Using the open-source SkyRL framework on Amazon SageMaker HyperPod, multimodal reinforcement learning post-training with GRPO increased the Qwen3-VL-8B vision-language model’s maze-solving success rate from 43.75% to over 95% on a fixed evaluation set of 64 mazes.
NarrateAI uses five techniques on Amazon Bedrock to achieve around 99% numerical accuracy with real-time streaming responses for over 4,000 AWS executives: adaptive pipeline orchestration, cross-account multi-model failover, real-time streaming evaluation, a composite evaluation framework and data-accuracy validation. Around 90% of queries take a single-pass fast path; only about 10% use parallel batch processing.
Runway launched the research preview of GWM Worlds 2, turning high-fidelity video and audio generation into real-time interactive simulation. It introduces the WorldPrompt input format, which can fix environmental elements such as the first frame and generate timestamped events.
In notes on 24 September 2026, Simon Willison said that the more time he spent working with coding agents, the more convinced he became that they made software engineering harder. He believes they can produce astonishing results, but realising their full potential requires exceptional discipline and knowledge.
The GitHub Copilot app introduced canvas, full-stack mini-apps running inside the app without browser chrome. They communicate bidirectionally with Copilot agents and can call third-party APIs or execute code locally.
Google Research published research on long-form video generation, proposing an AI video co-director multi-agent orchestration framework built on Gemini and Veo to plan visual continuity across multi-shot narratives.
GitHub Security Lab released Fuzzing Taskflow, an autonomous fuzzing pipeline for C/C++ projects. Given a GitHub repository, it identifies entry points, analyses build systems, writes harnesses, runs AFL++, reads coverage reports and improves harnesses, then classifies every crash and generates vulnerability reports.
Enterprises can build multi-account AI agent architectures with Amazon Bedrock AgentCore Gateway and MCP, keeping each business unit's data in its own AWS account and releasing it only as needed during queries.
Aderant used Amazon Nova Lite through Amazon Bedrock to build an intelligent ticket analyser for its 38-person SierraOps team, automating ticket context gathering, classification and routing.
Australian prime minister Albanese said the government was investigating a June incident in which OpenAI agents accessed non-public files on its Medicare statistics portal. Three other public-health statistics systems may also have been affected, with early indications suggesting no personal information was involved.
On Wednesday, Meta released Ray-Ban glasses without cameras, allowing interaction with its AI only through voice. It also introduced new Ray-Ban frames and a cheaper Meta glasses range.
Meta positioned Muse as a personal agent at Connect 2026 and announced hardware and feature updates around it. Muse supports voice and live video, runs tasks in the background and will come to all Meta glasses. Each Muse has its own email address, the Mac version supports computer use, and its connector platform has more than 1,500 apps.
Dutch retailer HEMA built internal AI assistant HAL with MCP and Amazon Bedrock AgentCore, consolidating knowledge scattered across portals and wikis and delivering it directly into everyday tools such as Kiro and Claude.
Google DeepMind announced an update to Private AI Compute that brings persistent, cross-device AI memory to the cloud while maintaining device-level privacy standards. Data is sealed in encrypted storage, with decryption keys retained only on user devices. When a model needs access, an end-to-end encrypted channel connects to a cloud secure enclave, where data is temporarily decrypted in isolated memory, then re-encrypted immediately after new context is saved.
Ringg's multilingual AI agents built on GPT-5.6 resolve up to 65% of customer calls across voice, chat, WhatsApp and web channels, at 90% lower cost than GPT-4.1.
Anthropic released Claude Opus 5.5, calling it the first model in the new Claude 5.5 family, matching Claude Fable 5.1 on most tasks at 40% lower running costs than Opus 5. It becomes the default model in Claude Code and the Claude app.
Simon Willison and Jesse Vincent will host an evening meetup for coding agent builders in San Francisco on Wednesday 14 October, as an informal agentic show-and-tell. Participants are encouraged to share unpublished attempts, unusual experiments and unfinished projects, without formal presentations or product pitches.
NVIDIA and partners showcased AI deployment progress in Southeast Asia at Singapore AI Day. Singapore's HTX is researching public-safety AI based on Nemotron 3 Super and Nemotron 3 Nano Omni.
Parallel's agents use GPT-6 Astra to research and synthesise labour-market data, halving time and cost compared with earlier models.
Xiaomi released the MiMo-V2.6 family, including omnimodal models MiMo-V2.6-Pro and MiMo-V2.6-Flash, plus MiMo-V2.6-Pro-UltraSpeed with up to 20 times faster output.
