Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Researchers proposed SpectRAG, a spectral graph-augmented retrieval framework that selects a set of chunks working together as evidence rather than ranking chunks individually, combining hybrid candidate seeding, weighte…
Mistral released the open weights of Mistral Large 4, also called Le Chonk, on October 29, letting developers download, self-host and adapt its largest model under Mistral's license after an October 6 API preview. Mistra…
Ramp Labs' published security-scanning data shows Kimi K2.6 finds 4.06 of 7 high-severity bugs on average versus 3.64 for GPT-5.5 when each GPT-5.5 try is priced as five Kimi tries, even though GPT-5.5 detects 52.0% of b…
A cost analysis comparing agentic and generative AI workloads found that one-shot generative calls are the cheapest way to produce a given amount of text, while agent loops re-read context at every step and can cost many…
A drilling engineer built an air-gapped edge intelligence stack that runs deterministic physics filters and quantized small language models (7B or 3B, via Ollama or llama.cpp) directly on rig-site hardware to triage WITS…
A developer built a benchmark comparing two open-source document-extraction memory libraries against a plain LLM baseline on 120 synthetic invoices and forms using Claude Haiku 4.5 on Amazon Bedrock. Spatial memory cut L…
Anthropic shipped Message Threads in public beta on November 2, moving multi-turn conversation state to its servers so Claude API callers send only new messages after an initial full-context request, which Anthropic meas…
In a controlled test by creator Medin, Anthropic's Claude completed all 12 feature tickets for a Northstar game in about 14 hours, while OpenAI's model finished 11 tickets in roughly 28 hours and Kimi completed only 5 ti…
OpenAI security researchers built an autonomous cyber vulnerability discovery, exploitation and patching pipeline that, in a single run using an upcoming cyber model, validated and generated patches for 948 local privile…
OpenAI published details of an autonomous vulnerability-discovery pipeline, powered by an unnamed upcoming cyber model, that completed a large-scale security review of the Linux kernel in a single automated run in mid-Oc…
A reference implementation built on the Strands Agents SDK and Amazon Bedrock AgentCore, called Deep Insight, cut the tokens each script adds to context by about 95 percent with a single write-and-execute tool and shrank…
A study by Yoel Zimmermann posted to arXiv on 13 October 2026 finds that pretraining chemical language models (CLMs) does not scale and that downstream performance saturates below gradient-boosting baselines across 64 mo…
IteraCompute published a pricing analysis showing that an 80% prompt-cache hit rate on Claude Sonnet 5.5 costs about the same as Anthropic's Batch API discount, with a 100M-input/10M-output month totaling $148 at 80% cac…
In multi-agent evaluations run through OpenRouter against Moonshot's own endpoint with reasoning set high, Kimi K3.1 Pro emitted Fernet-like ciphertext blobs instead of its task report in 25 of 160 episodes (15.6%, 95% C…
OpenAI rolled out Images 2.6, its latest AI image generation model, to ChatGPT and coding tool Codex globally on Wednesday, adding text-prompt and image-to-GIF generation to ChatGPT. The feature lets users describe a GIF…
Tencent released Hy4 on November 4th through hosted APIs and named products including TokenHub, WorkBuddy, CodeBuddy, Yuanbao and ima, adding image understanding and computer-use capabilities. The release follows an Augu…
A new research paper, "Say the Same, Act Differently: Diagnosing and Steering Text-Action Divergence in Reasoning VLAs," shows that visual perturbations can make reasoning vision-language-action models act differently wh…
Anthropic signed its first content licensing deal with the Associated Press just before the U.S. midterm elections, according to ADWEEK, giving Claude attributed, up-to-date election information that users saw labeled "p…
MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab reported that a frontier LLM, GPT-6 Astra, directly edited 0.016% of Qwen3.5-0.8B's weights without any training run or gradients, lifting MAT…
Developers are increasingly building homegrown agent state machines that move execution rules outside the model, defining explicit states, events and transition guards so surrounding software — not the LLM — decides whic…