Liquid AI LFM2.5-24B: 2B Active MoE on Apple M5 Max
Liquid AI claims 184 tok/s on Apple M5 Max with 2.45B active compute, but LFM2.5-24B requires 15GB RAM, crushing 16GB laptops and lagging on SWE-bench.
EyesTech: Frontier AI, Silicon & Defence Technology
Independent analysis, benchmarks, and field intelligence on AI models, chips, computing, and defence systems.
OpenCode 2.0: MCP Code Mode Compression vs. Daemon RCE
OpenCode 2.0 delivers 86.8% MCP token compression via Code Mode, but daemon configuration flaws and .well-known stagers expose enterprise RCE risks.
Grok Voice 3.0: 810ms Latency, Architecture & Reality
Grok Voice 3.0 hits 810ms latency and 86.9 on S2S benchmarks at $0.08/min, but real-world mobile jitter and echo feedback erode its synthetic speed.
Anthropic Scion: Inside the 30,000 Agent Merges Monorepo
Anthropic replaced Git with Scion for 30,401 daily agent merges, dropping status to 19ms and saving 25% fleet CPU amid Meta Mononoke prior art debates.
REA: What AI Reverse Engineering Can Actually Prove
REA connects coding agents to reverse-engineering tools. Learn what file hashes, runtime captures and behavior checks prove before rebuilding a feature.
Timescale PRISM: Run 100M Vectors on 1/4th Postgres RAM
Timescale pg_vectorsearch and PRISM replace HNSW graphs with RaBitQ-quantized partitions, serving 100M vectors at 21ms on 128GB RAM with 51x faster builds.
Muse Wonder vs Claude Sonnet 5.5: The Agentic Cost Trap
Claude Sonnet 5.5 beats Muse Wonder on SWE-bench (57.2% vs 48.2%) while halving token costs ($2/$10 vs $4/$20) and avoiding Active Focus retry loops.
Cursor The X Subscription: Grok Lock-In & $195/mo TCO Shock
Cursor’s X Subscription bundles Grok for $30/mo but pushes Claude to API rates, causing an 800% monthly TCO spike for developers running coding agents.
Anthropic OSS Scanner: Unreviewed Zero-Day Fast Track
Anthropic launched its Cyber Mission and free OSS Scanner, beaming unreviewed zero-day PoCs directly to maintainers with a 90% precision target across code.
Google TPU v7 Ironwood Architecture: 1.77 PB Memory & OCS Breakdown
Google’s TPU v7 Ironwood scales to 9,216 chips and 1.77 PB of HBM. We examine optical circuit switching, long-context KV cache demands and memory limits.