Daily digest
12 items · ~12 min · Week 2026-W28
Must-read (4)
Meta launches Muse Image and previews Muse Video from Meta Superintelligence Labs
Meta AIMeta Superintelligence Labs released Muse Image on July 7, an AI image generation model that operates as an agent—using search and coding tools to self-refine outputs and compose from multiple references. It handles complex multi-step prompts, photo blending, and in-image text rendering. Muse Image is available free in Meta AI, Instagram Stories (US, with 30+ AI effects), and WhatsApp chats; a Muse Video preview was also announced, sharing the same pretraining foundation with native audio support.
InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization
InternRoboticsInternVLA-A1.5 addresses semantic drift during robot manipulation training by preserving VQA and subtask-prediction objectives alongside action learning. Future prediction is reformulated as latent-querying via learnable foresight tokens supervised by a frozen video generation model, giving the policy world-model dynamics priors without pixel-level generation. Pretrained on 1.2M robot episodes and 3M multimodal samples, it achieves state-of-the-art results on six simulation benchmarks and strong compositional generalization on real robots at real-time speeds.
LLM-as-a-Verifier: verification as an independent scaling axis for LLMs
Stanford University / UC Berkeley / NVIDIAThe paper proposes verification as a new scaling axis for LLMs, complementing pre-training, post-training, and test-time compute. Instead of discrete scores, the framework computes the expectation over scoring-token logit distributions to produce continuous verification signals, scaling along three dimensions: score granularity, repeated evaluation, and criteria decomposition. Evaluated across coding (SWE-Bench Verified 78.2%), robotics (RoboRewardBench 87.4%), and medical domains (MedAgentBench 73.3%), setting new SoTA results.
Multiplayer Interactive World Models with Representation Autoencoders
This paper introduces the first world model that conditions on simultaneous action streams from multiple agents, learning to correctly attribute scene changes to individual players rather than treating co-players as background. Trained on 10,000 hours of Rocket League gameplay, the 5B-parameter latent diffusion model generates four-player matches in real time at 20 fps on a single Nvidia B200 GPU. Despite training on short clips, rollouts remain stable for over five minutes in testing.
Worth knowing (4)
GitHub Copilot desktop app goes GA on all plans including Free, with BYOK support
GitHubGitHub released the Copilot desktop app as generally available on July 7, available on macOS, Windows, and Linux for every Copilot plan—including Copilot Free and GitHub Education. The app enables agent-driven development sessions from the desktop; a bring-your-own-key (BYOK) mode lets users connect to their own model provider without a Copilot subscription. Business and Enterprise plans require admin policy enablement.
Kimi K2.7 Code (1T params, 32B active) expands to GitHub Copilot Business and Enterprise
Moonshot AIGitHub expanded Kimi K2.7 Code to Copilot Business and Enterprise on July 7, a week after its July 1 rollout to consumer plans. The model is a 1T-parameter, 32B-active-per-token sparse MoE coding model from Moonshot AI, hosted by GitHub on Azure and billed at provider list pricing under usage-based billing. The model is off by default for organizations and requires admin opt-in.
AutomationBench-AA: 657-task independent benchmark for AI agent SaaS automation
Artificial AnalysisArtificial Analysis launched AutomationBench-AA on July 6, an independent evaluation of AI agents across 657 tasks on 40 simulated SaaS applications (Gmail, Slack, Salesforce, Jira, HubSpot, Zendesk, and others) in six business domains. The benchmark separately tracks task completion rate and guardrail violations. Claude Fable 5 (48.6%) and Opus 4.8 (48.5%) lead on task success; Gemini 3.5 Flash leads on efficiency ratio at 15.0 objectives per guardrail violation. Every model tested violated at least one business rule.
OpenAI Codex v0.143.0: first stable release with remote plugins on by default and Amazon Bedrock support
OpenAICodex v0.143.0 (July 8) is the first stable release in the 0.143 series after 39+ alpha builds. Remote plugins are now enabled by default; system proxy is supported on macOS and Windows; Amazon Bedrock models with configurable reasoning effort are added; MCP tool handling is improved. Bug fixes cover Windows input handling, stale TUI prompts, and WebSocket request failures.
For reference (4)
Anthropic extends Claude Fable 5 included access through July 12 after user backlash
AnthropicAnthropic announced on July 7 that included (no-extra-charge) access to Claude Fable 5 for Pro, Max, Team, and Enterprise subscribers is extended five days to July 12, up from the original July 7 cutoff. Subscribers may use up to 50% of weekly limits on Fable 5 at no extra charge during this window; after July 12, usage shifts to prepaid credits at $10/M input and $50/M output tokens.
Claude Code v2.1.203–v2.1.204: background session fixes and hook streaming repair
AnthropicTwo Claude Code releases shipped July 7–8: v2.1.203 fixed macOS background session stalls (15–20 s false low-memory trigger), Windows PATH inheritance in background agents, and added login-expiry warnings with a manual-permission pause badge. v2.1.204 fixed hook events not streaming during headless SessionStart hooks, which caused remote workers to be idle-reaped mid-hook.
Ollama v0.31.2: MLX small-batch matmul kernel, llama.cpp build 9840, CUDA updates
Ollama v0.31.2 shipped July 7 after two release candidates on July 6. Key changes: MLX updated with a new small-batch matmul kernel contributing to ~90% Gemma 4 throughput gains on Apple Silicon, llama.cpp updated to build 9840, CUDA and GPU compatibility improvements, UTF-8 file handling fixes, and unsupported ROCm device entries removed.
OpenCode v1.17.15: Z.ai overflow fix and model tooltip restoration
SSTSST released OpenCode v1.17.15 on July 7, fixing a Z.ai context-window overflow error that was misclassified and causing silent failures, restoring missing model-details tooltips in the model picker, and addressing a macOS Sequoia titlebar appearance issue.