Google DeepMind's AI readiness chief, Lila Ibrahim, says the benefits of AI will outweigh the risks. She joined DeepMind in 2018 and moved into her new role this year. Ibrahim advises people to start using AI tools themselves to understand them better. She also wants schools to teach students about what AI can and cannot do.
Ibrahim believes people should have a voice in how AI gets used. She uses AI daily to summarize briefing documents and asks her teams three key questions about their AI usage. She wants her teenage daughters and all students to learn responsible AI use in schools.
Meanwhile, OpenAI and Synopsys announced a multi-year partnership to build GPT-Synopsys, a specialized AI model for chip design. The model will operate Synopsys EDA tools directly, helping engineers evaluate more options and deliver more sophisticated chips faster. It will run on OpenAI-hosted infrastructure and integrate with Synopsys.ai and Autopilot. The companies have not disclosed a release date or pricing structure.
A SailPoint study found that 79 percent of organizations now run AI agents in production, but only 2 percent have identity security tools designed to govern them. While human security has improved, 54 percent of organizations sit at the lowest maturity level for agent identity security. The report warns that enterprises must compress years of security maturity into a single year to keep up with the rapid rise of autonomous AI agents.
Researchers auditing agentic LLM pipelines found that label agreement does not measure whether an agent was authorized to propose a label or whether its output was complete. On COBRE and FBIRN neuroimaging datasets, they discovered that output can parse as JSON yet still drop required fields in 10 to 33 percent of cases. A row-order error also erased most diagnostic signal that none of the tested metrics detected.
The EU says its AI Act, agreed in 2024 and enforceable since August, is strong enough to protect Europeans. It has sent over 30 information requests to companies on issues like copyright and safety. However, EU lawmakers and experts warn of gaps, including the lack of AI liability rules and the AI Office having only about 125 staff. They also question whether Europe has the political will to enforce rules against large US-based AI providers.
Jay Shimshack, a professor at UVA's Batten School, says data and AI cannot replace human judgment in policymaking. He stresses that policymakers need deep context understanding before making decisions. Data and AI can offer valuable insights but cannot account for consequences or stakeholder impact on their own.
CoreWeave announced improvements in weight synchronization and data movement to keep GPU utilization high during AI post-training. These changes reduce delays between training rounds and help agents use tools more effectively. Customers like You.com report higher accuracy and lower total cost of ownership.
An article from Ethics.Dev highlights how AI agents and cybersecurity are increasingly colliding. It notes that AI attackers can have unlimited time and resources compared to human attackers with limited capacity, making AI-powered attacks a more formidable threat.
A commentary argues the military must re-engineer operations around AI rather than treating it as a software update. It calls for strong data governance and adoption of AI Fitness Standards, urging leaders to use enterprise platforms like JWCC and Cloud One instead of building siloed systems.
Key Takeaways
- Google DeepMind's Lila Ibrahim says AI benefits will outweigh risks and wants schools to teach students about AI
- OpenAI and Synopsys partner to build GPT-Synopsys, a specialized AI model for chip design
- 79% of organizations run AI agents in production, but only 2% have tools to govern them
- Researchers found label agreement fails to detect missing fields in 10 to 33% of cases in agentic LLM pipelines
- EU's AI Act is enforceable but faces enforcement gaps with only about 125 staff at the AI Office
- UVA professor Jay Shimshack says data and AI cannot replace human judgment in policymaking
- CoreWeave's GPU utilization improvements help AI agents use tools more effectively, lowering costs for customers like You.com
- AI attackers have unlimited time and resources compared to human attackers, changing cybersecurity economics
- The military must adopt AI-first thinking and modernize infrastructure rather than treating AI as a software update
- 54% of organizations are at the lowest maturity level for agent identity security
Google DeepMind's AI readiness chief says benefits will outweigh risks
Lila Ibrahim became Google DeepMind's AI readiness chief earlier this year. She told CNBC that her job is helping workers, schools, and communities prepare for AI. She advises people to start using AI tools themselves and says the benefits will outweigh the risks. She also wants schools to teach students about AI's capabilities and limits.
