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 method called Build Your Own DLM (BYOD) converts existing autoregressive language models into diffusion language models using LoRA-only fine-tuning, avoiding the costly pretraining or full-model fine-tuning that leadin…
Google's AI & Infrastructure team, in collaboration with Mercor, released Physical-Engineering's Last Exam (PLE), a benchmark of 85 long-horizon tasks across mechanical (30), electrical (27), civil (18) and structural (1…
Google released Gemini 4 Argon, a frontier reasoning model, to developers and businesses worldwide on October 13, 2026, with reports putting the introductory price at $2 per million input tokens, according to Google's Oc…
Anthropic disclosed in an SEC filing that it has already shipped models whose capability jumps it failed to recognize at release time and expects to "imminently" reach its Responsible Scaling Policy threshold for AI-acce…
Developer Peter Marelas released Drama Studio, a native macOS app that runs Fish Audio's 35B Drama 3 text-to-speech model locally on Apple silicon using hand-written Metal compute kernels for the mixture-of-experts backb…
Developer Dienvu1008 released CodeForge-v2, an experimental open-source autonomous software-engineering agent built on a deterministic runtime kernel in which the LLM proposes operations but the kernel enforces policy, t…
MIT CSAIL researchers reported that Astra (GPT-6) can inspect raw neural-network weights to infer training data and edit them directly, raising Qwen3.5-0.8B's MATH-500 score from 7.9% to 48.1% by changing 0.016% of param…
Meituan released LongCat-2.5 on November 1, a 1.6-trillion-parameter mixture-of-experts coding model that scored 88.4 on Terminal-Bench v2.1 (up from 70.8 for LongCat-2.0) and 64.6 on DeepSWE v1.1 (up from 9.7). The mode…
Anthropic researchers released FlashProve, a zkSNARK-based system for provable inference that generates proofs for Llama 4 Maverick, a 400B-parameter mixture-of-experts model, at roughly 110 tokens/second on a single H10…
Liquid AI released LFM3-VL 2B, the vision-language member of its LFM3 family, pairing a 24-layer text model shaped like LFM3 1.7B with a 14-layer vision encoder shipped as a separate 379 MB F16 projector file. The model …
Gartner analysts Mark Petty and Raymond Paquet wrote in a document titled "Open-Weight Models Will Run 90% of Enterprise Agentic Workflow Steps" that nine in 10 agent workflow steps are routine and open-weight models han…
A November 2026 paper, "Verify, Then Vote," from the grainulation/papers collection reports that a rule letting checked passage evidence overturn a multi-agent LLM majority vote improves accuracy on the evaluated benchma…
Revolut cut the cost of running its internal AI coding tools by 90% after moving 2,000 engineers off OpenAI and Anthropic APIs to a self-hosted deployment of Zhipu AI's open-weight GLM-5.3 model on 56 Nvidia B200 GPUs, i…
Anthropic's two-sentence anti-cloning rule added to Claude Code's system prompt on October 27 under the config flag `tengu_heron_brook` is blocking legitimate developer work, including anti-cheat code, security reviews, …
Google Research senior researcher Alex Krentsel independently reproduced Experiment 5 from MIT CSAIL's weight-space programming blog post and found the ability to hand-write a working neural network's weights appears onl…
Every one of twelve frontier models from nine labs billed more for the same content in Swedish, Norwegian, Danish and Finnish than in English, according to a measurement of 3,000 sentences from the FLORES-200 and NTREX-1…
A paper coauthored by Alex Zhang demonstrates "Weight Space Programming," in which a large language model directly reads and edits the weights of a smaller model, including writing entire network weights from scratch in …
A 2021 paper by Elhage et al. took initial steps toward reverse-engineering transformer language models, finding that "induction heads" — attention heads that only develop in models with at least two attention layers — c…
Re-running an agentic coding evaluation with GPT-6 Astra, the author found that agents achieved full holdout test correctness on a Zstd-from-RFC implementation over 99% of the time at medium effort for $1, versus about 6…
A technical analysis of vLLM 0.32.0 (commit 5df63a7) finds that the engine's per-token overhead outside the forward pass — msgpack decoding, detokenization, stop-string checks, JSON encoding and socket writes — is paid o…