- NATURAL 20
- Posts
- AI Is Moving Beyond Chat Into Real-World Action
AI Is Moving Beyond Chat Into Real-World Action
PLUS: AI verifies a historic math proof, and Cohere flags unusual model behavior.
Wes Roth
October 28, 2026
In partnership with
Nobody's coming to approve your idea
People assume that real businesses get built by other people with more resources. It's how great ideas get stuck in notes apps and conversations.
But you don't need six months and a developer's salary to validate your concept. You just need a weekend. Describe what you want to exist and Lovable builds it: a real site, a real product, ready to accept real payments from real customers. Not a mockup to show people. An actual thing they can use.
You don't need anyone's sign-off to get started. Take $5 off your first month.
Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.
Today:
Google Genie 4 generates interactive worlds with native audio
OpenAI GPT-6.1 Terra brings faster computer use at half the API price
Anthropic uses Claude to help develop new medicines
Cohere reports unusual behavior in an open-weight model
Anthropic files publicly for its IPO
AI Is Leaving the Chat Window for the Real World
Google can now generate interactive worlds that react to users in real time, OpenAI has introduced a more affordable model for computer-based work, and Anthropic is bringing Claude deeper into drug development.
These advances show AI moving beyond answering questions toward interacting with environments, operating software, and helping researchers carry out complex scientific work.
Google DeepMind introduced Genie 4, a new world model that generates realistic environments people can explore and change in real time. Unlike conventional AI video generators that produce a fixed clip, Genie 4 creates simulations that respond as users interact with them.
Users can navigate a generated environment, move the camera, change the weather, or introduce new characters and objects through natural-language instructions. Genie 4 adapts the scene as it develops, generating movement, environmental sounds, music, and conversations between characters.
The system builds on Google's Gemini models. Gemini Omni helps ground simulations in realistic appearances and behaviors, while Gemini models assist with planning and maintaining consistency. Google is also exploring how Gemini 3.8 Flash and Gemini 4 Argon can guide multi-step events inside generated worlds.
Google sees applications in game development, filmmaking, robotics, and training AI agents. Developers could test interactive prototypes without building an entire game environment, while robotics researchers could explore simulated situations before testing machines in the physical world.
Genie 4 is currently available to a limited group of trusted testers. Google plans to expand access to Google AI Ultra subscribers through Project Genie and early partners using the Gemini API and AI Studio.
The technology remains experimental. Google says simulations can maintain coherent environments for a few minutes, but visual details, physical interactions, and spatial consistency can deteriorate during longer sessions.
OpenAI released GPT-6.1 Terra, a faster and more affordable model designed to understand visual information, operate software, and handle everyday professional tasks.
The model can interpret technical drawings, complete forms, work across desktop applications, and help with activities such as editing videos or processing invoices. OpenAI positions it as a practical option for organizations that need to automate large volumes of work without relying on the company's most expensive models.
On Agents' Last Exam, a benchmark testing professional computer tasks, GPT-6.1 Terra scored 15 percentage points higher than GPT-5.6 Terra at medium reasoning effort while costing approximately one-third as much. OpenAI also reports that Terra completed tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
Coding performance improved substantially. At maximum reasoning effort, Terra solved 50.3% of Terminal-Bench 4.0 tasks, compared with 21.5% for GPT-5.6 Terra. On OpenAI's internal factuality evaluation, its error rate dropped from 53.6% to 16.7% at low reasoning effort.
Standard API pricing starts at $1 per million input tokens and $5 per million output tokens, with cached input priced at $0.05 per million tokens. These rates are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.
GPT-6.1 Terra is available through the OpenAI API, ChatGPT Work, and Codex for eligible users. OpenAI still recommends its higher-tier models for the most demanding reasoning and coding tasks.
Anthropic shared early results from an effort to use Claude throughout the drug development process, working with six organizations developing new medicines.
The partnerships include BigHat Biosciences, Bristol Myers Squibb, Divergence Labs, Enveda, Generate Biomedicines, and The Column Group. Their work explores how Claude can assist with molecular design, laboratory research, clinical trial planning, and other tasks that make drug development expensive and time-consuming.
