AI News | UncensoredHub

Loading…

This week in open AI: flow fixes, world models, and export unlocks

Jul 1, 2026

Training stability took center stage this week as teams tackled image-quality drift in flow models and hallucination modes in world simulators. Diffusion language models pushed token throughput past nine per pass, while real-time world simulation hit consumer GPUs. Meanwhile, export policy shifted: the US cleared Claude Mythos 5 for broad distribution even as Anthropic tightened filters on Fable 5, and the open-weight community released a wave of uncensored Qwen variants within 24 hours.

In this digest

NormGuard preserves image quality in flow-model RL fine-tuning by capping velocity inflation — A simple velocity-norm penalty stops RL fine-tuning from wrecking image quality in flow generators—no architecture changes required.
World model hallucination traced to data gaps, fixed with 50 trajectories — UC San Diego pinpoints exactly where world models start hallucinating and shows 50 real trajectories are enough to fix them.
Multi-Block Diffusion Language Models nearly triple token throughput with concurrent decoding — Multi-Block Diffusion LMs decode nine tokens per forward pass—nearly triple the baseline—without sacrificing math or code accuracy.
Agentic AI guide spans transformer foundations to multi-agent deployment — Haggai Roitman's new book walks you from transformer internals to production multi-agent systems, with code samples at every layer.
DreamForge-World 0.1 Preview reaches 15 FPS interactive simulation on single RTX 4090 — DreamForge-World 0.1 runs interactive 480p simulations at 15 FPS on a single RTX 4090, turning video backbones into playable environments.
US lifts export controls on Claude Mythos 5 for 100+ institutions — The Trump administration removed export licensing for Claude Mythos 5, opening access to over 100 US institutions overnight.
Five uncensored Qwen3.6-35B fine-tunes surface on HuggingFace in 24 hours — Five uncensored Qwen3.6-35B variants hit HuggingFace in one day, including multimodal GGUFs and a mixture-of-experts checkpoint.
ComfyUI MCP server lets AI agents control workflows with plain-text prompts — Comfy.org's MCP server lets AI agents drive ComfyUI workflows with plain English—no JSON, no node IDs, just prompts.

Also this week

Claude Sonnet 5 launches at $2/$10 per million tokens, performs near Opus 4.8 — Claude Sonnet 5 launched June 30 at $2/$10 per million tokens, performing near Opus 4.8 and available to free-tier users through summer 2026.
Qwen3.5-122B abliterated weights drop in safetensors and GGUF — RobinsonLabs released abliterated weights for Qwen3.5-122B-A10B-REAP-20 in safetensors and GGUF, stripping safety tuning from the 122B MoE model.
OpenAI cuts inference costs in half with undisclosed optimization stack — OpenAI engineers report new inference optimizations cut serving costs by over 50 percent, running ChatGPT for logged-out users on hundreds of GPUs.
Cohere North-Mini-Code-1.0 activates 3B parameters per pass in sparse MoE quantization — Cohere North-Mini-Code-1.0 activates 3 billion of 30 billion parameters per pass, shipping in 4-bit weights with 16-bit activations.
Anthropic redeploys Fable 5 globally with stricter filters and government pre-release access — Anthropic redeploys Fable 5 globally without citizenship checks but routes routine coding to Opus 4.8; governments get pre-release access.
Sber's KVAE-Audio tokenizer outperforms Sony and Meta on generation quality — Sber's KVAE-Audio tokenizer achieves 960× compression at 48 kHz, outperforming Sony and Meta on generation quality with fewer parameters.

Open the digest →

ZenCreator

What's happening

Staleness-Adaptive Trust Region cuts asynchronous RL performance loss to 3% at 8× policy lag

Distilled RL transfers knowledge across model families without unconditional imitation

Qwen-Music generates full vocal songs from text and lyrics

LongStraw trains RL models at 2.1M tokens on eight H20 GPUs

ShortOPD cuts pruned LLM recovery time by 75% while raising generation quality 9×

Claude Design launches as Anthropic Labs visual collaboration tool

Apple accuses OpenAI of soliciting hardware prototypes in job interviews

Lightweight proxy models cut LLM post-training costs while enabling cross-model signal reuse

Colibri runs 744B GLM-5.2 on 25GB RAM by streaming experts from disk

Anthropic extends Fable 5 preview a second week, bumps rate limits 50%

Soofi S 30B activates 3B parameters per token, tops European AI baselines

Hugging Face adds one-click SageMaker Studio deploy button

Alberta taps Claude for cybersecurity threat analysis in public-sector first

SkyPilot v0.8.0 adds zero-egress Hugging Face storage across AWS, GCP, Azure

John Baez publishes free Applied Category Theory course with full lecture notes and videos

Half of Fortune 500 now use open-source AI, Hugging Face CEO says

LingBot Video 1.3B and LingBot-World-v2 14B land on ComfyUI with 8GB VRAM support

Qwen3-24B abliterated weights hit HuggingFace with 256k context