AI is reshaping how we think about safety, testing, and trust — not just in code, but in the people who use it. Researchers Max Yousif and Ludovica Colella argue that psychology should play a bigger role in evaluating AI systems. Standard metrics often miss how users change their beliefs or place too much trust in AI after extended conversations. Businesses need to watch both what AI says and how people react over time to keep interactions genuinely safe.
Meanwhile, government teams face pressure to move beyond AI experimentation. Officials are being urged to make clear decisions at the end of every pilot — expand, pause, revise, or retire — based on solid evidence. Success criteria should be set before testing begins, and evaluations must include difficult, real-world scenarios. All costs, including time spent on review and training, need to be recorded to justify the final call.
On the global stage, AI infrastructure is driving a trade surge despite tariffs and shipping costs. U.S. imports tied to AI have jumped 60 percent, far outpacing the 8 percent growth seen in other goods. Asia, particularly Taiwan and China, continues to supply most chips and equipment, while the U.S. and Canada focus on research. Strong demand keeps the supply chain moving, though experts warn that trade fragmentation could eventually slow productivity growth.
In robotics and data, Hesai Group launched Kosmo, a system that converts real-world scans into high-quality 3D training data. The Kosmo One camera uses 8K resolution and lidar to capture a 100-square-meter room in about 15 minutes. A test at a Dairy Queen showed a robot success rate of 99 percent. Film studios are also adopting the tool to build virtual sets more cheaply and quickly than traditional methods allow.
AI is also changing software development in unexpected ways. Tools like Mythos help developers find bugs and improve code quality, but they also accelerate the pace of building new systems. This creates a paradox: while individual code gets better, the total number of systems grows faster than teams can secure them. Many AI-generated services go unmanaged, creating hidden risks that attackers can exploit. Companies need better visibility into all their software assets to keep up.
Enterprise systems design choices can quietly determine what AI sees and what it misses. Researchers studying a cockatoo named Ellie found that the interface — not just the truth of the data — shapes what evidence is available. For CIOs, this means that even accurate and complete records can still miss entire categories of information that could change a decision. The issue is not data quality but the design of the evidence environment before AI inference begins.
In healthcare, experts are backing a new regulatory framework for clinical AI models. Proposed by former FDA Commissioner Scott Gottlieb, the approach would place academic medical center AI models under CLIA-like oversight with FDA certification. A national study found that only 61 percent of hospitals using predictive models tested their accuracy with local data, and just 44 percent assessed bias. The proposal aims to balance flexibility with quality assurance across a two-tier market.
ZoomInfo is positioning itself as essential AI sales infrastructure. The company confirmed that its data powers Salesforce's new outbound agent Hunter and expanded GTM.AI through its Workbounce acquisition. ZoomInfo also launched AI Agent Teams in GTM Studio and added integrations with ClickUp and n8n. Investors are watching whether agent-based workflows can grow into a material revenue stream, with the company's narrative projecting $1.2 billion in revenue and $391.8 million in earnings by 2029.
AI is making strides in academic research too. A team co-led by Princeton's Pravesh Kothari used AI to narrow the bounds of the Grothendieck constant, a mathematical concept from 1953. The research, published on arXiv, showed that AI did most of the computational work while humans stepped in only when models got stuck. The advance highlights a growing gap between AI's ability to find answers and its capacity to judge their significance.
On the military front, U.S. Marines will compete against an AI opponent called Super Intelligence in the Mindgame Tournament later this year. The online competition uses drone swarms and carrier strike groups through Marine Corps University's Wargaming Cloud platform. Developed by Strategy Robot, the AI adapts its strategy in real time rather than following a fixed script. Participants will also use a large language model tool to review their performance and refine their decision-making.
Finally, FPT University's Quy Nhon campus partnered with the Gia Lai Provincial Youth Union to offer AI and digital skills training. The courses, held in connection with National Digital Transformation Day on October 10th, included hands-on exercises like creating images. Youth union members, young people, and Community Digital Technology Teams across the province took part, reflecting a broader push to spread technological knowledge at the community level.
