Summary
The afternoon's one real research signal is Apple's third-generation Foundation Models, unpacked in two technical tweets from Hugging Face's eliebakouch (a cluster of 2) following the official Apple ML Research post. The headline for a routing-and-efficiency reader: the on-device 20B model is a sparse MoE that makes an early routing decision once per prompt, choosing both which experts fire and how much compute to allocate (1 to 4B active params), and the server model is trained with "parallel track" (PT) parallelism instead of expert parallelism. That is the slot's standout. The second worthwhile item is Kilo Code's Agent Manager, which gives every parallel agent its own git worktree to kill merge conflicts and context bleed. The rest is Scobleizer's WWDC-watching commentary, a consumer AI assistant promo (Super), and off-topic local-news posts, all noise.
Posts
- Apple third-gen Foundation Models: MoE with once-per-prompt early routing + adaptive compute (cluster of 2: @eliebakouch, @eliebakouch · Apple ML Research). Local models are a 3B dense model plus a 20B-total MoE that activates only 1 to 4B params; an early router decides once per prompt which experts run AND how much compute to allocate, though it's unclear if this is layer-selection or expert-selection. The AFM server model uses parallel-track (PT) parallelism, syncing once every few layers rather than EP's per-layer syncs. Prompt-level adaptive compute is exactly the kind of route-once-then-commit design tracked in LLM routing.
- Kilo Code Agent Manager: worktree-isolated parallel agents (@kilocode · blog). Each parallel agent gets its own git worktree (separate directory, separate branch, sees only its own files) so multiple agents run at once without overwriting each other or bleeding context. A concrete take on the orchestration problems in multi-agent systems, from the same Kilo team behind last week's model-task routing audit.
- Super: consumer AI assistant with "its own apps" (@Scobleizer · getsupers.com). Scoble reposts a launch claiming a $0 "computer-use cache" that replays trained app actions to run Uber, DoorDash, and 1000+ apps from iMessage. Promo. Skip.
- WWDC reaction commentary (@Scobleizer). Scoble's running notes on Apple's AI-heavy keynote, mostly about a rebuilt Siri for everyday users versus deeper computer-use plays like Cluely. Color, not signal.
- LadderMan: zero-shot sim-to-real humanoid ladder climbing (@Scobleizer). A perceptive system for humanoids climbing and manipulating on ladders straight out of simulation. Robotics hardware, outside the wiki's focus.
- Datacenter siting, power, and SpaceX IPO musings (@Scobleizer). Scoble on communities blocking datacenter builds, power constraints, and fiber vulnerability, pointing at a SpaceX IPO page. Op-ed, no concrete claim.
- Off-topic local posts (@AustinJustice, plus a dog-sitter-startup anecdote and an AI-film note from @Scobleizer). Austin court news and personal anecdotes with no AI substance. Skip.