AI Daily · 2026-10-28

Oracle is today's central thread: it expanded its OpenAI partnership to bring Oracle applications and data into ChatGPT and announced AI Workspace, Security Mesh, an SMB bundle, and App Composer, while also planning with NVIDIA to scale OCI Superclusters to multi-gigawatt Vera Rubin NVL72 systems. On the research side, Google introduced a sleep framework that consolidates memory through a dreaming phase, letting LLMs keep learning after deployment with less catastrophic forgetting; LiquidAI released LFM3, a 2B-class multimodal and omni family; and NVIDIA shipped the open-weight Nemotron 3.5 Transcribe and Granary v2. Ai2's benchmark suggests LLM novelty judging is unreliable, with small prompt changes swinging accuracy by more than 50 points. Application stories clustered around agentic work: Legora, Retell, and LangChain showed efficiency gains in legal, call-center, and design workflows, Anthropic highlighted a brain-versus-hands split to isolate execution sandboxes, and Simon Willison called October 2026 an inflection point for vibe game development.

North America · First-hand

OpenAI

⭐⭐⭐ [Product Update] OpenAI expands partnership with Oracle to bring its apps to ChatGPT

OpenAI and Oracle are expanding their partnership to bring Oracle applications and data into ChatGPT and to power new Oracle products and services with OpenAI models and tools. The companies say this builds on an existing commercial relationship spanning the OpenAI API, ChatGPT Work, and Codex; Oracle reported 130,000 active ChatGPT users and more than 95,000 active Codex users, and said its recruiting team uses ChatGPT Work for market research that once took two to four days to compile and now takes 15–20 minutes. Oracle is developing Oracle AI Workspace to bring these integrations into ChatGPT, so people can describe what they need to accomplish, find information, and complete approved workflows within their organization's permissions and approvals — illustrated by a new finance team member investigating costs, purchasing decisions, and what remains before month-end close. Oracle also announced offerings built with OpenAI models and tools: Oracle AI Security Mesh, which uses OpenAI cyber models to find and fix vulnerabilities in applications and infrastructure; Oracle AI SMB Bundle, which combines ChatGPT Business and Codex with Oracle applications and recurring Oracle Cloud Infrastructure credits for small businesses; and Oracle App Composer, for building apps with Codex and other AI tools, connecting them to approved Oracle data and workflows, and deploying on Oracle Cloud Infrastructure. U.S. Oracle customers can purchase access to the OpenAI API, Codex, and ChatGPT Work through Oracle Marketplace using existing procurement processes and eligible Oracle Universal Credits.

⭐⭐ [Other] Legora is building the agentic operating system for legal work

OpenAI published a customer story on Legora, founded in Stockholm in 2023, which CEO Max Junestrand describes as building "the agentic operating system for legal work." Legora uses OpenAI models via API, Codex and ChatGPT to power core workflows of its legal AI platform, including agentic legal work and large-scale document review. Cited figures include a 95%+ reduction in time Legora's own legal team spends reviewing NDAs, 90% faster contract turnaround for a Fortune 100 customer, and 9.5 hours returned per lawyer per week. Internally, Legora uses ChatGPT Work, Codex and dots; CTO Jacob Lauritzen says engineering can explore 20 solutions at once and pick the best within days, and coding agents can search millions of lines of logs, sometimes going from bug report to working fix in about five minutes. General Counsel Emma Kadri Bergström says NDA first-pass review went from about 20% of the legal team's capacity to under 1%.

⭐⭐ [Other] Retell uses OpenAI models to help power 100M monthly calls

OpenAI published a customer story on Retell AI, a voice AI platform for contact center operations that now serves more than 4,000 businesses, with monthly call volume growing from 30 million in early 2026 to over 100 million by September. OpenAI models generate responses for 80% of calls on Retell, customers report resolution rates above 80%, and the company has consumed more than 1 trillion OpenAI tokens. Retell's own vCX benchmark compares dozens of models on difficult multi-turn conversations from real production scenarios; in its October 13, 2026 run, GPT-6 Astra ranked first on instruction following at 89.7%, and pre-launch evaluations showed GPT-5.5 leading on non-hallucination rate (88.0%) and tool-call correctness (88.3%) with reasoning enabled against Gemini-3.1-Pro and Claude-4.6. Retell engineers use Codex for code research, feature building and debugging, and GPT-5.6 Terra to analyze call transcripts afterwards; the company cites a $1.6M annual revenue run rate per employee.

