The Rundown · episode
WW3 prep: bots, cameras, and Starlink drama
Maticus sifts through new open AI toolkits, camera bugs, robot bot drama, and a spicy read of Starlink and battlefield uses—tech optimism with troubleshooting detours.
Chapters
- 28:26 Maticus runs down new image and robot toolkits
- 35:43 Maticus highlights Moss Live for real-time voice
- 37:42 Maticus troubleshoots a camera defaults bug in Actorcore
- 42:06 Camera controls change but image stays dark
- 52:07 Core Actor Core bot is stuck; they debate giving it a skill
- 58:53 Maticus runs through this week's open-model boom
- 61:58 Maticus reads Elon on massive Starlink growth and Ship 42 ti
- 64:02 AI in warfare, long-context models, and memory tricks
What happened
Maticus runs down new image and robot toolkits
Maticus summarizes several new open models and toolkits: KH3 Image (an image generator/editor based on Minimax H3), Open RoboKit (four robotics libraries), and related open-source releases, and says they'll link the main pages in the description.
Maticus highlights Moss Live for real-time voice
Maticus describes Moss Live, a real-time multimodal conversational voice system that separates foreground conversation from background reasoning and can handle voice distinction and meeting participation; they note there's a demo and API but no model weights mentioned.
Maticus troubleshoots a camera defaults bug in Actorcore
Maticus explained a bug in Actorcore's Mission Control where hitting 'default' made a buddy camera image super dark and the gain/exposure controls and hotkeys either don't move or don't change the image reliably.
Camera controls change but image stays dark
Maticus walks through a technical problem where camera exposure and gain values change in the UI but the image remains dark; they discuss driver calls, file locks, and a teammate sending a ping when a fix is ready. They plan to restart and test once the fix is applied.
Core Actor Core bot is stuck; they debate giving it a skill
Maticus describes trouble with Core (the Actor Core Engineer Bot) refusing to apply a bugfix for AfterCore and asks colleagues to inspect Core's thread; they suggest the bot might need a skill added to proceed.
Maticus runs through this week's open-model boom
Maticus summarized a long list of new AI releases and features—Meta's Muse real-time voice and avatars, motion model Mobius, the large open model KimiK3.1 release on Hugging Face, Baidu's ERNIE-like 5.5 Flash, Ant Group's multimodal Ling3, Ripple65B, Mage 1 (Longcat 2.5), Thinking Machines' Inkling2 Small, Vox Solution 3D generator, and a 3D reconstruction model called Solver—calling out sizes, capabilities, licensing and where to try each.
Maticus reads Elon on massive Starlink growth and Ship 42 timing
Maticus reads and reacts to Elon Musk's claims about Starlink, revenue forecasts, Ship 42 catching a power, and a Qualcomm/Starlink Mobile integration that would make a global cellular-satellite brand.
Maticus reads warning from a Kyiv interview about drones and mesh nets
Maticus relays and reacts to a report from Kyiv about Russian use of modified Shahed drones, mesh networks, loss of Starlink-like access, and concerns about large-scale drone attacks and satellite competition.
Could Starlink be used to hunt across borders?
Maticus reads and reacts to a Kyiv interview questioning whether Starlink-style satellite meshes could let one country ‘hunt’ targets inside another, and whether persuading Elon Musk to enable Starlink in Russia would be wise.
Launch schedules already put some projects on a fast track
Maticus reads numbers about recent satellite launches and notes that some groups already have launches scheduled that put their service timelines within the next few years.
AI is moving into battlefield planning and final targeting
Maticus walks through excerpts about how current machine vision and AI tools are being used for final targeting and decision support in war, including human-in-the-loop examples and the prospect of AI recommending strategic moves.
Maticus walks through memory tricks in long-context models
Maticus reads and explains a technical write-up about model attention designs—sliding windows, long layers, cross-layer attention (CLA) and Yoko (You Only Cash Once)—and how those tricks cut memory and speed up long-context inference. They follow the paper's reasoning about caches, which layers write or read them, and why some designs reduce test performance on long-context tasks despite similar perplexity.
Maticus walks through how long-context training affects agents
Maticus explained how long attention and training phases change a model's ability to recall far-back tokens and how that matters for agent-style, long sessions. They described average token perplexity, the penalty from losing long-range context, and how longer training and phase choices can reduce that gap.