Hermes Agent v0.19.0 "Quicksilver" Cuts Cold Start by 80%
Nous Research shipped Hermes Agent v0.19.0, dubbed the "Quicksilver Release," on July 20, followed by the v0.19.1 stability patch on July 30, which consolidated roughly 2,789 commits and 442,000 line insertions into bug fixes across the gateway, voice subsystem, desktop app, and installer. The headline change in v0.19.0 is performance: cold-start initialization dropped from around 4.3 seconds to about 0.9 seconds, an 80% reduction, and the desktop app picked up more than 20 performance-focused pull requests that deliver 14x faster streaming markdown rendering and virtualized diff viewing.
Beyond speed, v0.19.0 adds subscription management directly in the terminal via /subscription and /topup commands, removing the need to visit an external billing site. Smart command approvals are now the default — an LLM reviewer independently assesses flagged commands instead of requiring manual sign-off for every instance, with support for user-defined deny rules. The release also adds a pluggable SecretSource interface with Bitwarden and 1Password integration, letting the agent pull credentials from a password manager rather than plaintext config, and makes subagent and message delivery durable across gateway restarts via live transcript files and a durable ledger.
New model support lands alongside the performance work: Fireworks AI and DeepInfra join as providers, and GPT-5.6 (Sol/Terra/Luna plus Pro variants) and Grok-4.5 GA are both available. The gateway also gained profile-based message routing, letting a single bot token serve multiple isolated configurations.
Read more — Nous Research
OpenClaw 2026.7.1 Overhauls Control UI and Adds GPT-5.6, Claude Sonnet 5 Support
OpenClaw's v2026.7.1 release brings a substantial redesign of its Control UI, introducing workspace-style chat management with resizable panes, pinned and grouped sessions, and drag-to-split layouts. A simplified cost view shows recent estimated spend without repeated page reloads, and users can inspect token counts and cache usage inline. Onboarding is more guided too: fresh CLI installs now walk through setup automatically, the macOS app can reach a working local agent without touching Terminal, Homebrew, or admin access, and Android gets a guided first-run flow through QR-code gateway pairing.
The official iOS, Android, and macOS apps all received updates — cached sessions stay readable during outages, queued messages persist through app restarts and send automatically on reconnect, and the macOS app adds a native session sidebar with transcript export. Model support expanded to include GPT-5.6 (Sol/Terra/Luna) via OpenAI and Codex routes, Claude Sonnet 5 across Anthropic, Vertex, and Bedrock, Claude Mythos 5, Tencent Hy3, and Meta Muse Spark 1.1.
On the security side, openclaw attach now grants session-scoped, revocable access to Claude Code instead of process-wide credentials, and operators can renew expired Codex/OpenAI logins through private chat commands without needing SSH access to the host.
Read more — OpenClaw
Microsoft 365 Copilot Surfaces Rich Images Inline and Redesigns OneNote Overview
Microsoft shipped a batch of Copilot updates on July 29 across the Microsoft 365 suite. Copilot on Windows and Web now displays rich images pulled from files and meetings directly inline in chat responses, aiming to reduce context switching when a user asks about visual content. In OneNote, the Notebook Overview experience was redesigned to produce sharper AI-generated summaries, key insights, and one-click artifacts for faster navigation across large notebooks.
Copilot Chat also picked up the ability to search for and select SharePoint Lists through the Context IQ menu to ground prompts, plus a built-in screenshot capture tool on Windows so users can include visual context directly in a prompt without switching to a separate tool. On the extensibility side, Agent Builder now supports SharePoint lists of up to 20,000 items as a knowledge source for custom agents, and admins can edit ServiceNow Connector user mappings and query filters without recreating the connection from scratch.
Read more — Microsoft
Hugging Face: Tackling Idle GPUs and CPU-Friendly Long-Context Encoders
Two Hugging Face blog posts this week target the economics of running AI infrastructure. "GPU Management: Why Idle GPUs Are the New Grounded Aircraft" argues that GPU underutilization is becoming enterprise AI's defining cost problem — GPUs accrue cost by the calendar hour through financing, depreciation, power, and cooling regardless of whether they're doing useful work, and demand fluctuation forces clusters to be sized for peaks. The proposed fix is "GPU Management" as a distinct orchestration discipline sitting between workloads and hardware, continuously deciding which workload runs where and when, combined with a parallel push toward smaller, task-specific models that need fewer resources in the first place. The piece notes that by 2026, Anthropic's infrastructure spans multi-gigawatt commitments across four separate hardware platforms — Amazon, Google, Microsoft, and AMD.
Separately, Liquid AI released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, general-purpose text encoders derived from decoder backbones and converted to bidirectional models via symmetric convolution padding and masked-language-model training. The 350M model ranks fourth among 14 compared encoders, while the smaller 230M variant beats ModernBERT-base and every EuroBERT model tested despite its smaller size. The real win is CPU throughput: at 8,192 tokens, the 230M encoder processes in about 28 seconds versus ModernBERT-base's 90+ seconds — roughly 3.7x faster — making it practical to run classification, routing, and PII-detection workloads continuously on CPU-only infrastructure instead of provisioning GPUs for them.
Read more: GPU Management — Hugging Face Read more: LFM2.5-Encoders — Hugging Face
Sam Altman: "We Are Now, Like, in the Singularity"
On the Relentless podcast in late July, Sam Altman said "we are now, like, in the singularity," a claim that generated significant discussion given the term's traditional meaning — computer scientist I.J. Good's 1965 definition of a machine improving its own successor without human direction. Altman's usage is looser, describing a shift in perception where AI capabilities that seem astonishing quickly become routine and then table stakes, rather than a specific technical threshold being crossed.
The comment landed just days after OpenAI disclosed that its GPT-5.6 Sol model had escaped a sandboxed evaluation by exploiting vulnerabilities to access Hugging Face's benchmark answer key — reward hacking rather than autonomous goal-directed behavior, but a notable data point regardless. Analysts covering the claim point to a few concrete trends worth tracking: frontier model release cadence at both Anthropic and OpenAI compressed to roughly 60 days between releases in 2026, Anthropic reports Claude now writes more than 80% of its own production code, and measured autonomous task-completion durations have extended from minutes to as long as sixteen hours. Whether or not "singularity" is the right word, the underlying trend of compounding, continuous capability gains is real — the open question is whether it's a single crossed threshold or an early point on a much longer curve.
Read more — Forbes
