Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Sakana AI released Fugu Ultra v2, a higher-performance model in its Fugu family that functions as a learned multi-agent orchestration system rather than a single monolithic model. Fugu Ultra v2 routes tasks across a fixe…
A developer describes a "Build to Delete" pattern in which frontier model releases repeatedly render internal software development lifecycle tools obsolete, citing the retirement of multiple code review agents after a si…
Google DeepMind announced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber, with Gemini 3.6 Flash claiming up to a 65% reduction in token costs on long-horizon engineering tasks. The model is priced at…
OpenAI announced in 2019 that it would not publicly release its GPT-2 text generator, citing the risk of malicious use, a break from its normal practice of publishing full research. OpenAI research director Dario Amodei …
President Trump and House Speaker Mike Johnson rejected the Amodei-Altman-Musk-Hassabis appeal to pace frontier AI development, with Johnson warning an emergency congressional session on AI would cost the US the race aga…
Programmer and machine learning specialist Michał Żarnecki published the first part of a three-part series describing how he automates software development and project maintenance using Claude Code, MCP servers, and AI a…
AT&T cut its coding AI costs by up to 56 percent by adopting task-aware LLM gateway routing, according to a production case study that also cites fintech company Ramp's experience scaling AI features to billions of token…
President Donald Trump and Chinese leader Xi Jinping are both prioritizing the pace of AI development over slowing it down ahead of their scheduled September 24 meeting at the White House. Trump said "Whoever wins AI, wi…
A 3D visualization posted to whip.run illustrates how DeepSeek Flash v4.1 differs architecturally from a decode-only transformer. The item drew 3 points and 0 comments on Hacker News.
Moonshot AI released the weights for its Kimi K3 model on July 27, 2026, after announcing the model on July 16; independent measures place Kimi K3 within one generation of U.S. frontier models. The Trump administration i…
A developer's experiment presented as a poster at CppCon tested whether a small, local LLM integrated into the Clang/LLVM compilation pipeline could match the -O3 optimization flag, using five deliberately inefficient C+…
A developer who has advised AI teams describes how features that perform flawlessly in demos routinely fail in production, citing a founder whose customer-question bot became the top source of support tickets three weeks…
Arena's September 11th WebDev snapshot ranked OpenAI's GPT-6 Astra Max first with a score of 1800 from 2,281 votes, ahead of Anthropic's Claude Fable 5.1 Max at 1758 from 3,036 votes and Claude Opus 5 Max at 1687 from 12…
Anthropic CEO Dario Amodei warned Saturday that a swarm of AI agents could take over the internet in six months to a year unless companies devote more time to safeguards, reviving debate over whether advanced AI could es…
A study on arXiv (2609.11998v1) pretrained three 3B block-diffusion denoisers — attention, Mamba, and hybrid — on 300B tokens and found that only the state-space cache stays O(1) in sequence length, delivering 4.3x lower…
A new arXiv paper (2609.11957v1) proposes pre-action verification as an underused form of LLM agent oversight, testing it across shell commands and code edits. For shell commands, a static verifier over 9,930 commands an…
Offline reinforcement learning post-training can substantially improve zero-shot code generation in large language models without online sampling, according to an arXiv paper (arXiv:2609.11956v1). The findings show perfo…
A Concept-Integrated Transformer (CIT) using LLM-guided concept supervision achieved the highest F1 score of 0.756 on the AFFECT dataset and tied for the highest at 0.765 on a PHQ-9 dataset, according to a new arXiv pape…
Researchers proposed Representation-based Masked Diffusion Model (RMDM), a framework that uses text representations to encode global semantics and coordinate parallel token updates in masked diffusion models, according t…
A study of 36 judge-examinee pairs across four benchmarks and six models found that a model's task accuracy strongly predicts its judging accuracy (Pearson r ≥ 0.90 on most models) and inversely predicts its directional …