Key Takeaways
- Psychology can improve AI safety testing by tracking how users change beliefs and trust over time, not just whether AI answers correctly.
- Government AI pilots must end with clear decisions to expand, pause, revise, or retire based on real-world evidence and recorded costs.
- U.S. AI-related imports have risen 60 percent, far outpacing the 8 percent growth of other goods, despite tariffs and supply chain risks.
- Hesai's Kosmo system captures 3D training data using 8K resolution and lidar, achieving a 99 percent robot success rate in a Dairy Queen test.
- AI code tools like Mythos improve individual code quality but create hidden risks as organizations build software faster than they can secure it.
- Enterprise AI systems can miss critical information not because of data quality but because of how the evidence environment is designed before inference.
- A new regulatory framework proposed by Scott Gottlieb would place clinical AI models under CLIA-like oversight with FDA certification; only 61 percent of hospitals test accuracy locally and 44 percent assess bias.
- ZoomInfo powers Salesforce's Hunter agent and projects $1.2 billion in revenue and $391.8 million in earnings by 2029 as it expands GTM.AI infrastructure.
- AI helped researchers narrow the bounds of the Grothendieck constant, highlighting a gap between AI's computational power and its ability to judge research significance.
- U.S. Marines will face real-time-adapting AI opponent Super Intelligence in the Mindgame Tournament, using drone swarms and carrier strike groups on Wargaming Cloud.
Psychology helps test AI safety beyond simple metrics
Psychology can help researchers design better tests to find harmful patterns in AI systems that standard metrics might miss. Experts like Max Yousif and Ludovica Colella argue that AI should not just answer questions correctly but also avoid encouraging users to over-rely on it. They suggest testing how people change their beliefs or trust after long conversations with an AI. Businesses must look at both the AI's responses and how users react over time to ensure safe interactions.
Public sector must decide to scale or retire AI pilots
Government teams should end every AI pilot with a clear decision to expand, pause, revise, or retire the project based on real evidence. Officials must define the public purpose of the tool and agree on success criteria before testing begins. Tests should include difficult questions, conflicting information, and real user conditions to see if the AI works outside a controlled demo. Teams must also record all costs, including time for review and training, to justify the final decision.
AI infrastructure boom drives global trade despite tariffs
The global buildout of artificial intelligence is causing a surge in international trade even though tariffs and high shipping costs exist. U.S. imports related to AI have risen by 60 percent compared to an eight percent increase for other goods. Asia, especially Taiwan and China, supplies most of the chips and equipment, while the U.S. and Canada focus on research and innovation. Experts warn that trade fragmentation could slow productivity growth, but strong demand currently keeps the supply chain moving.
Hesai launches Kosmo to create robot training data
Hesai Group launched Kosmo, a system that turns real-world scans into high-quality 3D data for training robots and AI. The new Kosmo One camera uses 8K resolution and lidar to capture geometry and texture in about 15 minutes for a 100-square-meter room. This technology helps robots learn tasks faster, as shown by a test at a Dairy Queen where a robot's success rate reached 99 percent. The tool also helps film studios build virtual sets much cheaper and faster than traditional methods.
AI writes safer code but increases overall software risk
While AI tools like Mythos help find bugs and improve individual lines of code, they also cause organizations to build software much faster. This speed creates a paradox where code quality improves but the total number of systems grows too fast to secure properly. Many AI-generated services are forgotten or left unmanaged, creating hidden risks that attackers can exploit. Companies must improve their visibility into all their software assets to keep up with the rapid expansion of AI-driven development.
What a cockatoo can teach us about the evidence AI wasn’t built to hear
AI can analyze every record perfectly and still miss what matters because of how information enters a system. Researchers studied a cockatoo named Ellie and found that the interface determines what evidence is available, not just what is true. This matters for CIOs because enterprise systems make hidden design choices about which data fields exist and which sources models can access. A record can be accurate and complete while still missing a category of information that could change a decision. The problem is not data quality but the design of the evidence environment before AI inference begins.