Google

⭐⭐⭐ [Research] The sleep framework: A new way for LLMs to retain knowledge in continual learning

Google Research introduces a sleep phase for continual learning that lets large language models continuously integrate new facts and improve reasoning while mitigating catastrophic forgetting and static knowledge cutoffs. The post notes that continual parameter updates trigger catastrophic forgetting, while relying on in-context learning limits retention to the current context window. Inspired by how the brain stabilizes memory during rest, the framework has an active (wake) phase where the model processes inputs and stores temporary information in fast short-term memory, and a sleep phase where it pauses new input and redirects computation to memory consolidation and self-improvement within a fixed architecture and human-defined training rules, unlocking pre-allocated Mixture-of-Experts memory blocks and updating weights from self-generated practice examples. The sleep process has two stages, memory consolidation and dreaming; consolidation uses parameter expansion and a continuum memory hierarchy with the Hope architecture to integrate newly unlocked low-rank experts into stable lower-frequency memory blocks, and uses Generalized Knowledge Distillation with RL-based imitation learning to move fragile short-term memories into the expanded long-term parameters. The authors report that the framework lets models continually acquire new facts, retain long-context knowledge up to 10M tokens, and improve complex reasoning.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] Build specialized multimodal models with LFM3

LiquidAI released LFM3, a new generation of Liquid Foundation Models positioned as a starting point for specialized intelligence, bringing text, vision, audio, omni and decision models onto one shared backbone in the 2B parameter range. Compared with the LFM2 family, LFM3 makes four main changes: a new tri-hybrid architecture and training recipe let LFM3-1.7B outperform LFM2.5-2.6B with 35% fewer parameters; six 2B-class models, including a first omni model and video input for the vision model, are trained from LFM3-1.7B-Base; the text model fits in about 1 GB at 4-bit precision with new in-house vision and audio encoders, and its 256K context window targets on-device agents; and every model fine-tunes on a single GPU. LFM3-1.7B-Base has 24 layers combining 13 gated short convolutions with kernel size 7 (up from 3 in LFM2), 5 sliding-window attention layers with a 256-token window and 6 global attention layers, plus head-wise gated attention where a sigmoid gate scales each head's output so a head can switch off when it has nothing to add. The architecture came out of hardware-in-the-loop search covering 14 device and backend combinations, from phones and a Raspberry Pi 5 to laptops, cloud CPUs, GPUs and the Qualcomm Hexagon NPU.

⭐⭐ [Model Release] NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt

NVIDIA released Nemotron 3.5 Transcribe, an open-weight, promptable model available on Hugging Face and ModelScope that combines transcription, speaker diarization, translation, and audio question answering in a single checkpoint. It transcribes 40 languages, translates 31 of them to or from English, diarizes 13, handles recordings up to an hour long with up to eight speakers, and supports word-level timestamps, with natural-language prompts guiding output format and domain vocabulary. The post says it ranks first among open-weight models on the Hugging Face Open ASR Leaderboard for English, posts the lowest error rate on David AI's 21-language diarization leaderboard, tops VoiceArena's 14-system comparison, and runs on a single NVIDIA GPU. NVIDIA also released the Granary v2 multilingual dataset and curation pipeline plus NeMo Speech fine-tuning tools; in a medical case study, fine-tuning cut word error rate on held-out speakers by 45.8% relative to the best prompt. The post also notes Agora used word-level timestamps in a classroom demo.

Product Update

⭐⭐ [Product Update] Thin, Light and Built for Everything: NVIDIA RTX Spark Windows Laptops Gain New Models, Agent Tools and Games

NVIDIA published a blog post on its RTX Spark line of thin-and-light laptops, positioned for the work developers, creators and gamers do. The laptops are now available from ASUS, Dell, HP, Lenovo, Microsoft and MSI. According to the post, these systems combine an NVIDIA Blackwell RTX GPU and an NVIDIA Grace CPU in a single superchip with up to 128GB of unified memory. The headline also references new models, agent tools and games, though the post body offers limited detail.

⭐⭐ [Product Update] How We Built Our Design Agent

LangChain describes how it built Content Studio, an in-house design agent that lives in Slack: a team member describes what they want and attaches content or briefs, and the agent returns finished assets such as videos, social graphics, recreated UI images and slide decks, with every asset written in code. The post lists three requirements: writing designers' judgment into the agent's skills and checks, a custom environment and sandbox spanning a video engine, image tools and a shared canvas, and a multiplayer UX that shows the plan, live progress and an approval step for jobs that run for hours on async subagents. Content Studio is built on Managed Deep Agents, and the team has open-sourced the harness so others can add their own brand tokens, fonts and skills and deploy it to Slack. Reported results include more than 3,000 assets and over 80 slide decks produced, roughly 4.5M views on X for posts featuring its work, and about 2x likes and 3.7x (figure truncated in the source) on the LangChain account. The post also advises deleting rules written to compensate for model weaknesses as models improve, keeping only the design principles.

