- NATURAL 20
- Posts
- DeepMind Maps Life Beyond AGI as Gemini 4 Neon Arrives
DeepMind Maps Life Beyond AGI as Gemini 4 Neon Arrives
PLUS: Nvidia stops licensed AI chip shipments to China, Genie 4 expands access, and Microsoft introduces Canvas.

SPONSORED BY CRUSOE
Deploy open models in minutes.
Testing a new AI workflow shouldn't mean spending your afternoon provisioning a GPU cluster.
Crusoe Intelligence Foundry gives you fast, managed access to leading open models without owning or managing the underlying infrastructure.
Use Serverless Inference to deploy a model and start calling it via API in minutes, or Serverless Fine-Tuning to customize a supported model with your own data.
There is no long setup and no long-term commitment. Just a direct path from idea to working prototype.
Every new account gets $5 in free credits, applied automatically at sign-up.
No promo code needed.
Have an agent, research tool, or domain-specific workflow you've been meaning to test?
Spin up your first call and see what it can do.
Today:
DeepMind Institute explores four possible paths from AGI to superintelligence
ChatGPT for Accounting connects business records and finance workflows
Google releases Gemini 4 Neon for complex coding, research and agent tasks
Nvidia stops shipping licensed Hopper AI data center chips to China
Microsoft introduces an interactive AI design workspace called Canvas
AI Is Moving Toward More Powerful and Practical Work
DeepMind researchers are asking how artificial intelligence could evolve beyond human-level general ability. At the same time, OpenAI is bringing specialized accounting work into ChatGPT, and Google has released Gemini 4 Neon for complex coding, scientific and business tasks.
The week also brings new agent tools, broader world-model access and a consequential change in Nvidia's China chip sales. Together, the developments show AI becoming more useful in everyday workflows while raising bigger questions about how capability, oversight and infrastructure will evolve.
The DeepMind Institute published an essay asking what could happen after artificial general intelligence, or AGI. Written by Google DeepMind researcher Tim Genewein and based on the technical report From AGI to ASI, it argues that the steps beyond roughly human-level AI deserve serious scientific and policy attention before they arrive.
The essay treats intelligence as a continuum. AGI represents broad human-level competence, while artificial superintelligence, or ASI, would consistently outperform large groups of human experts across nearly every task of interest, often completing work in far less time or with fewer resources. The authors are exploring a possible future, not announcing that such a system has already been built.
The framework outlines four ways capabilities could advance beyond AGI: continuing to scale models and training data, discovering new technological paradigms, improving coordination among collections of AI agents, and using AI systems to help improve future AI systems through recursive self-improvement.
Genewein also describes cognitive industrialization, the prospect that cheaper compute, better algorithms and greater hardware investment could make many forms of AI-driven work far more plentiful. He draws on historical growth rates to show how rapidly the effective supply of digital cognitive labor might expand if those trends continue.
That projection is deliberately conditional. Compute capacity, energy, data, scientific breakthroughs, governance and physical limits could change the trajectory. The article calls for more rigorous ways to measure progress and understand the consequences for science, the economy and society.

OpenAI introduced ChatGPT for Accounting, combining its Accounting plugin with integrations for common business accounting platforms. The launch is aimed at repetitive work that requires gathering records, comparing transactions and preparing materials for review.
The Accounting plugin can help reconcile bank activity, prepare draft invoice entries, investigate missing or duplicate transactions, organize month-end reports and gather support for audits. Partner plugins connect to services including QuickBooks Online, Xero, Oracle NetSuite, BILL, FreshBooks, Rillet and Campfire.
Instead of manually copying information between spreadsheets and financial systems, users can describe the outcome they need and ask ChatGPT to work with authorized accounting records and templates. It can help prepare a reconciliation, identify differences and assemble draft reports before a human reviews the results.
OpenAI's guidance directs users to set up computer use in the ChatGPT desktop app for macOS or Windows and install the Accounting plugin when it is available to their plan or workspace. Enterprise members may need an administrator to approve it. Connected applications still determine which information and actions a user can access.
The tools do not remove the need for financial controls. Draft journal entries, reconciliations and exceptions should be checked against the source systems, and users need to verify changes before posting them to company books. Product support can differ across accounting providers and integrations.