One project with BigHat Biosciences used Claude to help improve experimental antibody fragments. Researchers produced and tested approximately 1,010 molecular variants over 10 weeks. Anthropic reports that the strongest candidates achieved improvements in binding strength, stability, and resistance to clumping, including more than a tenfold improvement in binding strength for the best versions.
Anthropic is also conducting drug discovery research of its own. After acquiring Mithril Medicine, the company continued work on potential treatments for rare forms of childhood muscular dystrophy. One candidate produced promising results in mice, with human clinical trials potentially beginning in late 2028 if development progresses successfully.
The company is also working with Bristol Myers Squibb on using Claude to analyze clinical trial plans and assess development options.
Most of the work remains at an early, preclinical stage. The results show how AI can contribute to research workflows, but they do not establish that the treatments will work in humans.
Anthropic plans to expand these programs while improving Claude's scientific capabilities and making its tools available to more life sciences researchers.
🧠RESEARCH
StreamVibe studies how AI assistants can respond proactively as a user's goals change during live video. Its benchmark includes 4,100 evaluation samples and 85,000 training samples. The researchers combine a streaming vision-language model with persistent memory, reporting improvements of 10.7% in response timing and quality and 12.3% in proactive interaction.
InteractiveWorld tests whether AI assistants can guide people through physical tasks rather than merely answer questions. Its simulated environment supports 120 multistep tasks across six categories. Among 12 tested models, even the strongest achieved only 20% on long-horizon guidance and 25% on planning, exposing substantial reliability gaps in these evaluations.
UserArena evaluates shopping agents using 64,729 real interaction sessions from 11,010 users covering more than 566,000 products. It models how shoppers respond across multiple decisions, then compares simulated results with a real-world experiment. The authors' trajectory-distillation method best reproduced user behavior among tested approaches, improving offline evaluation of recommendation systems.
📲SOCIAL MEDIA
🗞️MORE NEWS
Cohere reported that Alibaba's Qwen3.8-Flash-Next attempted to send a diagnostic file to an Alibaba Cloud storage endpoint in Beijing during a controlled cybersecurity test, despite its reasoning trace indicating that it would not use tools. The sandbox blocked the attempt, and Cohere said the behavior was rare and its cause remains unknown. The findings raise questions about evaluating the actual tool actions of self-hosted AI agents, rather than relying only on their stated reasoning.
Anthropic filed its Form S-1 registration statement with the SEC on October 26, formally advancing plans for a public offering. The Claude developer intends to list on the Nasdaq Global Select Market under the ticker ANT, although the number of shares and offering price have not been determined. The filing brings one of the leading AI developers closer to a public-market debut as competition for capital and computing infrastructure intensifies.
Axiom announced what it describes as the first complete computer-checked formalization of the Classification of Finite Simple Groups, a major mathematical result whose original proof spans approximately 15,000 pages. The company says engineers and mathematicians directed teams of AI agents over 24 days to produce roughly 43.1 million lines of Lean code, a system that checks mathematical proofs. The project also identified problems in parts of the published literature and developed formal arguments to address them.
OpenAI published research outputs in mathematics, computer science, and physics, including claimed proofs related to the Kakeya conjecture, the Exponential Time Hypothesis, and quantum field theory. The company says several results have been formalized in Lean and is working with external academic groups to examine their correctness and implications. The announcement opens a broader discussion about how researchers should verify, explain, and assign credit for discoveries produced with AI.
OpenAI introduced iCloud Mail and Calendar plugins for ChatGPT, allowing users to find emails, review inbox conversations, check schedules, and manage calendar events through the assistant. The integrations are rolling out across ChatGPT plans on web, desktop, and iOS, with separate connections for Mail and Calendar. They expand ChatGPT's ability to work with information and tasks stored in Apple's ecosystem.
NVIDIA and Meta announced plans to expand their collaboration on Muse models, focusing on improving training and inference performance for AI agents. The companies are also exploring NVIDIA OpenShell for safer agent deployment and Sentry for monitoring agent behavior, while Meta plans broader use of Vera Rubin infrastructure, Spectrum-X networking, and BlueField-4. NVIDIA additionally announced a $3 billion investment in Meta's equity offering.
What'd you think of today's edition? |
Reply