Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
A developer argues that AI memory should be structured as handoffs rather than summaries, identifying four key classes of information to carry between sessions: framing shifts, boundaries, unresolved states, and correcti…
Apple hardcoded a restriction into Siri AI's system prompt in iOS 27 beta 2, causing it to refuse to summarize, read, or extract information from URLs. The move aims to prevent Siri AI from repackaging web content withou…
Loop engineering is a new paradigm for building self-prompting autonomous AI systems that iterate through cycles of task generation, execution, observation, and refinement until goals are achieved. The approach replaces …
Chinese AI startup DeepSeek plans to at least double its staff of roughly 170 after raising $7.4 billion in its first external funding round, one of the largest ever for a Chinese startup. The round, which values the com…
LightSeek Org open-sourced TokenSpeed-Kernel, a portable API and high-performance kernel subsystem for multi-silicon LLM inference, decoupling runtime from hardware-specific code to simplify backend complexity. The syste…
Chinese AI lab DeepSeek is doubling its headcount across departments, including a new Harness team for agent products, as it shifts from a research-focused operation to a product company. The expansion aims to build a na…
A developer explains that AI agent memory is not a single system but several distinct stores—working memory, vector retrieval, episodic traces, and persistent facts—each solving different failure modes. The goal is to us…
Anthropic CEO Dario Amodei was replaced by co-founder Tom Brown in White House meetings after officials found Amodei difficult to engage with, according to a report. The change comes as the Trump administration imposed e…
OpenAI released gpt-realtime-2 on May 7, offering speech-to-speech reasoning at $0.25-$0.35 per minute, which collapsed the AI voice wrapper layer by eliminating the need for stitching multiple vendors. Vapi raised $50M …
Subwave launches a personal internet radio station powered by an agentic AI DJ that selects tracks, writes and speaks intros, station idents, time checks, and weather updates. The service streams a single shared broadcas…
A technical writer argues that diverse technical systems—from data engineering pipelines to RAG-based AI products—share a common underlying structure, using stages of refinement to transform raw inputs into usable output…
Vicki Boykis, a machine learning engineer who has worked at Tumblr, Automattic, Duo, and Mozilla.ai, delivered a talk at the Applied Machine Learning Conference in Charlottesville, VA in April 2026, reflecting on the sta…
DeepReinforce open-sourced Ornith-1.0, a family of self-improving coding models ranging from 9B to 397B parameters, which learn to generate their own task-specific scaffolds during reinforcement learning. The models achi…
A developer investigating low prompt processing throughput with Multi-Token Prediction (MTP) in llama.cpp created a proof-of-concept that recovers the overhead by processing only the output row of the last layer MoE FFN …
The European Union held talks with the White House to regain access to Anthropic's advanced AI models after a US export-control order cut off European access in mid-June. The order, citing a jailbreak concern, forced Ant…
KV Cache reduces duplicated computation in autoregressive LLM inference by storing previously computed Key and Value tensors, but creates a memory bottleneck as context length grows. To address this, Multi-Query Attentio…
A developer building a CLI AI agent found that planning mode was harder than direct execution. The agent required a separate planner with read-only tools to avoid making plans based on assumptions rather than real data. …
Investor Chamath Palihapitiya said Meta “completely fumbled” its lead in artificial intelligence, arguing the company failed to capitalize on its data and distribution advantages to become the dominant open-weight AI lab…
Baidu released Unlimited OCR, an open-source 3-billion-parameter model that uses Reference Sliding Window Attention to parse multi-page documents in a single pass, eliminating the need for page-by-page OCR pipelines. The…
Multi-agent systems that perform well in demonstrations often fail in production due to memory inconsistencies, according to a new analysis. The article highlights the challenges of maintaining coherent memory across mul…