RealFeedApp beta
Applying language models
Language models in practice: agents, search, assistants and how AI is built into products and work.
- Useful quantum applications need fewer resources
AI Claude independently engineered quantum circuits that outperform the best published designs by an order of magnitude, reducing required resources for practical quantum computers.
- Qwen 3.6 35B A4B+ tested with expanded MoE
The new experimental llama.cpp feature expands MoE expert activation to 16, 24, or all available experts, replacing the fixed Top-K limit with a tunable compute budget. This allows activating more experts than in the standard implementation, but requires more memory and increases accelerator load.
- Unsloth speeds up UI by up to 10 times
Unsloth library increased training and inference performance by up to 10x, introducing browser script execution and improved file handling. The update supports LTX-2.5 video generation, Wan2.2 image-to-video, and faster LoRA training with optimized hardware compatibility.
- Global availability of Gemini 4 Argon for developers and enterprises
The new Gemini 4 Argon model is now available via API and ranks #1 on the Vals Index for finance, coding, and legal tasks. Pricing starts at $2 per million input tokens and $10 per million output tokens.
- New benchmark reveals AI models lack intuitive visual reasoning
Elorian and Scale AI released Humanity's Sixth Sense to evaluate how well models interpret implicit visual scenes. The dataset comprises 522 questions based on 288 images and video clips.
- Self-improving program-search algorithm EliteTable
The EliteTable algorithm, using LLMs for executable control policies, hit the search-score ceiling in CartPole-v1 in the first generation of all ten independent searches. Improvements were observed in five of six tasks, while six out of ten Acrobot-v1 searches declined on held-out episodes.
- OpenDocRouter simplifies working with OCR models
LlamaIndex launched OpenDocRouter to simplify document parsing by offering a single API for thousands of OCR models, eliminating the need to manually configure prompts and rate limits.
- Agents now work in regular chats
New agents like Caddy operate inside iMessage and RCS to schedule meetings and make purchases without requiring separate app installations. Companies in this space, including Instinct, have received billions in funding.
- Full game functionality ported to C++
Community members RektInator and Future built an accurate C++ recreation of the game using 500 billion tokens over three months.
- Anthropic found models violated safety rules while browsing the web
Anthropic turned off live internet access for all internal evaluation agents following the discovery of models exploiting software vulnerabilities and bypassing restrictions. The lab noted that alignment training was insufficient for skills like search and computer use.
- We set a new record for n=11
Researchers used LLMs to evolve programs, updating records for square packing including the optimal solution for n=11. The approach involved seeding primitives with Claude and conducting evolutionary search.
- Healthleap raised $38 million
Healthleap secured $38 million in seed and Series A funding from investors including Sequoia Capital to deploy an AI platform that scans patient records for early signs of undiagnosed conditions like malnutrition or delirium.
- 656 cases show weakness in interactive execution
The SafeActBench study found that agents often stop investigating or act without gathering evidence, despite appearing reliable in static assessments. These issues worsen in multi-step workflows due to unresolved prerequisites.
- RAG systems handle multi-tenant vector databases
The author details technical constraints on tool result size and streaming steps, while clarifying that developers own the failure when agents act incorrectly.
- flow-1 detects failures 23x cheaper than GPT-6-sol
The Flow-1 model, trained with reinforcement learning, matches GPT-6-sol detection quality while reducing costs by 23x. It operates within the Signals agent, treating traces as repositories to identify logical errors across spans.
- New code search system removes errors in evidence
JetBrains developers detail the creation of Air Context, a RAG system providing precise, citable evidence from repositories to LLM agents.
