The Pulse — August 2, 2026
The signals that entered our radar, organized with sources and context to understand what changed.
Gemini Robotics 2: whole-body control, embodied reasoning, and on-device adaptation
WHY IT ENTERED THE RADARGoogle is splitting the robot stack into a VLA controller, an embodied-reasoning planner, and an on-device model. It claims new robot bodies can be adapted in a few hours with fewer than 200 examples, while ER 2 handles multi-minute, multi-step tasks and multi-robot coordination.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“AI robots just got a brain, a body, and an offline mode — why that three-layer stack is the important part.” Show planner → motion model → hardware as the visual.
Claude Opus 5 is now the default Opus model in Claude Code — with 1M context and a new fast mode
WHY IT ENTERED THE RADARAnthropic positions Opus 5 as near-Fable capability at half the price; the Claude Code release makes it the default Opus model, exposes a 1M-token context window, and lists $10/$50 per million input/output tokens for fast mode. The story is less “new benchmark winner” and more “long-running agent work is becoming a default developer workflow.”
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The real Opus 5 upgrade isn’t IQ. It’s that coding agents can hold an entire project in context — but should they?”
Gemini 3.6 Flash: efficiency becomes a competitive feature
WHY IT ENTERED THE RADARGoogle says 3.6 Flash uses 17% fewer output tokens than 3.5 Flash, while listing a 1M-token input window, 64k output, tool use and computer use. Its published table frames the model as a practical agent worker, not a cheap “lite” model.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The next AI race is not only smarter models; it’s models that finish the same job with fewer tokens.” Explain why output-token efficiency directly affects latency and agent cost.
Seedance 2.5 moves AI video from clips toward editable scenes
WHY IT ENTERED THE RADARSeedance 2.5 generates up to 30 seconds of synchronized audio-video per pass, supports multi-round extensions, and accepts up to 30 images, 10 video clips, and 10 audio clips as references. Timestamp-level editing and stronger reference control are the creator-relevant claims—not just prettier generations.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“AI video is becoming an editing system, not a prompt slot machine.” Demo the production workflow: references → one-take scene → targeted edit → extension.
Kimi K3 on AMD MI355X: memory capacity can beat the CUDA default for giant MoE inference
WHY IT ENTERED THE RADARWafer reports Kimi K3 (2.8T parameters) reaching ~952 aggregate tok/s per MI355X node and 118 tok/s single stream. Its claimed performance-per-dollar advantage over B300 is driven by 288GB HBM per GPU and lower rental cost; it also shows how a small ROCm fix plus speculative decoding and an attention-kernel shape workaround materially changed results.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“The GPU winner for giant open models may not be NVIDIA.” Lead with the caveat: these are one vendor’s measurements, then explain why memory and prefill matter as much as headline decode speed.
A 2.78T Kimi K3 model running on one CPU with 8.24GB RAM — technically real, practically slow
WHY IT ENTERED THE RADARThis C99 project streams routed experts directly from a 1.56TB checkpoint, keeps only a working set in RAM, and reports byte-identical output across memory budgets. The 8GB preset takes roughly 33 seconds per token: it is not a consumer serving recipe, but an unusually clear demonstration of how MoE architecture changes the memory problem.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“A 2.8-trillion-parameter AI on 8GB of RAM? Yes—at one word every half minute.” Explain the difference between fitting, running, and being usable.
Mference brings SSD-streamed MoE inference to Apple Silicon
WHY IT ENTERED THE RADARThe Swift/Metal runtime keeps the shared core and cache resident while streaming selected experts from SSD. It reports Gemma 4 26B at ~2GB memory and Qwen 3.6 35B at ~1.45GB, plus an experimental DeepSeek-V4-Flash path. The key practical signal: local inference tools are increasingly architecture-specific rather than generic wrappers.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Your Mac’s RAM is no longer the full limit for local AI—but your SSD speed becomes part of the model.” Make the trade-off tangible: RAM, disk footprint, tokens/sec.
Open Agent Teams: a tiny protocol for reliable multi-agent coding workflows
WHY IT ENTERED THE RADARThe upstream skill proposes detached tmux executors, a result file, and a file-based done sentinel instead of fragile terminal signals. This is the unglamorous operational layer that turns “many agents” from a demo into something observable and recoverable.
SUGGESTED EDITORIAL ANGLEOpen original source ↗“Multi-agent workflows don’t fail because agents are dumb; they fail because orchestration is sloppy.” Show the simplest reliable pattern: coordinator, isolated worker, artifact, completion signal.