artificial-intelligence.
302 writings found
Page 2
Speculative Decoding Meets Vision: LFM2.5-VL Gets the Speed Treatment
Liquid AI's DSpark draft model accelerates vision-language inference by 2.3x on-device and 20x on GPU, but Amdahl's law reveals the real bottleneck.
Speculative Decoding Hits Vision Models: What Developers Need to Know
Liquid AI's DSpark draft model accelerates vision-language inference 2-3x on edge devices and 20x on H100s. Here's what it means for your stack.
Scaling MuJoCo to 2048 Parallel Simulations on GPU
How MJWarp bridges MuJoCo and NVIDIA Warp to parallelize robot simulations for reinforcement learning at scale, without rewriting physics code.
Making AI Evaluations Reproducible: AISI and EvalEval's Open Infrastructure
How AISI and EvalEval are standardizing AI evaluation reporting through shared schemas and open platforms to improve reproducibility and research reliability.
Why Your AI Agent Works Once But Fails the Next Time
Most benchmarks hide consistency problems behind averages. A new diagnostic tool exposes why capable agents produce unreliable results and how to fix it.
Why Your AI Agent Works Once But Fails the Second Time
Consistency gaps in LLM agents are hiding in plain sight. Why average accuracy masks unreliability, and how to measure what users actually care about.
Why Your AI Agent Succeeds 77% of the Time (But Only 53% Reliably)
Most agent benchmarks hide inconsistency behind averages. Here's why that gap matters and how to measure real-world reliability.
OpenAI and Microsoft Knew They Were Breaking the Web
Unsealed court docs reveal internal warnings about AI training destroying publisher economics. What this means for the future of content and web infrastructure.
Why Your AI Agent Works Once But Fails the Next Time
Consistency gaps in LLM agents matter more than average accuracy. Introducing consistency guidelines to stabilize agent decisions.
Snap's Specs Intelligence Shows AI Assistants Are Getting Personal
Snap launches Specs Intelligence, an anticipatory AI service that connects to your apps and knows your context. Here's what it means for AI development.
Why Your AI Agent Works Once but Fails Twice
AI agents achieve high average accuracy but fail inconsistently on identical tasks. A new consistency measurement and guideline system reveals why, and how to fix it.
AI Slowdown Promises Need Teeth, Not Just Talk
Major AI labs are pledging to slow development, but without enforcement mechanisms and global coordination, these commitments risk becoming regulatory capture dressed in safety language.
AI Safety vs Speed: The Republican Pushback on Responsible Development
Trump and House Republicans reject calls to slow AI development, citing national security concerns about China. What this means for developers and the industry.
Gradio Workflows Replace ComfyUI's Node Graph for Most Builders
Workflow1111 shows how gr.Workflow delivers ComfyUI-style node graphs with automatic REST APIs, MCP endpoints, and free parallelism. A practical look at what this means for AI pipeline builders.
NeoMME: The Efficient Multimodal Encoder That Skips the VLM Bloat
NeoMME ditches separate vision towers for unified multimodal encoding. I break down why this matters for document retrieval and what it means for building efficient AI systems.