Neetx/ai-research-radar

Tracks AI-ecosystem trends (research + engineering) and curates a study shelf for an AI researcher / AI-systems engineer who works with AI daily — generated from TRENDS.md.
**Since last scan (2026-10-16):** Meta's Muse Spark Open — an 816B/42B-active open-weight MoE billed as the most capable US open-weight model to date, engineered to run end-to-end on two Mac Studios — lands as evidence on both the open-weight wave and small/CPU-first models · Self-Evolving Agent Harness and IBM's Granite Switch 4.1 (an 8th independent org on the fast-decision-model pattern) both become agent-harness evidence · queue burndown drops the stale 10-09 below-bar batch while adding 4 new watch items; Muse Spark Open joins the study shelf.
⭐ Pinned topics
| trend | stage | latest signal |
|---|---|---|
| ⭐ Small / CPU-first models | 📈 accelerating | 2026-10-15 |
| ⭐ Ultra-low-bit quantization | 📈 accelerating | 2026-10-10 |
| ⭐ Latent-space reasoning | 🌱 emerging | 2026-10-14 |
| ⭐ Latent-space communication | 🌱 emerging | 2026-09-30 |
Trends
🌱 8 · 📈 11 · 🚀 0 · 🌊 0 · 🏔 4 · 📉 0 · 💤 0
| trend | stage | latest signal |
|---|---|---|
| ⭐ Small / CPU-first models | 📈 accelerating | 2026-10-15 |
| Prefill/decode disaggregation | 📈 accelerating | 2026-10-14 |
| On-policy distillation | 📈 accelerating | 2026-10-14 |
| Subquadratic/sparse attention | 📈 accelerating | 2026-10-12 |
| ⭐ Ultra-low-bit quantization | 📈 accelerating | 2026-10-10 |
| World/action models from video | 📈 accelerating | 2026-10-07 |
| Verifiable RL environments | 📈 accelerating | 2026-10-07 |
| Multi-agent engineering as product | 📈 accelerating | 2026-10-07 |
| Diffusion language models | 📈 accelerating | 2026-10-07 |
| MCP as the standard integration layer | 📈 accelerating | 2026-10-04 |
| Remote sandboxes for agents | 📈 accelerating | 2026-09-30 |
| ⭐ Latent-space reasoning | 🌱 emerging | 2026-10-14 |
| VLM-scaffolded embodied navigation | 🌱 emerging | 2026-10-13 |
| Deployment-grounded agent eval | 🌱 emerging | 2026-10-06 |
| Parametric injection | 🌱 emerging | 2026-10-01 |
| ⭐ Latent-space communication | 🌱 emerging | 2026-09-30 |
| Dense credit assignment for agentic RL | 🌱 emerging | 2026-09-30 |
| Agent payments (x402) | 🌱 emerging | 2026-09-30 |
| Decentralized multi-agent coordination | 🌱 seed | 2026-09-29 |
| Agent harness/runtime as self-improving object | 🏔 mainstreaming | 2026-10-16 |
| Open-weight frontier MoE wave | 🏔 mainstreaming | 2026-10-15 |
| Agent security: architectural turn | 🏔 mainstreaming | 2026-10-08 |
| AI agents doing open-ended research | 🏔 mainstreaming | 2026-10-07 |
Worth studying
- Muse Spark Open — Meta's 816B/42B-active open-weight MoE, engineered to run end-to-end on two Mac Studios via MLX rather than just a datacenter, with day-0 agentic demos.
- OpenCode 2.0 — a production-grade reference design for the architectural questions (client/server split, tool-composition layer, background orchestration) anyone building an agent harness has to answer.
- RoboJEPA: Scaling Robotic Latent World Models — a rare genuine scaling-law study (not just a bigger model) applied to embodied/robotic world models.