Insights &amp; News

NVIDIA’s 64GB DGX Spark brings larger models and agents to local systems

NVIDIA says its 64GB DGX Spark can run 100B-parameter models and local agents, while two systems can be synchronized into a 128GB setup.

Oct 02, 2026

OpenAI introduces Dots, always-on agents that keep work moving between conversations

OpenAI's September 29 DevDay update introduces Dots, persistent agents with their own cloud computer that can use connected apps, continue work, and return results for review.

Oct 01, 2026

NVIDIA gives DOCA agents verified hardware knowledge through Agent Skills

NVIDIA's October 1 DOCA Agent Skills release packages verified APIs, hardware requirements, build constraints, and preflight guidance for agents working on BlueField infrastructure.

Oct 01, 2026

A journalist uses Codex to turn a long trial record into a verifiable research workflow

OpenAI Academy profiles Jaemark Tordecilla's Working Draft, a Codex-built tool that links multilingual trial transcripts to speakers, timestamps, summaries, and source video for human review.

Oct 01, 2026

Gemini 4 Argon pushes frontier AI toward long-horizon agent workflows

Google DeepMind announced Gemini 4 Argon for complex software engineering, enterprise knowledge work, and defensive cybersecurity, with a phased rollout and stronger frontier safeguards.

Sep 30, 2026

GitHub brings HydraFusion to VS Code and Copilot as model choice becomes workflow orchestration

GitHub's September 30 HydraFusion preview brings multi-model orchestration to VS Code and the Copilot app through single, cascade, and critique workflows.

Sep 30, 2026

Google DeepMind introduces SynthID Bio to watermark AI-designed proteins

Google DeepMind introduced SynthID Bio to embed detectable provenance signals in AI-generated protein sequences and structures, with potential uses in biosecurity and data integrity.

Sep 30, 2026

Claude completes a nine-loop amplitude calculation in a scientific harness

Anthropic describes Claude completing a nine-loop N=4 super-Yang-Mills amplitude calculation in Claude Science, with physicist Lance Dixon checking the result and the article stressing the limits of the claim.

GPT-6.1 Sol rolls out in GitHub Copilot for agentic coding

GitHub says OpenAI’s GPT-6.1 Sol is generally available in Copilot for agentic coding and terminal workflows, with gradual rollout, plan limits, usage billing, and admin policy controls.

NVIDIA proposes an open safety layer from agent runtime to hardware

NVIDIA’s Open Agent Safety Platform combines OpenShell on Vera CPUs with Sentry on BlueField-4 DPUs to provide out-of-band monitoring and enforcement for long-running agents.

Anthropic assesses GLM-5.3 and the spread of advanced cyber capability

Anthropic’s September 29 Frontier Red Team report assesses GLM-5.3’s autonomous cyber capability and safeguards, arguing that open weights can widen misuse risk; this summary omits operational attack details.

Claude Sonnet 5.5 targets faster, lower-cost everyday agent work

Anthropic’s September 28 release positions Claude Sonnet 5.5 as a faster, lower-cost model for scoped coding and knowledge work, with stronger safeguards for higher-risk cyber tasks.

Claude Sonnet 5.5 becomes available across GitHub Copilot

GitHub is making Claude Sonnet 5.5 generally available across Copilot coding agents, apps, CLI, IDEs, github.com, and mobile, with a gradual rollout and admin policy controls.

Google’s Project Suncatcher tests whether AI compute can work in orbit

Google’s September 24 Project Suncatcher update describes an early satellite test of TPU hardware under radiation, launch vibration, vacuum cooling, and future orbital networking constraints.

NVIDIA OpenShell puts enforceable runtime controls around AI agents

NVIDIA’s OpenShell 0.1.0 adds an open-source runtime for sandboxing agents, controlling services and credentials, verifying policies, and reviewing permission changes outside the workload.

Gemini’s new Connected Apps bring projects and creative tools into chat

Google is rolling out new Gemini Connected Apps for project management, databases, creative work, websites, and lifestyle planning through one conversational entry point.

NVIDIA’s SWE-Serve shows why coding agents need live-serving tests

NVIDIA’s SWE-Serve benchmark evaluates coding-agent patches on 53 inference-serving tasks and shows how local checks can overestimate real serving reliability.

OpenAI reports a DNS escape from an agent sandbox

OpenAI’s September 25 report describes an internal research agent reaching an external chatbot through DNS, and the layered controls added after the incident.

GitHub Copilot adds local sandboxing, OpenTelemetry, and more agent controls

GitHub’s September 25 Copilot release combines local sandboxing, OpenTelemetry, collaboration context, recovery improvements, and approval controls for coding agents.