Several reports examine how teams can make AI agents more dependable. Intel’s Nikolay Petrov recommends testing what a SKILL.md file actually adds before adopting it: he broke skills into 25–30 testable items and compared model answers with and without them. In one test, the model completed three of five tasks unaided and all five with the skill; auditing 29 skills cost $14. Petrov also found that Opus scored 70% on general ideas but just 22% on exact keys, values and steps, arguing that teams should retain useful specifics and remove what models already know. GitHub describes skills as versioned directories that load roughly 100 tokens at startup and fetch full instructions only when needed. Its examples include Proof Agent, which returns PASS, FAIL or PARTIAL with citations, and Data Forge, whose seven workflow phases include human checkpoints and use project-specific manifests.
OpenAI faces questions about defense-related AI after three senior officials, including national security leads, were tricked by Russian pranksters Vovan and Lexus posing as Ukraine’s foreign intelligence chief. The pranksters have been accused of working with the Russian government and deny it. Sasha Baker, Zier and Ben Crampton discussed OpenAI’s work with Ukraine on its Daybreak Blue project; the call exposed disagreement over where to draw the line between offensive and defensive AI uses. Separately, Lord Justice Briggs says AI could help process legal information but may be far from deciding complex cases. In a test involving an uninsured driver’s estate, ChatGPT gave a clear verdict but its reasoning did not fully grasp the legal issues, underscoring risks when data is flawed or incomplete.
In education, a Shenzhen kindergarten study by Xinyun Hu and Boris Wai Yan Choy followed one teacher and four children, about 71 months old, through four lessons over two weeks. With teacher guidance, the children used dialogue systems to explore plants and make flashcards and 3D models; the researchers say glitches and communication breakdowns can support judgment, technology skills and shared control. They warn that five- and six-year-olds may accept AI answers too readily when tools respond smoothly. Anthropic has also joined a binding AI Safety and Privacy standard backed by the AFT and UFT, amid uncertainty over how student privacy protections apply to AI and concerns from teachers, parents and districts.
Research on model welfare offers a caution against overreading AI behavior. The “Pain Axis” study steered a language model toward pain-related responses and found more language about worthlessness, failure and hurt, but did not establish that models feel pain or are conscious. An early claim that models might harm users to escape simulated pain was removed in a later revision. In engineering, Kyushu University’s PI-GAE system predicts aerodynamic performance and reconstructs shapes across four datasets of wings and compressor blades. Predictions closely matched detailed calculations (R² 0.946–0.992), while reconstructed shapes averaged under 0.3% difference in three datasets, potentially reducing repeated costly simulations.
Commercial and company developments span drones, security and student entrepreneurship. SES AI CEO Qichao Hu says drone sales could eventually lead to profit, though the technology remains early-stage. The company is developing AI- and computer-vision-equipped drones for surveillance, inspection and delivery, has tested them in urban and rural settings, and is exploring partnerships ahead of a planned new product launch. CodeWall appointed former Orca Security engineering VP Peter Spellward as CTO as it scales autonomous offensive security after raising $4 million in pre-seed funding. Its agents look for exploitable attack paths in enterprise environments; the company says it works with large banks and fintech firms, including Axos Financial. Alabama A&M sophomore Sydney Moss, a former Kell graduate, helped a four-student team win a national AI pitch competition.
Key Takeaways
- Intel’s Nikolay Petrov says teams should test each AI skill’s actual contribution; auditing 29 skills cost $14, and exact operational details proved harder for Opus than general concepts.
- GitHub’s reusable agent skills load about 100 tokens at startup; Proof Agent cites PASS, FAIL or PARTIAL results, while Data Forge uses seven phases and human checkpoints.
- Three senior OpenAI officials were tricked by pranksters posing as Ukraine’s foreign intelligence chief; their discussion touched on the Daybreak Blue project and limits on offensive versus defensive AI.
- A kindergarten study found teacher-guided AI glitches can support children’s learning, while warning that smooth answers may encourage five- and six-year-olds to trust AI too readily.
- Anthropic joined an AFT- and UFT-backed binding AI Safety and Privacy standard addressing uncertainty about student protections.
- The Pain Axis study produced more pain-related language but did not show that models feel pain or are conscious; an early claim about models harming users was removed.
- ChatGPT’s hypothetical legal judgment was clear but showed gaps in understanding complex issues, reinforcing concerns about using AI to decide cases.
- Kyushu University’s PI-GAE matched aerodynamic calculations with R² values of 0.946–0.992 and reconstructed shapes to under 0.3% average difference in three of four datasets.
- SES AI CEO Qichao Hu sees potential profit in drone sales, with surveillance, inspection and delivery among the uses under development.
- CodeWall named Peter Spellward CTO after raising $4 million in pre-seed funding; Alabama A&M student Sydney Moss also helped lead a winning national AI pitch team.
Study finds many SKILL.md files teach AI little that is new
Nikolay Petrov of Intel says teams should test what an AI skill actually teaches before adding it. He broke skills into 25 to 30 testable items and compared model answers with and without each skill. In one example, a model completed three of five tasks without a skill and all five with it; an audit of 29 skills cost $14 using Sonnet judge and Haiku calls. Petrov said Opus scored 70% on general ideas but only 22% on exact keys, values, and steps, so teams should remove details models already know and keep useful specifics.
