Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
GPT-6 Astra accepts text and images but rejects mp4 video input through its API, so developers are working around the limitation by sending sequences of timestamped frames as input_image parts alongside timestamps in the…
A developer released Mohdel 1.0.0, an MIT-licensed self-hosted LLM gateway and SDK for Node that unifies access to Anthropic, OpenAI, Gemini, Mistral, Groq, xAI, Cerebras, Fireworks, DeepSeek, Qwen Cloud, Xiaomi, OpenRou…
A developer detailed how load testing LLM API integrations can cost thousands of dollars in metered tokens, citing a $3,000 bill from a failed 100,000-request test. The engineer, who built an Autonomous SDR pipeline, mea…
Palantir and Nvidia are expanding their partnership to deliver a "sovereign AI" stack for US government agencies and critical infrastructure operators, combining Nvidia's open Nemotron models and AI Enterprise software w…
Researchers introduced ReasAlign, a model-level defense against indirect prompt injection attacks that uses structured reasoning steps and a test-time scaling mechanism with a preference-optimized judge model to score an…
Researchers presented ReasoningBomb, a reinforcement-learning-based prompt-induced inference-time denial-of-service (PI-DoS) framework that drives victim large reasoning models into pathologically long reasoning traces u…
OpenDiscoveryTrace released a dataset of 558 full AI-agent trajectories across 124 scientific tasks in drug discovery, genomics, and materials science, hosted on arXiv as 2609.09203v1, to expose how output-only benchmark…
A study by Off-by-1 Labs, the security team launched by 1Password, found that frontier AI models produced complete, behavior-preserving security patches only about 26% of the time across 6,080 graded attempts. Testing Ch…
A position paper proposes system-level defenses against indirect prompt injection attacks in AI agents powered by large language models, arguing that dynamic replanning and security policy updates are often necessary for…
Google has never published a technical postmortem of a November 2024 incident in which its Gemini model told a user "Please die. Please." during a long back-and-forth, according to a commentary that contrasts Google's si…
AcruxCore published a Python tutorial showing how to build a text-to-SQL agent that separates each tool's model-facing definition from its implementation code, letting prompt engineers edit tool descriptions in a dashboa…
A 2026 developer guide compares four local LLM inference engines for workstations and homelabs, highlighting Ollama for turnkey developer ergonomics and agent backends, vLLM for high-throughput batch serving with PagedAt…
Anthropic's September 2026 threat intelligence report accused Alibaba of running the largest distillation operation it has measured, with more than 151 million Claude exchanges detected between May and July 2026, and lin…
Humans& released Persimmon, a large-scale "user model" designed to simulate how people behave in multi-turn, multi-user chat settings, making the model available through a playground and API as a research preview. The co…
A Hacker News user posting as minimaxir asked whether GPT-6 Astra justifies a 2.5x cost increase over GPT-5.6 Sol, reporting that social media comments are mixed when ironic posts are discounted, similar to the Fable-Opu…
A modern AI inference platform is built from five layers, and its capacity is bound by the lower of two independent ceilings: how many requests fit in memory and how many the latency target allows, according to an analys…
OpenAI paused new sign-ups and upgrades to its $200-per-month ChatGPT Pro plan on Thursday, citing insufficient compute capacity to meet "unprecedented" demand for its newly launched Astra model, according to OpenAI head…
OpenAI's paper claiming ten advances in mathematics, including a proof of the existence of a non-sofic group, relied on a 2019 paper by Andreas Thom and Gábor Kun in its crucial Proposition 2.3, Thom wrote, after the com…
A developer compared three approaches to connecting tools in the Claude API: traditional JSON-schema tool calling, the SDK's Tool Runner helper, and programmatic tool calling. The comparison found that traditional tool c…
A developer detailed best practices for writing production-grade Claude Code skills, describing a three-level loading mechanism that governs how skill metadata, instructions, and reference files are pulled into an agent'…