DeepMind's Lila Ibrahim on preparing for AI's impact on society
Lila Ibrahim joined DeepMind in 2018 and moved from chief operating officer to her new AI readiness role. She believes people should have a voice in how AI is used. She uses AI daily to summarize briefing documents and asks her teams three key questions about their AI usage. She also wants her teenage daughters and all students to learn responsible AI use in schools.
Label Agreement Does Not Measure Authorization
Researchers audited agentic LLM pipelines that handle label ontology and metadata. They found that label agreement does not measure whether an agent was authorized to propose a label or whether its output was complete. On COBRE and FBIRN neuroimaging datasets, they discovered that output can parse as JSON yet still drop required fields in 10 to 33 percent of cases. A row-order error also erased most diagnostic signal that none of the tested metrics detected.
EU vows AI rules will protect Europeans. Will they?
The EU says its AI Act, agreed in 2024 and enforceable since August, is strong enough to protect Europeans. It has sent over 30 information requests to companies on issues like copyright and safety. However, EU lawmakers and experts warn of gaps, including the lack of AI liability rules and the AI Office having only about 125 staff. They also question whether Europe has the political will to enforce rules against large US-based AI providers.
79% of organizations run AI agents; only 2% have tools built to govern them
A SailPoint study found that 79 percent of organizations run AI agents in production, but only 2 percent have identity security tools designed to govern them. While human security has improved, 54 percent of organizations are at the lowest maturity level for agent identity security. The report warns that enterprises must compress years of security maturity into a single year to keep up with the rapid rise of autonomous AI agents.
OpenAI and Synopsys build AI model for chip design
OpenAI and Synopsys announced a multi-year partnership to create GPT-Synopsys, a specialized AI model for designing chips. The model will operate Synopsys EDA tools directly, helping engineers evaluate more options and deliver more sophisticated chips faster. It will run on OpenAI-hosted infrastructure and integrate with Synopsys.ai and Autopilot. The companies did not disclose a release date or pricing structure.
AI is changing the economics of cyberattacks
An article from Ethics.Dev highlights how AI agents and cybersecurity are increasingly colliding. It notes that AI attackers can have unlimited time and resources compared to human attackers with limited capacity. This shift makes AI-powered attacks a more formidable threat. The author argues this less dramatic but critical economic change deserves more attention in security discussions.
Data and AI cannot replace human judgment in policy
Jay Shimshack, a professor at UVA's Batten School, explains that data and AI are not substitutes for human expertise in policymaking. He stresses that policymakers need deep context understanding before making decisions. Data and AI can offer valuable insights but cannot account for consequences or stakeholder impact on their own. Shimshack urges critical thinking and sound judgment based on available information.
Better GPU use improves AI agent training
CoreWeave announced improvements in weight synchronization and data movement to keep GPU utilization high during AI post-training. These changes reduce delays between training rounds and help agents use tools more effectively. Customers like You.com report higher accuracy and lower total cost of ownership. The improvements make advanced training accessible beyond frontier labs.
Military must adopt AI-first thinking
A commentary argues the military must re-engineer operations around AI rather than treating it as a software update. It calls for strong data governance and adoption of AI Fitness Standards. The article urges leaders to use enterprise platforms like JWCC and Cloud One instead of building siloed systems. It says cleaning data and modernizing infrastructure are urgent for national security.
Sources
- Google DeepMind's AI readiness chief says benefits of AI will outweigh the risks
- Google DeepMind's Lila Ibrahim, world's top AI 'readiness' officer, on preparing for what's to come
- Label Agreement Does Not Measure Authorization
- EU vows AI rules will protect Europeans. Will they?
- 79% of organizations run AI agents; only 2% have tools built to govern them.
- OpenAI and Synopsys partner to build "GPT-Synopsys" for autonomous chip design — specialized AI model will operate EDA tools, allowing engineers to deliver more sophisticated chips faster
- The part of the AI security story I keep thinking about
- Q&A: How are data and AI changing the future of policymaking?
- GPU utilization in post-training shapes better AI agents
- Military AI dominance: Breaking ossified logic and re-writing institutional dogma
Comments
Please log in to post a comment.