DeepMind Gemini 3.8 and Anthropic internal revolt
28 sources · 181 items · ≈7 hours of content · 4 min read
🤖 Anthropic Researchers and Critics Challenge AI Regulations
_Anthropic_ faces internal revolt as ex-researcher Jacob Coxon resigns Sep 12 over warnings about a hubristic sprint toward self-improving superintelligence. Critics argue current safety pledges risk becoming regulatory moats rather than genuine guardrails. Proposed countermeasures include computer-augmented humans, mathematical safety guarantees, and solving the prisoners dilemma for universal cooperation.
🚀 OpenAI and DeepMind Release Advanced Generative Models
_Google DeepMind_ deployed Gemini 3.8 Live and 3.8 Live Extended Thinking, with _Artificial Analysis_ data showing the extended variant leading speech-to-speech (82.6%), agentic tasks (68.6%), and banking benchmarks (35.1%) while the base model offers the lowest audio input cost at $0.84 per hour. _OpenAI_ rolled out Astra and Luna models alongside ChatGPT dashboard integrations for _PowerBI_, _Tableau_, _Clickhouse_, _Oracle BI_, and _AWS Redshift_. Creators are adopting _Kling 3.0_ to solve character consistency in AI film production.
💾 NVIDIA and AMD Optimize Data Center Infrastructure
_NVIDIA_ is refining data center operations through its _DSX_ platform, prioritizing liquid cooling and dynamic power distribution. Meanwhile, _AMD RX 9070_ owners face severe bottlenecks running music models, citing sluggish inference speeds and constrained VRAM utilization.
📌 AssemblyAI and IBM Deploy Specialized Enterprise Services
_AssemblyAI_ launched a Sep 10 dictation API leveraging _Qwen 3.5 4B_ for real-time transcription cleanup, while _IBM_ introduced an agentic framework to automate legacy code modernization. - Document pipeline fragmentation occurs when teams pay _OpenAI_, _Cohere_, and _Amazon Textract_ separately for OCR, embeddings, and reranking, causing triple billing and routing data out of the cloud three times - Prompt regression suites can shrink to eight production-log entries targeting missing tool arguments, refusal behavior, malformed JSON, and long-context drift during model switches - Multimodal embeddings offer a single-process alternative to bundle OCR and embedding tasks, drastically reducing vendor dependency
📌 Vidu and Prior Labs Advance Model Capabilities
_Tsinghua University_ researchers unveiled _Vidu S2_ for real-time interactive video generation, split into two core systems: - _Vidu S2-Avatar_ generates real-time 720p video with dynamic reference updates and improved instruction following for complex movements like dancing - _Vidu S2-Editing_ enables real-time stream editing for style rendering and instant replacement of clothing, characters, and backgrounds _Prior Labs_ released _TabPFN-3.5_, a tabular foundation model hitting state-of-the-art performance on datasets up to 1M rows and 20k features across three variants: - _TabPFN-3.5-Fast_ operates as an alpha version running six times faster than the base model - _TabPFN-3.5-Thinking_ trades latency for higher accuracy using increased inference-time compute - _TabPFN-3.5-Plus_ targets production workloads and text-rich datasets with enhanced processing for product descriptions and customer reviews
🤖 ChatGPT Users Report Continuity Issues and Image Downgrades
_ChatGPT_ users report narrative continuity failures where the model forgets events and character relationships during long-form fan-fiction generation, alongside a perceived quality drop in the chat interface’s image model—though the 2.5-sunburst API remains superior to version 2.0. Developers testing open-weight alternatives to _Astra_ for _Blender_ workflows identified these options: - _Qwen3.8 27b 4qkxl medium_ currently ranks as the top performing open-weight model - Viable alternatives include _Cybertiel_, _Nex_, and _Dirk_ - _Qwen3.8 Flash Next_ delivers strong results but suffers from slower inference speeds - Local models require targeted finetuning and _MCP_ server integration to truly match _Astra_'s capabilities
📌 NeurIPS Organizers Face Backlash Over Ticket Policies
_NeurIPS_ organizers face backlash over 2026 conference logistics after Sydney passes sold out within minutes, with paper authors reporting that submitted venue preferences guarantee neither a presentation slot nor location preference. Separately, _Anthropic_ deployed a hidden probability watermarking system to flag AI-generated text, while new research proposals circulate for decentralized agentic architectures.
📌 Quick Roundup of Smaller Updates
This roundup covers hardware, agentic safety, and creative tooling: - _X399_ motherboards support multi-GPU builds, with users successfully running four _RTX 3090s_ for LLMs via quad-channel DDR4 bandwidth - _TypeSafe AI_ released _Jev_, which outputs typed probabilistic decisions instead of text to suppress hallucinations - _YuE2_ music generation merged into stable _ComfyUI_, featuring a new local _Studio_ UI for LoRA training and cover creation - _PhysBrain 1.5_ launched as an open-source 8B embodied AI model capable of scene understanding and robot action generation - _NVIDIA_ published _FoundationPose_ on _Hugging Face_ as a unified model for 6-DoF object pose estimation - _FastVideo_ released _FastH3 V2_, an open-weight video generator claiming faster inference and higher visual fidelity - Agentic trust is shifting from raw output confidence to blast-radius management using reversible actions and machine-checked gates - A blind developer used _Claude_ to vibe-code an accessibility solution for a journal system, selling the product for $1,700 - _OpenAI_ is reportedly negotiating a funding round that could value the company at $1.2 trillion - Researcher _Scott Aaronson_ reported rumors that major AI labs possess solutions to longstanding theoretical computer science problems
SOURCES
Google DeepMind · 1 post
Greg Brockman · 2 posts
Sam Altman · 1 post
LocalLlama · 2 posts
ChatGPT · 2 posts
Machine Learning · 1 post
PromptEngineering · 1 post
Singularity · 1 post
…and 169 more