AI & robotics news brief ·
Millions of attacks hit unpatched servers on day one
Good morning. Millions of attacks began hitting unpatched servers on day one, with 1.5 million still exposed. Also today: Anthropic introduced Claude Duet, and a new paper looks at continued pre-training for embodied agents in virtual worlds.
Today in AI
Millions of Attacks, 1.5M Servers Unpatched on Day One
Why it matters: The rapid exploitation of PHP2Shell vulnerabilities poses significant risks to millions of web servers, highlighting the urgent need for widespread security measures and prompt software updates.
Year-old AI lab Flapping Airplanes raises $1B at a $6B valuation
Why it matters: This significant funding underscores the growing interest and investment in AI technologies that promise to enhance efficiency in model training and deployment.
Scaling in Games: Continued Pre-Training For Embodied Agents In Diverse Virtual Worlds
Why it matters: Advancements in pre-training for embodied agents could enhance their performance in varied virtual environments, important for the future development of generalist AI systems.
The Social Society Helping Women Join the AI Boom
Why it matters: The emergence of the Aged Intelligence Society highlights the importance of integrating women into the AI discourse, thereby addressing specific challenges they face, especially in midlife.
Anthropic Introduces Claude Duet: A New Agentic AI App for Professionals
Why it matters: The introduction of Claude Duet could significantly enhance productivity for professionals by automating task management and prioritization within their daily workflows.
Robotics
Uber rival inDrive eyes 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia
Why it matters: The development of robotaxis in Kazakhstan could significantly advance autonomous vehicle technology in challenging environments, influencing future deployments in similar markets worldwide.
NVIDIA Details Kazakhstan GPU Deployments and Robotaxi Plans
Why it matters: The deployment of GPUs and the exploration of robotaxi services in Kazakhstan signify key advancements in the nation's AI capabilities and infrastructure modernization.
Agent uses to try
Research & Learning · Claude Code
Have a coding agent check papers before each experiment
A Claude Code agent with paper search found useful ideas that a plain run missed while tuning a small language model. The version with paper search improved the model by 4.05%, while the baseline reached 3.67%. The poster says roughly 15 of the 25 paper-based ideas proved useful. They also said the stronger setup remained ahead by 12+ minutes in a 2-hour run, but commenters said the result may not generalize well.
Try this prompt
I want to improve [model/project/benchmark]. Before each attempt, search [paper corpus/search tool] for methods related to [problem]. For each candidate idea, tell me the paper it came from, why it might help, and how to compare it with the current baseline of [score]. Apply only one change at a time, pause and ask me before any irreversible action or final submission, and keep a log of each idea, its paper source, and the result. After [number] experiments, summarize which ideas were worth keeping.
Coding & DevOps · Replit Agent
Build a CRM with email, video, and scheduling
A tech sales professional used Replit Agent to make a working CRM with contact, email, video, and calendar features. They described the project as a v0.2, with the main functions in place and more features still planned. People who tried it said the mobile layout was weak, one tester could not create a contact, and there was no login page.
Try this prompt
Help me build a simple CRM for my [business type]. I need a place to add contacts, track notes and deal stage, and schedule meetings on a calendar. First build the core contact and stage flow, then add [email sending / video recording / another feature] after the basics work. Before you do anything irreversible such as sending an email, placing a call, deleting data, or submitting a form, stop and ask me to approve it. Once the app can [key test action], show me how to try it and what to fix next.
Money & Bills · Claude Code
Turn brokerage exports into a portfolio action plan
Use an agent to parse brokerage files, compare holdings to your goals, and draft buy, hold, and stop-buying lists with tax-loss harvesting alternatives. The agent produced a ten-part plan with four stages, plus tax repayment tables and fund research. Matt said he is close to finishing the changes, and his own review put the plan at 7.5 out of 10. One improvement he found was to place the overweight position in tax-free accounts.
Try this prompt
Read the brokerage CSV exports and balance notes in [folder path]. Use my constraints in [constraint list] to compare my holdings against my target setup in [goal description]. Research each fund I own, then give me three lists: add, stop adding, and keep. For any position you think could be sold for a tax loss, suggest a comparable partner fund and explain the match. Save the plan to [output file]. Before you do anything irreversible, such as placing orders or submitting forms, stop and ask me to approve the next step. Also critique your own plan and call out any weak spots or assumptions.