Research

⭐⭐⭐ [Research] Can we trust LLMs to judge whether a research idea is novel?

Ai2 released a study, 「Old Ideas, Novel Problems」, examining how reliably large language models can serve as novelty judges for scientific research ideas, publishing a paper along with a benchmark, data, code, and an automated pipeline for building new evaluation sets. The benchmark was built from ICLR 2026 peer reviews: 154 high-novelty ideas came from accepted papers in the top 10% of their area whose reviewers explicitly affirmed novelty with no dissent, while 145 lower-novelty ideas came from rejected papers in the bottom 10% whose reviewers raised originality concerns and no reviewer affirmed novelty; in a 50-item check, an annotator agreed with the extracted novelty-signal polarity in 47 cases. Across six LLM judges (GPT-5.1, GPT-5.2, GPT-5.4, Claude Sonnet 4.5, Claude Opus 4.5, Claude Opus 4.6), small changes to the evaluation setup produced dramatically different verdicts on the same ideas — changing just a few words in a prompt shifted accuracy at identifying the more novel of two ideas by more than 50 percentage points. Retrieval and additional reasoning did not consistently improve judgments, and two purpose-built novelty evaluation systems performed worse than a simple prompted Claude Opus 4.6 baseline. The authors conclude that novelty evaluation itself is far from reliable and that more dependable methods are needed to tell whether AI is actually proposing something new.

Opinion

⭐⭐ [Opinion] The October inflection point for vibe game development

Simon Willison argues on his blog that October 2026 marks an inflection point for vibe game development, comparing it to November 2025, which he had called an inflection point for coding agents (Opus 4.5 and GPT-5.1). He says Claude Opus 5.5 shows meaningful improvement in 3D modeling, music generation (citing Scrimshaw Jukebox) and game design, and that telling it to 「Make it a LOT more fun」 actually worked. Over a few weeks he built four games entirely by prompting without looking at the code — including Tree Appreciator, Emberwick and others — all written in HTML and JavaScript and running in the browser; Tree Appreciator was produced in a single shot in the regular Claude chat interface with a prompt asking for three.js and to 「think hard about how to make sure it is both fun and beautiful.」 He says GPT-6 Astra can build games but its sense of fun does not match Opus 5.5, and he expects OpenAI to get there by the end of the year.

⭐⭐ [Opinion] How Anthropic ships Claude with Claude: Separating the Brain from the Hands

This Latent Space podcast episode features Anthropic's heads of Engineering and Product, Katelyn Lesse and Angela Jiang, discussing the Opus 5.5 launch and how Anthropic uses Claude to build Claude. The launch emphasized both model quality and cost: Anthropic says the new model costs less to run than its predecessor, with the exact saving depending on workload and use of cached context, and it addresses feedback that earlier Opus generations were too verbose in pursuit of stronger intelligence per dollar. Less visible launch work involves API integration changes, such as preserving thinking blocks across turns to protect reasoning from extraction for distillation, a new requirement that can also break a customer's existing harness. Early access programs in practice have customers running evaluations in their own harnesses and reporting surprising behavior, which surfaces the model in environments Anthropic's own products do not cover. The episode's favored mental model is the split between an agent's brain and hands: in Managed Agents, the session log and harness are separated from the sandbox that executes code, so a failed sandbox does not take the whole session with it, and credentials are held separately and attached through a proxy so no credential sits in the sandbox. The guests say an earlier approach put the whole coding agent in a sandbox behind an API, adding work before the first response and losing the running agent if the sandbox disappeared; separating the pieces let the platform pursue lower latency and better reliability while keeping execution isolated.

Other

⭐⭐⭐ [Other] NVIDIA and Oracle Expand NVIDIA Vera Rubin to Multi-Gigawatt Scale and Bring NVIDIA Vera CPU to Oracle Cloud Infrastructure

NVIDIA and Oracle announced plans to expand Oracle Cloud Infrastructure (OCI) Superclusters with NVIDIA Vera Rubin NVL72 systems to multi-gigawatt scale. The companies say the expansion addresses growing customer demand for AI training, inference and agentic applications. The announcement also covers bringing the NVIDIA Vera CPU to Oracle Cloud Infrastructure.