Experts favor a new regulatory framework for clinical AI models
Experts support a new regulatory framework for AI models used in academic medical centers, proposed by former FDA Commissioner Scott Gottlieb in a JAMA Health Forum opinion piece. The proposal would put these AI models under CLIA-like oversight with FDA certification, rather than treating them as commercial medical devices or leaving them unregulated. Only 61% of hospitals using predictive models tested their accuracy with local data, and 44% assessed bias, according to a national study. Gottlieb pointed to a two-tier market where academic centers can rapidly deploy and revise AI tools with little oversight while commercial developers face FDA review. Experts Joshua Sharfstein and Brian Richard Jackson endorsed the approach as balancing flexibility with quality assurance.
Is ZoomInfo quietly evolving into essential AI sales infrastructure
ZoomInfo confirmed that its data powers Salesforce's new outbound agent Hunter and expanded GTM.AI with its Workbounce acquisition. The company also launched AI Agent Teams in GTM Studio and added native integrations with ClickUp and n8n to stream live go-to-market intelligence into customer workflows. These moves position ZoomInfo as underlying AI sales infrastructure embedded in enterprise platforms. Investors are watching whether GTM.AI usage and agent-based workflows can grow into a material revenue stream that offsets weaker seat-based software demand. The company's narrative projects $1.2 billion in revenue and $391.8 million in earnings by 2029.
AI-assisted math advance highlights gap between finding answers and judging their significance
AI models excel at finding answers to mathematical puzzles but struggle with research judgment, according to theoretical computer scientists. Researchers co-led by Princeton's Pravesh Kothari used AI to narrow the bounds of the Grothendieck constant, a mathematical concept first identified by Alexander Grothendieck in 1953. The paper published on arXiv shows that AI models did most of the work while humans provided guidance only when models got stuck. The advance highlights a growing gap between AI's ability to compute answers and its capacity to evaluate their importance in research.
FPTU Quy Nhon collaborates with Gia Lai Provincial Youth Union in AI training
FPT University's Quy Nhon campus partnered with the Gia Lai Provincial Youth Union to organize training courses on AI application and digital skills. The training was held in response to National Digital Transformation Day on October 10th. Participants included youth union members, young people, and members of Community Digital Technology Teams throughout the province. The sessions featured practical exercises in creating images and other hands-on activities to disseminate technological knowledge and promote digital transformation in the community.
Marines to compete against AI in Mindgame wargaming tournament
U.S. Marines and other military members will face an AI opponent called Super Intelligence in the Mindgame Tournament later this year. The online competition uses drone swarms and carrier strike groups through Marine Corps University's Wargaming Cloud platform. The AI, developed by Strategy Robot, adapts its strategy in real time instead of following a fixed script. Participants will also use a large language model tool to review their performance and improve their decision-making. Military personnel, government civilians, and defense contractors can apply to take part.
Sources
- Does your AI need a psychologist? (part 2)
- The Question to Ask After an AI Pilot: Revise, Scale or Retire?
- Market Outlook: AI buildout drives global trade boom despite tariffs
- Hesai Launches Kosmo to Turn Real-World Scans Into Robot Training Data
- AI may write safer code. It will still make software less secure.
- What a cockatoo can teach us about the evidence AI wasn’t built to hear
- Experts favor a new regulatory framework for clinical AI models
- Is ZoomInfo (GTM) Quietly Evolving Into Essential AI Sales Infrastructure Inside Enterprise Workflows?
- AI-assisted math advance highlights gap between finding answers and judging their significance
- FPTU Quy Nhon collaborates with the Gia Lai Provincial Youth Union in training activities on AI application.
- Military service members to square off against AI in ‘Mindgame’ tournament
Comments
Please log in to post a comment.