Results for “context engineering”

…and reusable skills. Harnesses are becoming first-class engineering artifacts: A recurring theme across tweets is that the useful part of agent systems is increasingly the runtime/harness, not the base model alone. DSPy 3.2 shipped RLM…

…include Olmo and AMD's 1B model. The model has a 17-minute context window and could theoretically be fine-tuned like other transformers using audio datasets. Real-world latency (120ms) outperforms GPT-4o (320ms average) and approaches human…

…space” and recursion as a capability multiplierCore idea (2 context pools): RLMs are framed as giving models a second, programmatic context space (files/variables/tools) plus the token space, with the model deciding what to bring into…

…Day 2 will focus on Agent Engineering, while Day 3 will have IRL workshops and the new Online track. {% if medium == 'web' %} Table of Contents [TOC] {% else %} The Table of Contents and Channel Summaries have been moved to the web version …

…Special Shoutout: Swyx will be curating the Data Council AI Engineering Track in Oakland on Apr 22. You can use LATENTSPACE20 for a little discount. {% if medium == 'web' %} Table of Contents [TOC] {% else %} The Table of Contents and…

…estimates. The design idea—memory as capability, not context—is one of the more practically interesting agent architecture directions in the set. Masked diffusion test-time scaling: @SakanaAILabs shared UnMaskFork, accepted to ICML 2026,…

…priority than where custom instructions are placed in the model context; shared chat examples of prompt engineering and NPC generation. Model Guessing vs Training: One member shared their process by presuming that the model either guesses a…

…in and introspect.App Shots, Accessibility, and Better Computer ContextSwyx [00:10:21]: Yeah. A lot of really hard engineering.Ari Weinstein [00:10:23]: Yeah.Swyx [00:10:23]: I mean, app shots in general, right? Like, I think people don’t…

…and code as an output modality every model bumped up to 128k context length (up from 8k) Training utilized a cumulative of 39.3M GPU hours of computation on H100-80GB (TDP of 700W): 1.5m for 8B, 7m for 70B, 31M for 405B. Llama 3.1 was…

…model? Yann LeCun vs Jim Fan • Google Gemini Pro 1.5 - 1m Long Context, Video Understanding • Groq offering Mixtral at 500 tok/s at $0.27 per million toks (swyx vs dylan math) • The {Gemini | Meta | Copilot} Alignment Crisis (Sydney is…

…and Performance: Gemini 2.0 Flash is noted for its larger context window, which some users argue gives it an edge over models like Claude and o1. This capability is seen as crucial in handling real-world software engineering tasks,…

…blogpost goes into more detail on why this is useful: Context: Old BERTS had ~500 token context; ModernBERT has 8k Data: Old BERTS were on older/less data; ModernBERT was trained on 2T, including "a large amount of code" Size: LLMs these…