AI funding hit records, a new search tool skips embeddings, and a Qwen model runs on Raspberry Pi.
Global AI venture funding reached all-time records yesterday for startup deals and valuations, providing market context for founders and teams planning their next moves. In other news, Mantic.sh offers a way to search large filebases fast without embeddings, a different approach for specific search needs. We also saw a 30B Qwen model running in real-time on a Raspberry Pi, showing what's possible for on-device ML.
Global startup funding surged to $425 billion in 2025, up 30% year-over-year, marking the third-highest year on record. AI captured $211 billion, half of all funding, driven by record deals like OpenAI's $40 billion round. Analysts expect this concentration of capital into AI to continue.
Read full story→A new framework proposes building ethical AGI by allowing it to hold contradictory values without internal conflict. It uses "non-dual motivation" with paraconsistent logic and nonlinear resonance, organizing around axes like "Acceptance ↔ Compassion." This aims for stronger value alignment and internal coherence.
AI agents running code can execute anything from harmless tests to malicious installs. This creates a security gap. A new field guide covers architectures and best practices for sandboxing AI agents, ensuring secure programmatic access.
ByteShape's Shapelearn method achieved 8 tokens/second on a Raspberry Pi 5 with a 30B Qwen model, maintaining 94.18% BF16 quality. This quantization approach optimizes for memory constraints and learns optimal bitlengths, outperforming Unsloth and MagicQuant.
Daniel Miessler reviewed his AI predictions from 2016-2026, noting his accuracy on "ecosystem over models" – human needs are more predictable than tech. He also details misses, like Apple's AI timeline and Hollywood disruption. It's a look at what makes tech forecasting hard.
Advanced AI models like Opus 4.5, GPT-5.2, and Gemini 3 are now generating the majority of software code. This shift diminishes the value of traditional coding skills, moving engineers towards product-mindedness and tech leadership, though it also raises concerns about increased code complexity.
AI doesn't understand product intent; it pulls answers from existing info. This makes product explainability a core problem for product teams. The article argues explainability needs to be a system, not just docs, with a single source of truth and update triggers.
Mantic.sh is a code search engine for AI agents that ranks 480k files in 0.46s. It works without embeddings or vector databases, reducing context retrieval overhead and token usage. The tool infers intent from file structure, integrates with platforms like Cursor, and runs locally for privacy.