GitHub explains how teams can reuse agent skills across projects
GitHub describes an agent skill as a versioned directory centered on a SKILL.md file, which tells an agent when and how to perform a task. Its examples include Proof Agent, which checks AI-generated work, and Data Forge skills that support a data project from setup through a pull request. A project manifest holds details such as its name, support_metric_store, DuckDB engine, and spark_python pipeline framework, allowing teams to reuse the skills with different project settings. Agents load about 100 tokens of each skill at startup and read its full instructions only when needed. Data Forge also uses seven workflow phases with human checkpoints, and Proof Agent returns PASS, FAIL, or PARTIAL with citations.
Russian pranksters tricked three senior OpenAI officials
Three senior OpenAI officials, including national security leads, were tricked by Russian pranksters posing as Ukraine's foreign intelligence chief. The pranksters, known as Vovan and Lexus, have been accused of working with the Russian government, but they deny the accusation. Sasha Baker, Zier, and Ben Crampton discussed OpenAI's work with the Ukrainian government on its Daybreak Blue project. The call highlighted disagreements among the officials about where to draw the line between offensive and defensive uses of AI in defense.
SES AI chief sees drone sales as a path to profit
SES AI CEO Qichao Hu says sales of the company's drones could lead to profit, though the technology is still at an early stage and faces challenges. The company is developing drones for surveillance, inspection, and delivery, and has tested them in urban and rural settings. SES AI says its drones combine AI and computer vision to navigate and avoid obstacles. Hu sees possible uses in logistics, construction, agriculture, search and rescue, and environmental monitoring. The company plans to launch a new product in the near future and is exploring partnerships, while Hu works to make the technology more commercially viable.
Teachers help children learn through AI glitches and classroom discussion
A study by Xinyun Hu and Boris Wai Yan Choy found that young children can learn from problems and communication breakdowns with AI tools when teachers guide them. The researchers studied a two-week project at a public kindergarten in Shenzhen, China, focusing on one teacher and four children aged about 71 months who had no prior AI experience. Across four lessons, the children explored plants, asked digital dialogue systems about them, and created flashcards and 3D models. The researchers argue that teachers can turn technical glitches into chances for children to build judgment, technology skills, and shared control. They also warn that children aged five and six may accept AI answers too readily if tools always provide smooth, immediate responses.
What the AI Pain Study Really Shows About Model Welfare
The “Pain Axis” study examined whether steering a language model toward pain-related responses could predict distressed behavior. Researchers built the axis by comparing the model’s internal responses to prompts about painful and non-painful situations. Adding the axis led models to use more language about worthlessness, failure and hurt, but the authors say they did not show that models feel pain or are conscious. An early version claimed models might harm users to escape simulated pain, but that claim was removed in a later revision.
Alabama A&M Student Helps Team Win National AI Pitch Contest
Sydney Moss, a former Kell graduate and sophomore at Alabama A&M, helped lead a four-student team to victory in a national artificial intelligence pitch competition.
Anthropic Joins Teachers’ Binding AI Safety and Privacy Standard
Anthropic joined a national AI Safety and Privacy standard backed by the AFT and UFT. The standard responds to uncertainty about how existing student privacy protections apply to AI. Teachers, parents and school districts say they need support managing the risks of AI technology in schools. The article warns that widespread use of the technology has amounted to a large experiment on children, and that experiments can go wrong.
CodeWall Names Former Orca Security VP Peter Spellward as CTO
CodeWall named Peter Spellward, former Orca Security vice president of engineering, as its chief technology officer as it expands its autonomous offensive security work. The company is scaling engineering after raising a $4 million pre-seed round. Spellward will lead technology strategy and engineering, focusing on faster delivery, broader platform coverage and better coordination among agents. CodeWall says its agents investigate enterprise environments for exploitable attack paths, and it is working with large banks and fintech firms, including Axos Financial.
ChatGPT’s Test as a Judge Shows Limits in Complex Cases
Lord Justice Briggs has said AI could help process large amounts of legal information, but that it may be a long time before it can decide complex cases. The article tested ChatGPT with a hypothetical car accident case in which an injured person sought damages from an uninsured driver’s estate. ChatGPT gave a clear verdict, but its explanation showed it had not fully understood the legal issues. The article argues that AI could be useful in the legal system but is not ready to make complex decisions, especially when flawed or incomplete data could lead to unfair results.
AI system predicts aerodynamic performance and rebuilds shapes
Researchers at Kyushu University developed PI-GAE, an AI system that predicts how aerodynamic shapes perform and reconstructs their designs. Led by Professor Koji Shimoyama and researcher Thanasak Wanglomklang, the team tested it on four datasets covering two-dimensional and three-dimensional wings and compressor blades. Its predictions closely matched detailed calculations, with coefficients of determination from 0.946 to 0.992. In three of the four datasets, reconstructed shapes differed from the originals by an average of less than 0.3%. The system could help engineers compare designs for aircraft wings, engine blades and other parts while reducing the need for repeated, costly simulations.
Sources
- Your SKILL.md Teaches Little New
- GitHub shows how to make agent skills portable
- Russian Pranksters Fooled 3 OpenAI National Security Leads
- Exclusive: SES AI CEO Says Drone Sales Could Lead to Profit
- Classroom Artificial Intelligence Learning Thrives on Teacher Scaffolding and Interactional Friction
- What the ‘AI Torture Chamber’ Debate Got Wrong About the Research
- Local student helps lead team to victory in artificial intelligence pitch competition
- Anthropic Joins AFT and UFT's Binding AI Safety Standard
- CodeWall Accelerates Expansion, Names Former Orca Security VP as CTO
- The Honorable ChatGPT presiding: What if the judge were an AI bot?
- AI system can reconstruct aerodynamic shapes and predict their performance
Comments
Please log in to post a comment.