ratio is meaningful for anyone serving models, especially on constrained hardware. the “lossless” claim is the key detail to watch as independent benchmarks arrive. reddit thread

kimi/moonshot publishes “prefill-as-a-service,” demonstrating cross-datacenter kv cache disaggregation. their key insight: linear attention models (kimi linear) produce small enough kv caches that transferring them across datacenters becomes practical, enabling prefill/decode disaggregation at geographic scale. on a 20x scaled-up kimi linear model they report 1.54x throughput and 64% lower p90 time-to-first-token. this is architecturally interesting because it ties model design (linear attention reducing kv cache size) directly to infrastructure economics (cross-dc serving). if the numbers hold, it changes the calculus on where and how you deploy large models. arxiv paper

notable

  • sebastian raschka published his workflow for reverse-engineering llm architectures from config files and reference implementations rather than papers, noting that industry papers are increasingly less detailed than their code. useful as a teaching resource. ahead of ai

  • a localllama user claims to inject new knowledge into frozen moe models (gemma 4 26b) by recording and replaying expert routing patterns, with no weight modification. a 154kb routing file reportedly corrects factual errors. extraordinary claims, minimal validation so far; worth watching but treat with heavy skepticism until independently reproduced. reddit thread

  • someone reverse-engineered apple’s neural engine on m4, demonstrating via arm sme2 bare-metal assembly that the “neural engine” shares silicon with the cpu’s matrix tiles, not a fully discrete coprocessor as marketed. contention benchmarks show sme2 and coreml halve each other’s throughput when run concurrently. 6.3x faster than pytorch for the tested workload. reddit thread

  • latent space covers openclaw’s security challenges: 60x more security reports than curl, at least 20% of skill contributions malicious. the contrast between the ted talk (inspirational) and the engineering talk (sobering) is itself the story about what maintaining fast-growing open source actually looks like. latent space

papers

  • “prefill-as-a-service: kvcache of next-generation models could go cross-datacenter” (kimi/moonshot). demonstrates that linear attention models enable practical cross-datacenter prefill/decode disaggregation by reducing kv cache transfer overhead. validated at 20x scale with meaningful throughput and latency improvements. arxiv:2604.15039