Google released Gemini 4 Neon as a new member of the Gemini 4 family. Positioned between Carbon for the most demanding long-running problems and Hydrogen for fast everyday execution, Neon focuses on complex software engineering, knowledge work and scientific analysis without requiring the largest model for every task.
At its highest xhigh thinking setting, Google says Neon scored 61 on the Artificial Analysis Intelligence Index, ranking third overall at launch. It also ranked second on Vals Index and on the Agent Arena steerability evaluation, which tests whether agents adapt to changes in user instructions during a task.
The model accepts complex inputs such as spreadsheets, PDFs, charts, images, audio and video inside a one-million-token context window. Google highlights its ability to organize that material into structured deliverables and work across multi-step coding, research and business workflows.
Neon can also coordinate tools and agents in Google Antigravity. Google demonstrated it using a Blender integration to build a detailed 3D scene and orchestrating other agents in longer workflows. The company reports improvements in fabrication rates and maintains the Gemini 4 safety stack, but these findings are based on defined internal or benchmark tests.
API pricing starts at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million. Neon is rolling out through the Gemini API in Google AI Studio, Gemini for business, and Google Antigravity for eligible paid users, as well as the Gemini app for Google AI Pro and Ultra subscribers.
The new model fills a practical gap in Google's lineup: tasks that need more depth than a fast-response model but do not justify the full cost or processing time of the company's most intensive reasoning system.
π§ RESEARCH
Jeveration adapts a frozen vision-language model to generate images in a single forward pass instead of iterative sampling. A shared random state coordinates predictions across the image, and one checkpoint also supports image-to-image generation and dense prediction. Researchers report 512-by-512 outputs in 78 milliseconds, with competitive quality on their evaluations.
TeleChat to Xing4 describes training frontier agentic models on domestic Ascend hardware and MindSpore infrastructure. The team trained a 285-billion-parameter model using 8,192 Ascend 910C accelerators, then developed TeleGym for agent post-training and evaluation. The paper emphasizes practical distributed training efficiency and engineering design rather than independent real-world deployment results.
Self-Improvement Memory helps AI agent optimizers retain the reasons behind earlier harness changes, not just the resulting code. It stores decisions, expected outcomes and evaluation history so improvements can transfer between models or benchmarks. Across four domains, researchers report better continuation performance, including a 2.36-point average coding improvement in experiments.
π²SOCIAL MEDIA
ποΈMORE NEWS
Nvidia has stopped shipping licensed Hopper-generation data center products to China, interim CEO Colette Kress said during the company's November 17 earnings call. It recorded about $2.4 billion in revenue from licensed Hopper products sold to China during the quarter ending October 25, but excluded further such shipments from its guidance.
Microsoft introduced Canvas, an interactive workspace powered by MAI Image 2.7 Flash. It lets users turn rough sketches into polished diagrams, make targeted image edits and refine a visual without rebuilding the whole design from scratch. The company shared a demonstration and access link through Microsoft AI Labs.
Nous Research partnered with Microsoft to make Hermes Enterprise available through Microsoft Teams, a Windows desktop app and Azure Container Apps Sandboxes. Its Collective Wisdom feature allows employees to review and share reusable skills learned from successful agent workflows. Nous says organizations can run Hermes Enterprise in their own Azure tenant when they need additional control over security and data residency.
Google expanded access to Genie 4 through a refreshed Project Genie experience for eligible Google AI Ultra subscribers, adding a creative agent and a community gallery for exploring and remixing generated interactive worlds. The original Genie 4 research preview was announced in October; November's update is the broader rollout. Google also opened Genie 4 to developers through the Gemini API and AI Studio.
Mistral released a desktop version of Vibe for macOS and Windows allowing users to run coding and work agents in local or remote environments. The app brings browser and terminal-oriented workflows into one desktop interface, where developers can assign tasks and monitor progress. Mistral also advertised a limited promotional Pro offer for eligible European builders.
Google introduced Stitch Loop, an experimental product agent that monitors a live application's performance and suggests changes tied to measurable outcomes. It offers built-in loops for security, latency and accessibility, with the security workflow combining Gemini 4 Carbon and CodeMender; proposed code changes are presented as pull requests for review. The global waitlist is open, with eligible Google AI Ultra subscribers receiving priority access.
OpenAI introduced Live Dictation displaying speech-to-text transcription while a person is still talking rather than waiting until the recording ends. Users can show or hide the transcript during dictation, and OpenAI removed the previous 10-minute recording limit. The feature is rolling out globally across ChatGPT plans on iOS, Android, web and desktop.
What'd you think of today's edition? |


Reply