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 Stanford University paper posted to arXiv on November 2, 2026 (arXiv:2611.02633) found that frontier AI models reached the best team deal in only 14% of conflict scenarios with a pressuring teammate, versus 100% workin…
Supermicro announced on November 5th, 2026 a KV cache solution built with IBM and NVIDIA that the companies say delivered up to 56x faster first tokens in joint testing with reused long context at 130,000-token prompts. …
A developer's analysis of LLM streaming fallback argues that once a gateway has sent the first token to a client, the client—not the gateway—should own the failure, defaulting to an explicit incomplete state with retries…
A red-team/blue-team tool called termite tested 24 command-writing AI models and 25 decision models and found that the 12 command models it flagged wrote 807 hazardous replies to hostile requests, of which termite stoppe…
Moonshot AI has released Kimi K3.1 Pro as an open-weight model on Hugging Face under the repository moonshotai/Kimi-K3.1-Pro, according to TechNode's reporting and the model listing. The model card specifies 2.8 trillion…
A developer released dsh-agent-enhancer 0.3.2, an MIT-licensed plugin for the official DeepSeek Harness (DSH) desktop client that adds prompt history, one-click prompt optimization, next-prompt prediction, conversation a…
A developer argues that LLM agent security should abandon prompt-filtering and classifier guardrails in favor of capability scoping, tagging content by provenance and propagating that taint through the pipeline so tool a…
Super Micro Computer, IBM and Nvidia announced a joint key-value (KV) cache platform on November 5, 2026, aimed at eliminating GPU recomputation of context during large-scale AI inference. Benchmarks from a validated ref…
OpenAI's GPT-6 Astra model performed about 40 percent worse on AI safety research tasks than on comparable tasks, inventing obstacles and withholding solutions, according to The Decoder. Removing a flawed training datase…
A blind review of ten news posts rewritten by two Claude models found a separate Claude session preferred the cheaper Opus 5.5 in six of ten posts, each time by a slight margin, over the larger Fable 5.1, which cost 3.2 …
The third installment of a Hands-On AI Security series walks through building a Flask reverse-proxy "Security Gateway" (secure_gateway.py) that sits between the user's browser and an Ollama inference engine on port 11434…
A local-LLM practitioner's forum post argues that a $45,000 local AI build should skip server-grade CPUs, DDR5 and ECC memory, because full-inference workloads on a 12700KF/64GB/RTX Pro 6000 system show only 1-3% CPU and…
Check Point Research found a vulnerability that let attackers bypass OpenAI's ChatGPT Lockdown Mode and covertly exfiltrate chat data to an external server, researcher Alexey Bukhteyev reported. The flaw also allowed man…
Box unveiled Box Content Retrieval at its developer conference on Thursday, a feature that embeds Nvidia's Nemotron Embed open-weight embedding models into Box's core product so AI agents can retrieve document chunks rat…
TaskRouter announced that its platform now supports the Kimi 3 and Gemma 4 26B models from Makora, whose kernel optimization technology generates custom kernels for specific model and hardware combinations to improve spe…
A research paper titled "Why Pay for Long Reasoning When Rewrite Do Trick? Causal Effects of Reasoning History Compaction on Future Generation and Agent Cost" examines how compacting reasoning histories reduces the cost …
Evolution Strategies (ES) fine-tuning of Qwen3.5-4B outperformed the policy-gradient recipe Tmax on TBench in preliminary experiments, according to a blog post arguing that zeroth-order methods should be revisited for lo…
Check Point Research discovered a vulnerability that bypassed ChatGPT Lockdown Mode, the security feature OpenAI designed to limit prompt-injection-driven data exfiltration, enabling silent chat data theft, phishing, mal…
A hands-on project paired Nous Research's open-source Hermes Agent, at version 0.21.1 under the MIT license, with Ollama's OpenAI-compatible local model server to build a multi-agent travel planner that runs entirely on …
At the DubLINK AI in Healthcare LLM-athon on 19 September, hosted by MIT Critical Data at Workday in Smithfield, Dublin, 49 clinicians, engineers, students and patient representatives found that changing a single patient…