AI Agents Become Persistent Workspaces
In Brief
- AI agents are becoming persistent workspaces: Perplexity, Meta, and Amazon are moving beyond individual prompts toward continuous, context-aware tasks.
- Infrastructure is running into political limits: Election results and local protests are intensifying pressure on new data centers and their energy needs.
- Capital and revenue tell different stories: OpenAI is receiving a huge new valuation, while stagnant spending and skepticism about Anthropic are raising questions about monetization.
- Robotics is moving closer to real-world operations: New humanoids, simulation tools, and more realistic training environments are targeting use in factories and homes.
🔬 Models & Research
Inkling 2 Small aims to be the strongest Western open-weight model
06.11.2026
Thinking Machines is presenting a Mixture-of-Experts model with 594 billion parameters, 17 billion of which are active at a time, and an index score on par with Gemini-3.8-Flash. Its strong gains on agentic tasks make the model interesting, but its high resource requirements remain a practical consideration.
Ripple-65B Preview targets agentic coding
06.11.2026
Ricursive Intelligence pretrained Ripple-65B Preview on 12 trillion tokens and reports a score of 80.52 on TerminalBench 2.1—the highest among open US models. The locally deployable alternative could make coding agents more accessible, although benchmarks and harnesses also influence performance.
METR sees evidence of accelerated AI research at Anthropic
06.11.2026
After 11 weeks of access to internal research data, METR concluded that the pace of model improvements had roughly doubled, with about 80 percent of the acceleration attributable to the use of AI in research. If confirmed, models’ ability to speed up their own development would become a central safety and governance risk.
Underdog showcases a speech model just 32 MB in size
06.11.2026
Startup Underdog is demonstrating Woof, a speech model that is around 32 MB in size and is used in an interactive gaming environment. Such compact models could enable speech features directly on smartphones and other edge devices, rather than requiring cloud infrastructure for every request.
💼 Companies & Markets
OpenAI reportedly reaches a $1.25 trillion valuation
07.11.2026
OpenAI has reportedly already raised $11 billion in an ongoing funding round, with the total potentially reaching around $20 billion. The valuation underscores the enormous capital demands of the AI race, while the smaller amount compared with the original target also reflects reliance on market sentiment and investors.
Falling valuations put Anthropic’s planned IPO under pressure
06.11.2026
Investors reportedly consider Anthropic’s target valuation of $1.5 trillion too high, which could delay a planned roadshow. The case shows that even highly sought-after AI companies increasingly have to align their growth plans with profitability and public-market conditions.
Companies are using more AI, but spending is flat
06.11.2026
According to Ramp data, AI spending grew by just 1.5 percent in August and September, even as usage continued to rise when measured by token volume. For providers and investors, this signals that falling prices and more efficient customers could constrain revenue more than model usage alone suggests.
AI data centers face pressure over energy needs and grid expansion
06.11.2026
The expansion of energy-intensive data centers is intensifying conflicts over power grids, local costs, and environmental impacts; several tech companies are now promising closer cooperation with communities. Location, energy supply, and public acceptance are thus becoming as important to AI strategies as chips and computing power.
Palo Alto Networks in talks to buy Clutch Security
06.11.2026
Palo Alto Networks is reportedly considering acquiring Clutch, a security startup specializing in AI agent access, for $500 million plus possible milestone payments. The potential deal signals that companies need to treat the identities and permissions of autonomous agents as a distinct part of their security infrastructure.
AI is becoming a dealbreaker for some people on dating apps
06.11.2026
Dating profiles are increasingly expressing clear opposition to AI, particularly among people in creative and social professions. For providers and employers, this suggests that public acceptance of AI depends not only on product performance, but also on personal values and perceived impacts.
⚖️ Politics & Regulation
Report: China allegedly recruited an AI executive to target Nvidia CEO Jensen Huang
07.11.2026
According to a New York Times report, China’s state security allegedly recruited an executive at an Nvidia customer to gather information about Jensen Huang and US AI chips. The allegation underscores how semiconductor expertise and business relationships have become geopolitical security risks.
US midterm elections intensify political headwinds for Big Tech
06.11.2026
According to the report, voters backed stricter restrictions in 40 of 41 votes on data centers, while AI and Big Tech also became national campaign issues. New political majorities could slow infrastructure expansion and impose additional requirements on companies and operators.
Protest against Microsoft data center in Amsterdam escalates
06.11.2026
Around 100 activists entered a construction site for a new Microsoft data center and protested against the planned development on the outskirts of the city. The incident shows that local disputes over land, energy consumption, and the role of tech companies pose concrete project and reputational risks.
🛠️ Tools & Products
Perplexity launches always-on AI agents for complex tasks
06.11.2026
Perplexity’s new Teammates are designed to work on ongoing projects independently, each with a virtual computer equipped with a desktop and browser. Integration with Threads and Slack shifts agents from one-off assignments to persistent workflows, though the report says a Max subscription is required.
Meta Spaces turns Muse conversations into customizable workspaces
06.11.2026
Users can ask Muse to create themed Spaces from a description, such as shopping plans, music, or a dashboard that draws on connected services. This develops the assistant from a chat window into a personalized interface, whose usefulness depends on data access and reliable updates.
Alexa now accepts shopping requests by text message
06.11.2026
Select US customers can ask Alexa to search for products by text message or product photo and place orders directly. The SMS channel lowers the barrier to agentic shopping, while also strengthening Amazon’s control over product discovery and checkout.
Claude gets a personal finance assistant
06.11.2026
In the US beta, users can connect bank and card accounts through Plaid and ask Claude questions about spending, categories, and upcoming payments. Read-only access limits the immediate risk of transactions, but privacy and the secure handling of sensitive financial data remain central.
Akai and Moog open their devices to modders and developers
06.11.2026
Parent company inMusic plans to provide developers with tools, access to device code, and an SDK for modifications. Controlled access to hardware can foster community innovation while reducing the risk that changes render devices unusable.
Hundreds of publicly accessible instances of an AI hacking tool discovered
06.11.2026
Security researchers found hundreds of freely accessible installations of Artex, an open-source agent for automated penetration testing that has been linked to several cyberattacks. Poorly secured agents can make offensive capabilities available to third parties, highlighting why deployment controls matter as much as model safeguards.
🤖 Robotics
Embodied AI unveils AIR, a humanoid factory robot
06.11.2026
Embodied AI developed AIR in nine months; the humanoid robot is designed to work safely alongside factory teams and learn directly on the job. Tests with global manufacturers bring the central question into practical focus: whether learning-capable robotics can handle repetitive tasks reliably without excluding people from the workspace.
ToyBlocks creates a controlled test world for robot learning
06.11.2026
With ToyBlocks, AllenAI is introducing a framework for controlled environments in which training data and model design can be deliberately varied and errors traced. Reproducible tests help robotics teams systematically measure progress and limitations in tasks such as sorting or handling unfamiliar objects.
Physicl raises $10 million for more realistic robot simulations
06.11.2026
Startup Physicl has raised a $10 million seed round to develop realistic simulations for world models and robotics. High-quality digital environments could make training data more scalable and narrow the gap between simulated behavior and real objects.
Humanoid robot cleans apartments, but is still remotely controlled
06.11.2026
A new cleaning service in San Francisco uses humanoid robots that are still controlled remotely by people. The trial shows that the hardware is already being tested in everyday environments, but robust autonomy and economically viable operations have yet to be achieved.