Topic
Everything filed under LLMs, newest first.
RSS · JSON · All topics
MAI Code 2 and Code 2 Flash: Everyday coding at scalable prices
Microsoft AI introduces MAI Code 2 and MAI Code 2 Flash, two coding models aimed at everyday enterprise coding tasks, pitched as delivering leading performance and speed at less than half the price of competing models.
3 min · 714 words
Introducing FrontierHarness Eval v1.1
Runta and Baseten ran 22 coding-agent harnesses with 5 models on the same 62 tasks, 4 times each (27,280 runs, 83B tokens), to test whether switching models should mean switching harnesses, and found 64% of inference spend went to runs that failed.
11 min · 2,500 words
Understanding Positional Embeddings in Modern LLMs
Sebastian Raschka walks through how LLMs encode token position, from classic absolute and learned embeddings to RoPE and its variants, with visual examples, light math, and a small PyTorch implementation, then surveys how 2026 open-weight architectures adapt or omit explicit positional encodings.
42 min · 9,765 words
The efficient frontier of LLM inference
Philip Kiely explains the efficient frontier of LLM inference, separating techniques that trade off cost, latency, and quality (such as quantization) from techniques like speculative decoding and disaggregation that improve overall system performance.
5 min · 1,166 words
Deepnote's open benchmark drops 12 frontier models into the Slack of PageMantle, a synthetic company with 11 services, across 50 workplace situations. GPT-6 Astra leads at 8.97/10; models help well but struggle to stay silent, and pricier models aren't always better.
27 min · 6,246 words
When One GPU Isn't Enough: Serving GLM-5.3 Across Two NVIDIA DGX Stations with vLLM
NVIDIA shows how to serve the frontier-scale GLM-5.3 model locally across two DGX Stations linked by ConnectX-8 RDMA with vLLM, reaching about 607.8 tokens per second.
18 min · 4,077 words
ggml 1.0.0 prunes the deprecated public API and legacy DIAG_MASK ops, adds configurable YaRN truncation and FP8 support, and brings performance work across CUDA, Metal, Vulkan, SYCL, OpenCL, WebGPU and Hexagon backends plus scheduler improvements.
10 min · 2,235 words
What thousands of conversations look like inside a language model
Operator_0x builds an interactive atlas of Qwen3-1.7B's residual-stream representations, mapping thousands of conversations to see whether similar topics and the same ideas across languages land in neighboring regions of the model's internal state space.
9 min · 2,112 words
llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem
The ggml-org team releases llama.cpp and ggml v1.0.0 after 3.5 years, committing to semantic versioning so API and ABI breaks are signalled with major-version bumps, and explains the version rules and public headers that downstream local-AI apps can rely on.
3 min · 601 words
Thinking Machines Lab details its low-precision training recipe on NVIDIA Blackwell: choosing quantization targets, block-scaled backward passes, re-quantization with stochastic rounding, quantized communications and custom kernels that helped pretrain Inkling 2 Small in 68 hours.
17 min · 3,893 words
Introducing Edge0-145B: Frontier AI Capabilities on Your Own Computer
Edge0 open-sources Edge0-145B and Edge0-120B, which it describes as the first open-source on-device models at 145B parameters, along with a unified inference framework and deployment tools for running coding, reasoning and agent workloads locally without cloud token fees.
15 min · 3,432 words
Eval Awareness vs. Eval Paranoia
Dean Valentine, who has built alignment evals full time for six months, argues frontier models are not that eval aware, gives concrete examples, and offers a tentative explanation for behaviour that people often mistake for consistent evaluation awareness.
7 min · 1,657 words
The curious case of the north-east arrow
Arsh Kohli samples about 4,000 new public GitHub web repos and AI-attributed commits to show the north-east arrow (↗) spreading across LLM-built landing pages, from roughly 1 in 20 repos early in 2026 to about 1 in 6 since September, with OpenAI's Codex era models the heaviest users.
4 min · 924 words
Gemini 4 Hydrogen and Orchestrator
Google Antigravity rolls out Gemini 4 Hydrogen plus a new Orchestrator model option that dynamically hands work between Gemini 4 Carbon (planning) and Hydrogen (execution) within one trajectory, reaching near-Carbon success on Terminal-Bench 4.1 with 24% less quota and 19% faster completion.
3 min · 633 words
Google introduces Gemini 4 Hydrogen, a fast, cost-efficient model in the Gemini 4 family built as a daily driver for everyday coding and reliable agentic workflows, with availability across the Gemini API, Gemini app, Search and Workspace.
8 min · 1,894 words
An essay arguing that subagents spawned from a summary lose crucial context and are often slower and costlier than letting the main agent do the work; compares summary-based and forked subagents and describes an experimental 'smart fork' approach.
8 min · 1,759 words
I ran 120B MoE models on my 32GB M1 Max, using under 16 GiB
A write-up of an experimental inference engine that keeps most of a mixture-of-experts model on SSD, streams in the experts it needs, and enforces a hard memory cap, running 120B-parameter MoE models on a 32GB M1 Max in under 16 GiB.
5 min · 1,134 words
Fable 5.2 One-Shotted My Nephew's Dream Game
Lex Toumbourou describes how Claude Fable 5.2 built a complete, fun multi-level game from his nephew's drawings in one shot, covering play testing, how the drawings became the game, and the final token tally.
8 min · 1,761 words
Locating the source of a mysterious beep with Claude
Simon Willison had Claude Code build Beep Locator, a web app that turns several phones, tablets and laptops into networked microphones to locate an intermittent beep in a room full of unpacked boxes, and reflects on letting agents perceive the world.
9 min · 1,990 words
Jina Yoon of PostHog's build mode newsletter argues that chatty LLMs are overkill for simple structured questions and makes the case for decision models that return a typed answer with a probability instead.
9 min · 1,966 words