Results for “context engineering”

…(in coinciding with Claude Tokyo). It is a feat of incredible engineering (and commitment to access) to make these research models GA, and the benchmarks are great… with asterisks. Here they are on yesterday’s brand new, out of…

…They also published a post on Generalizing an LLM from 8k to 1M Context using Qwen-Agent. {% if medium == 'web' %} Table of Contents [TOC] {% else %} The Table of Contents and Channel Summaries have been moved to the web version of this…

…SOTA claim on ARC-AGI-2: Extra API additions: effort control, context compaction, and advanced tool use. And Claude Code is now bundled with Claude Desktop, with Claude for Chrome and Claude for Excel rolling out to even more users. The…

…on AI scaling over this decade. Documents, OCR, retrieval, and context engineering for real workflowsDocument AI is trending toward end-to-end multimodal parsers with grounding: Baidu introduced Qianfan-OCR, a 4B end-to-end document…

…[TOC] Nous Research AI Discord Summary Breaking the LLM's Context Window Limit with Activation Beacon: @kenakafrosty shared an arXiv paper on Activation Beacon, a new solution that could potentially solve Large Language Models (LLMs)…

…among open weights they evaluated). Details include 200K context, text-only I/O, 355B MoE total / 32B active, MIT license, and practical deployment note: ~710GB BF16 weights means it won’t fit on a single 8×H100 node (~640GB)…

…tasks, e.g. @hqmank’s globe dashboard test. On software engineering, DataCurve said K3 debuted at #3 on DeepSWE, calling it the first open-weights model with frontier-level results there. ARC and cyber remain useful reality checks: ARC…

…the challenge of catastrophic forgetting remains unresolved. Context and Inference: There is interest in the potential for Titans to achieve a 200k context window with high accuracy, though concerns remain about inference speed and accuracy…

…announced that Claude can now reference past chats to maintain context, a feature @swyx calls instructive for how they solve problems with transparency and user control. On a related note, Google Cloud provided a guide on implementing…

…portal, not both (thsottiaux). Agents, Benchmarks, and Harness EngineeringAgent evaluation is becoming more long-horizon and workflow-grounded. Bespoke Labs released AutoResearchExam, a benchmark spanning 29 open-ended ML and engineering…

…and code generation capabilities of models like DALL-E. Prompt Engineering and Custom Instructions: In prompt-engineering and api-discussions, users discussed creating and applying custom instructions to GPT models, tool recommendations for…

…(Instruct & Thinking): MoE with ~3B active params, 256k–1M context, multilingual (32 languages), aiming at GPT‑5‑Mini/Claude Sonnet parity and shipping FP8 variants. Artifacts across chat, GitHub/cookbooks, API, ModelScope, Hugging Face,…