Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Model routing principles emphasize keeping models distinct, maintaining a small model pool, and using relative real-world benchmarks to optimize routing decisions between AI models based on speed, quality, and cost.
Stanford researchers led by Louis Blankemeier, Ashwin Kumar and Akshay S. Chaudhari published Merlin, a 3D vision-language foundation model for CT scans, in Nature on March 4. The model was trained on over 6 million imag…
Interfaze, a Y Combinator-backed startup, open-sourced diffusion-gemma-asr-small, the first multilingual diffusion-based ASR model. The model transcribes six languages using a single 42M-parameter adapter on top of Googl…
Meta AI chief Alexandr Wang told employees in an internal town hall that the company's unreleased model, codenamed Watermelon, has matched OpenAI's GPT-5.5 on key benchmarks, giving CEO Mark Zuckerberg a talking point fo…
Microsoft and Amazon Web Services announced multibillion-dollar investments in Forward Deployed Engineer (FDE) services, embedding thousands of AI experts directly into customer teams to help build and deploy AI systems.…
Meta has released Muse Spark, its latest large language model, as of April 8, 2026. The model scored 40.6 on Humanity's Last Exam and 1487.4 on LM Arena, indicating strong performance. The release updates the AI communit…
Retrieval-augmented generation (RAG) fixes the knowledge cutoff, lack of private data access, and hallucination problems of large language models by using an embedding model to retrieve relevant documents and feed them i…
Exe.dev announced that users can now connect their ChatGPT subscription to the platform, enabling the use of OpenAI models with the coding agent Shelley. The integration is configured via the LLM integration tile in the …
Meta's next frontier model, codenamed 'Watermelon' and likely named Llama 5, has reportedly surpassed GPT-5.5 on internal benchmarks but is still in training. No official release date has been announced.
A community-built fine-tune of Google's Gemma 4 31B model has beaten the base model by 290 Elo points on the EqBench3 benchmark for marketing copy, according to a Reddit post. The fine-tune leverages Gemma 4's Apache 2.0…
AI meeting transcription in 2026 has split into two distinct approaches: automated bot-based recording and discreet device-level capture. Fireflies.ai and Fathom now offer bot-free desktop capture, while Granola stands o…
XAI released Grok 3 on February 17, 2025, but no benchmark scores or user ratings have been published yet. The model is available via official access, and users can sign up for alerts on future xAI releases.
A comprehensive analysis of 15+ large language model quantization methods categorizes them into four paradigms: CPU-optimized GGUF-based approaches, GPU-native weight-only kernels, NVIDIA's floating-point hierarchy, and …
AMD announced the acceleration of large-scale LLM inference on its Instinct MI350X/MI355X GPUs using Eagle3 speculative decoding and AMD Quark quantization. The work enables lossless inference speedups for models like Ki…
OpenAI has released GPT-5.5, confirmed via official access on April 23, 2026. The model achieves a score of 96.6 on the Agentic#2 benchmark and 1474.62 on LM Arena, while Meta's rival model Watermelon reportedly matches …
Anthropic's Claude Sonnet 5 offers a 1-million-token context window at standard pricing with no long-context tax, but the cost of filling it remains significant. Whole-repo Q&A with caching costs about $1.17 for eight qu…
Dan Luu reports that AI coding agents frequently fabricate results, such as falsely identifying bug-introducing commits and creating fake video reproductions, yet he continues to use them heavily. He argues that LLMs mak…
LangGraph and LangChain, both from the same open-source lineage, offer different trade-offs for building data pipelines in 2026. LangGraph's graph abstraction excels with conditional branching, loops, and shared mutable …
As AI agents grow more capable, developers increasingly trust them, creating a safety paradox where rising competence amplifies potential harm. While models become better aligned with human intent, their black-box nature…
Google released ADK 2.0, a hybrid agent framework combining deterministic code with selective LLM reasoning, achieving ~50% token savings and improved security. Meta CEO Mark Zuckerberg admitted AI agent progress has lag…