Large language model (LLM) news — GPT-4, Claude, Gemini, Llama, Mistral and the latest research on training, fine-tuning, RLHF, and deployment of LLMs.
Shipd.ai, operated by Datacurve, runs a contributor platform where engineers from 40+ countries solve STEM problems — software engineering, machine learning, and data science — to produce training data for frontier LLMs.…
A proposed method called exploration-guided prompt scaffolding aims to fix the equal rollout budget that standard online reinforcement learning gives to training prompts, which vary widely in how informative they are for…
A developer building an AI technical interviewer replaced LLM-driven conversation flow with a finite state machine after the model repeatedly skipped stages, ended interviews early, and mishandled transitions roughly 20%…
A developer has designed a deterministic AI agent execution runtime in TypeScript that replaces the naive LLM agent loop with a Finite State Machine architecture built around four phases: Observe, Propose, Verify, and Co…
A calculator built by Sunk Cost estimates that a $3,499 Apple M5 Max Mac Studio with 64 GB of unified memory running Alibaba's quantized Qwen3.8 27B model would need 44 years to break even against hosted inference for a …
Researchers introduced LLaDA-UI, a block-wise diffusion vision-language model designed for GUI agents that must repeatedly perceive screen states and emit actions. The work applies diffusion large language models' block-…
A code review vendor's evaluation of 50 public benchmark pull requests from Cal.com, Sentry, Discourse, Keycloak and Grafana found that a cheap model, Luna, cost $0.0030 per verified bug versus $0.061 for the frontier mo…
A comparison published by Entelligence benchmarked GPT-5.6 Luna and GPT-6 Astra on identical pull requests, finding Luna cost $0.0041 per review versus Astra's $0.113 but caught 69 verified bugs at 74% precision against …
Anthropic cut Fable 5.1's cache read price to $0.25 per million tokens — a quarter of Fable 5's $1 per MTok and half of Opus 5's $0.50 per MTok — extending the break-even for keeping a prompt cache warm from the 62.5-min…
Tencent researchers introduced T-Mem, a long-term memory system for LLM agents that pre-saves contextual triggers at memory write time rather than relying on semantic similarity at retrieval. The approach sets new state-…
A Show HN project called Sunk Cost calculates how long it takes for a local large language model rig to pay for itself versus using API access. The tool estimates local inference speed as memory bandwidth divided by byte…
A Hugging Face forum user asked how to improve a Google Gemini-Pro LLM product-query model built with the LangChain framework's Agent and Tools, and whether to switch to a Meta Llama model with fine-tuning. A respondent,…
A developer released synthadoc, a tool offering 30 domain-specific templates for LLM knowledge bases that install with a single command. Each template provides six structural components — query routing tables, domain-tun…
Google confirmed that Gemini Live now supports Deep Research, letting users request and discuss in-depth research reports through voice conversations. The integration, announced as part of Google's Back to School 2026 up…
MindMapAny generates an intermediate indented Markdown outline before converting it into the editor's node structure, keeping mechanical bookkeeping like node IDs and sibling order out of the language model's task. The p…
In a recent Agent Factory podcast episode, host and guests discussed "harness engineering" for coding agents, with engineer Ryan Lopopolo, who coined the term "agentic harness," explaining how to sculpt smaller feedback …
An interactive site, One Million Tokens, visualizes what a 1 million-token context window represents, converting it to roughly 750,000 words, 3,000 printed pages, 83 hours of conversation, or 75,000 lines of code. The si…
An analysis of 2,218 instrumented production calls over 30 days found that phi4:14b delivered only 25.3 tokens per second to callers versus its 124.7 tok/s Ollama-reported decode speed, losing 79.7% of throughput with a …
MobileVLA-R1 2.0 has been released, an update to the vision-language-action system that adds reinforcement-learning-enhanced reasoning for mobile robot control. The system targets the gap between high-level semantic reas…
A user running hybrid GPU+CPU inference on 4x Nvidia V100 GPUs reported beating an Nvidia RTX 5090 in token generation (TG) for Qwen 3.8 27B and outperforming two DGX Spark systems with Qwen 3.8 Flash Next, according to …