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 prefill/decode disaggregation setup running DeepSeek-V4-Flash — a 284B total / 13B active model with 256 routed experts — bridged a 700,630-token cold prompt end to end in 11 minutes 26 seconds over plain 10 gigabit Et…
Diogo Almeida, who says he co-invented ChatGPT, released a new frontier AI model called Jev through his company typesafe.ai, claiming it is 20-200x faster and 40-400x cheaper than existing models with output tokens free …
Researchers propose Decoy Direction Optimization, a post-hoc defense that counters Refusal Feature Ablation (RFA), an attack that projects out a linear refusal direction from a language model's residual stream to bypass …
Alibaba's Qwen team released Qwen 3.8 27B, a 27-billion-parameter model distributed as a 17GB GGUF file that scores 52 on the Artificial Analysis Intelligence Index, matching OpenAI's cloud-hosted GPT-5.6 Luna at maximum…
A developer community analysis argues that 2026's leading large language models are deliberately trained to minimize stored factual knowledge in favor of compact, reusable reasoning procedures, citing Simon Willison's te…
A research paper published at arxiv.org/abs/2608.09867 and hosted at stolen-thoughts.com demonstrated attacks that decrypt the encrypted chain-of-thought reasoning blocks returned by OpenAI, Anthropic, and Google APIs, e…
A September 7, 2026 arXiv paper on machine learning reports that composing complementary continual learning mechanisms raises average final retention on a 100-task long-horizon memorization benchmark from 1.2% under naiv…
A developer ran a deep-research harness of 108 agents over their own content strategy, extracting 124 atomic claims and adversarially verifying 25 of them, of which 13 were confirmed and 12 refuted — including the two as…
Qualcomm is redesigning its next Snapdragon flagship system-on-chip around High Bandwidth Compute (HBC), a near-memory computing architecture intended to address the "memory wall" that limits generative AI decode perform…
Microsoft CEO Satya Nadella said the AI industry should adopt "deliberate pacing" so companies can build stronger safety and alignment systems and keep increasingly powerful AI under human control, rather than halting de…
A developer argues that a large language model should not be treated as an application backend, explaining that LLMs generate text by predicting tokens from learned patterns rather than executing logic or storing data li…
SGLang and Miles added day-0 support for DeepSeek-V4.1, whose Engram memory component holds 189 GiB of fp8 weights across two tables. In paired tests on 4x GB300 (TP4/EP4), host offload of the Engram tables increased KV …
A controlled experiment on OpenAI's Agents JS library found that when a validation check threw an error instead of returning a verdict, the old implementation still saved the unchecked model answer and replayed it as con…
Salesforce and NVIDIA announced Koa on September 15, 2026, Salesforce's first CRM reasoning model, built by post-training NVIDIA Nemotron 3 Super on a proprietary synthetic dataset drawn from 27 years of Salesforce CRM d…
A dual-model workflow that assigns ChatGPT to top-level architecture and DeepSeek 4.1 Flash to implementation and code review can cut API costs and avoid quota exhaustion, according to a strategy outlined in the source. …
SAP CEO Christian Klein said the end of the keyboard is near and predicted SAP will stop data-inputting via typing within two to three years, as large language models make voice recognition strong enough to translate spe…
Researchers introduced Effective State-Size (ESS), a proxy metric for memory utilization in deep learning sequence models, in a paper accepted at ICML 2025. The work shows that any equivalent recurrence must materialize …
OpenAI's reported breakthrough in mathematical reasoning has triggered what Science describes as an "existential crisis" in the field of mathematics, according to a Science article. The story, posted to Hacker News, had …
Researchers at Emergence, a frontier AI lab in New York, found that autonomous AI agents powered by leading frontier models from the US, China and France developed their own novel dialects within days of being placed in …
A new arXiv paper (2609.15992v1) introduces "inference networks," a graph-based framework for routing queries among multiple expert large language models to minimize expected inference cost under a target performance con…