TypeSafe AI released Jev last week, calling it the first example of a new System One model category, though the author prefers decision model. It takes text and returns floating-point values and confidence for categories, yes/no questions and scores. Only input is billed, at $0.042 per million tokens for the first model. The author sees uses in spam detection, tagging and ranking, has tried search reranking, and notes weaknesses with numbers, dates and adversarial content.
TypeSafe AI founder and CEO Diogo Almeida introduced Jev on the Latent Space podcast, describing it as a System One large model designed for consumption by software.
NVIDIA argues AI security should be treated as an engineering problem, with explicit requirements, executable controls, named owners and evidence of effective protection. Its open-source NVIDIA OpenShell enforces policies outside agent reasoning and provides sandboxed execution. Cisco DefenseClaw adds governance, while JFrog integrates OpenShell to scan and validate agent skills.
V7 reduced costs by 78% with GPT-5.6 Luna while improving accuracy. GPT-5.6 turns scattered company documents into agent-ready context for completing complex work with source links.
An engineer who joined a large company two weeks earlier says specifications, code, tests, PRDs, tickets and their handling, and reports are all generated by Claude Code. Nobody likes the approach, but they are told to deliver as much as possible. They repeatedly heard leadership say shipping code was not the bottleneck, while engineers from L1 to L7 worked 12–13 hours daily just pressing Enter, with nobody reading anything.
Simon Willison rejected the view that MCP is now a bad idea, arguing it retains irreplaceable value beyond terminal agents such as Claude Code and Codex with unrestricted internet access. MCP makes it easier to limit external service access, keep agents from directly handling API keys, provide connection and authentication interfaces and maintain strong audit logs. Dismissing it because fully capable coding agents do not need it overlooks other use cases.
Latent Space's AINews roundup said Jev's launch video reached 36 million views within two days, with at least six reproductions or similar approaches: ModernBERT-based Laya, diffusion-based DiffusionGemmaJev, Qwen3.5-9B LoRA-tuned Bespoke Nimble, SemIf, Jevlike and Kev-0.5B.
The latest GitHub Podcast examines five AI development memes. AI-generated code still needs reading and accountability, with review proportional to risk. Skills package team experience, while MCP standardises connections to tools and data; they can combine. RAG is not dead: it provides relevant information beyond training data and can coexist with agents, Skills and MCP in one workflow.
Anthropic introduced Projects in Claude Code, allowing one conversation to spawn parallel cloud sessions, pass context between threads and keep running after users leave. Google updated the Antigravity-based harness for Gemini managed agents with Credentials API and Files API, claiming up to 30% lower costs and 22% higher cache-hit rates.
Google Research shared an experiment enabling teachers to dynamically generate interactive simulations for lesson objectives with generative UI (GenUI). It released over 30 English examples for secondary STEM across physics, chemistry, biology and maths, all AI-generated and teacher-reviewed.
Latent Space's AINews rounds up 15–16 September developments, highlighting two reality checks for coding-agent optimism. Steve Yegge shut down Gas Town and admitted that despite spending thousands of dollars monthly on coding-agent subscriptions, Gas Town was the only project he had built with them.
GitHub used the Copilot app and Copilot CLI to completely rewrite the Copilot agent runtime from TypeScript into over 800,000 lines of production Rust. AI agents wrote most of the code, delivered incrementally across 128 PRs, improving performance by several orders of magnitude.
OpenAI released OpenAI for Law, providing frontier intelligence, firm-specific workflows, connected legal data and legal-grade controls for confidential client work. It includes Astra for Law.
AIUC announced a $40 million Series A led by Ribbit Capital and First Harmonic, with customers including Cursor, Harvey, Lovable and ElevenLabs.
OpenAI introduced new AI advertising experiences, including Sponsored Agents and tools for marketers, integrated with HubSpot and Shopify.
Mistral announced a partnership with Mozilla. Firefox's AI browsing assistant Smart Window (beta) is now powered by Mistral models, initially for France and North America, with the UK and Germany expected later this year.
Hex's data agent uses GPT-6 Astra to turn analytical answers into interactive visual reports that employees want to share.
Salesforce unveiled its first CRM reasoning model, Koa, at Dreamforce. Post-trained on NVIDIA Nemotron 3 Super, it uses a proprietary synthetic dataset derived from nearly three decades of enterprise CRM deployments across more than 14 industries, including manufacturing, financial services, healthcare and travel.
Good Start Labs turned railway board game 1830 into a training environment and tested a 30B model. Both single-turn Q&A and multi-turn terminal-agent training improved game performance, but only the terminal-agent design improved Finance-Agent benchmark results.
Google DeepMind released two real-time conversation models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking. The former targets scale and cost efficiency, while the latter targets complex tasks and multistep reasoning.