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 built a 5-layer, 512-wide TinyStories language model that runs natively on Apple's M4 CPU by treating ARM Neon instructions—including XOR, popcount, and SDOT—as trainable architectural primitives rather than …
A developer running nine parallel Claude Code sessions rebuilt agent memory recall as server-side plumbing rather than a prompt instruction, injecting semantically retrieved project notes into each session's first messag…
A developer reports that more than 50% of traffic to their Mintlify-hosted API documentation now comes from AI agents, prompting a shift in how OpenAPI specifications are written. The engineer argues that endpoint summar…
A developer detailed a production-grade autonomous clinical concierge agent deployed for a Tier-1 enterprise hospital network, replacing naive probabilistic prompt chains with deterministic finite state machines, idempot…
A developer outlined a method for configuring DeepSeek as the default model in OpenAI's Codex coding agent by editing the local config.toml and models.json files and setting a DEEPSEEK_API_KEY environment variable. The s…
DeepSeek released DeepSeek v4.1-Flash, a 763B-parameter model with a novel causal encoder-decoder architecture that splits 8B parameters to prefill and 16B to decode, per the model's tech report on Hugging Face. The rele…
A developer benchmarked classical game AI algorithms against LLM-based agents after asking ChatGPT and Grok to build the same comparison tool, then actually executing the code. ChatGPT produced a runnable browser tool wi…
Edge0-AI has open-sourced Edge0, an Apache 2.0 streaming inference engine that memory-maps sparse mixture-of-experts weights from SSD so a 35B model runs in roughly 2.9 GiB of peak active memory. The project ships two 4-…
Scalattice released Don't Hit Send, an open-source chat interface that fires a streaming chat completion roughly 350ms after the user pauses typing, with no send button. The tool aborts the prior request if the previous …
Codex can run against a local LM Studio server hosting the 27-billion-parameter gemma-3-27b-it model, but a blocking HTTP proxy test found the tool still contacts four external destinations — chatgpt.com, github.com, api…
A developer analysis of Artificial Analysis benchmarking data finds that GLM-5.3-Flash scores 42 at $0.25 per task while Kimi K3 scores 44 at $2.00 per task, an eight-fold cost gap for a two-point capability difference. …
DeepSeek is retiring its flagship deepseek-v4-pro model, routing all API requests to V4.1-Flash as of September 14, 2026, until V4.1-Pro ships. According to the company, testing by multiple parties found V4.1-Flash beats…
A developer spent two days attempting to reorganize an ESP-IDF project into a components folder using several AI coding assistants, including Gemini Flash Extended, Gemini Pro Extended, DeepSeek v4 Flash, DeepSeek v4 Pro…
AI tutoring startup Aristotle argues that most AI tutors are generic chatbots that fail students because they hand over answers rather than guide reasoning, and says its voice-first system is built on the ICAP engagement…
Bridgewater Associates co-CIO Greg Jensen said on Bloomberg's Odd Lots podcast on September 11, 2026 that there is a 30 to 60 percent chance AI causes a major financial or fatal disaster within two years, warning that "u…
A developer has published a set of conversational reasoning patterns designed to make other agentic coding assistants reason in an "Opus style." The approach prescribes specific verbal markers for flagging caveats, ackno…
Researchers introduced the KuaiRP series of role-playing models in a technical report, using a multi-stage training pipeline to inject deep domain knowledge without losing general agent capabilities. The pipeline combine…
A meta-agent variant achieved the highest Avg@3 reward on five of six heterogeneous benchmarks in a study of task-agnostic environment preprocessing, according to an arXiv paper (2609.10824v1). The research formalizes a …
A tail-risk-aware turn release scheduling method for agentic LLM workflows reduces the P95 of workflow flow time under contention by up to 3.50x, according to an arXiv paper (2609.10964v1). The method, which uses a mean-…
A new arXiv paper (2609.10992v1) analyzes the privacy-utility trade-off in Large Language Model interactions and introduces an intent-driven local protection framework built on a distilled lightweight model, Veilmind-4B.…