AI News Summary 2026-08-17

También disponible en:

English

.. lang: es

AI News Summary — 2026-08-17

GAFAM y grandes empresas de IA

  • OpenAI published The builder’s guide to GPT‑5.6 and Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed on Aug 13. The messages are clear: GPT‑5.6 pushes price-performance for agent workflows, and Ultrafast brings very high API throughput for latency-sensitive use cases. Sources: builder’s guide · Ultrafast preview
  • Google announced Sheets canvas, a Gemini-powered layer that turns spreadsheet data into interactive mini-apps, and also expanded connected apps/services for Gemini. The overall direction is familiar Google AI: move more prompt-driven capability into everyday Workspace and app surfaces. Sources: Sheets canvas · Gemini connected apps

Influencers y blogs técnicos

  • Simon Willison published Qwen 3.8 27B is excellent, but it defaults to wildly overthinking things on Aug 16. It’s a useful early read on a capable model that may need tighter tuning for agentic use. Source: post
  • He also published Markdown SVG upgrades on Aug 16, another signal from a consistently high-signal technical source. Source: post

Imagen generativa

  • Midjourney’s updates page shows the banner “V8.2 is live!” and links to the new version post. The page itself is sparse, but it’s an official sign that Midjourney has shipped a new version in the image-model line. Source: updates page
  • Anthropic’s watermarking explainer matters for the broader generative-content ecosystem because provenance and compliance are becoming part of how AI output is treated. Source: Claude text watermarking

Chatbots y agentes

  • Hermes Agent v0.20.2 landed on Aug 16 with a broad patch release across desktop, CLI, gateway, cron, auth, and installer hardening. For agent builders, this is the most relevant agent-framework update of the day. Source: release notes
  • OpenClaw’s latest release is still 2026.7.1-2 from Aug 4, so it stayed on watch but did not make the top tier today. Source: release page

IA local y serving

  • llama.cpp’s latest release, b10456, shipped on Aug 17 with a sycl kernel-launch fix for quantized cpy kernels. The headline number is striking: the q4_0→f32 path on Arc 70 improved from 20.21 GB/s to 158.19 GB/s. This is the day’s lead story. Source: release notes
  • vLLM published DSpark adaptive verification on Aug 14, showing another concrete push toward better speculative decoding and serving efficiency at higher concurrency. Source: blog post
  • Ollama announced NVIDIA Nemotron 3.5 Lightning, a local-friendly 30B/3B-active agent model with a 1M token context. It’s older than today’s freshest release, but still relevant for local inference readers. Source: blog post