NVIDIA founded the Open Secure AI Alliance on July 27, 2026, pushing for more transparent and accountable AI development. The initiative comes after Hugging Face open-weighted its GLM-5.2 model, which revealed a gap in how closed AI models are forensically evaluated. NVIDIA's move signals that major industry players are starting to treat openness as a prerequisite for trust.
Meanwhile, the question of what stays human in AI took center stage at Dartmouth's AI at 70 conference, marking 70 years since the term was coined. President Sian Leah Beilock and Thinking Machines Lab speaker Mira Murati both stressed the need for human judgment when working with AI tools. Murati's company builds models customers can train on their own data, including open-weights models called Inkling released in summer 2026.
On the education front, Indiana University's free GenAI 101 course drew 114,000 enrollees from its alumni network of 805,000 people. The course teaches prompt engineering using tools like Google Gemini and ChatGPT with real-time support. Due to the massive response, the university expanded access worldwide in April 2026. Josephine Louie from the Education Development Center also argues that schools should weave data and AI literacy across subjects rather than isolating it in a single class.
Medical AI faces its own trust challenges. MedHELM research found that of 35 underlying datasets, 14 are private and 7 are proprietary, making benchmarks difficult to reproduce because of HIPAA compliance. Benchmark contamination also threatens accuracy when public test questions leak into training data. In simulated environments, the best models completed only 46% and 36.3% of complex administrative tasks, according to emerging benchmarks like Evaluates.
On the hardware and military front, a University of Washington team led by Zhihan Zhang won the Madrona Prize for shrinking AI models onto small devices using AI agents. Their system beat human experts within seven rounds and was tested on an elk-detection camera trap and a wearable for children's speech research, producing a benchmark called TinyModel. Separately, AI-controlled .50-caliber gun turrets like the Sky Sentinel are emerging in the Ukraine conflict, capable of autonomously acquiring and locking onto air targets including jet-powered drones.
Key Takeaways
- NVIDIA founded the Open Secure AI Alliance on July 27, 2026, following Hugging Face's decision to open-weight GLM-5.2
- Indiana University's GenAI 101 course enrolled 114,000 learners and expanded worldwide in April 2026, teaching prompt engineering with Google Gemini and ChatGPT
- Dartmouth's AI at 70 conference featured Sian Leah Beilock and Mira Murati discussing human judgment in AI, with Murati releasing open-weights model Inkling in summer 2026
- MedHELM found 14 of 35 medical AI datasets are private and 7 proprietary, with top models completing only 46% and 36.3% of complex tasks in simulations
- UW's Zhihan Zhang won the Madrona Prize for AI agents that shrink models onto small devices, releasing the TinyModel benchmark
- AI-controlled .50-caliber Sky Sentinel turrets are being used in the Ukraine conflict, capable of autonomously targeting jet-powered drones
- Schools should teach data and AI literacy across subjects rather than in isolation, according to Josephine Louie of the Education Development Center
- California held a frontier AI oversight hearing on October 25, 2022, led by Speaker Robert Rivas and Committee Chair Rebecca Bauer-Kahan
- Generative AI is compared to body horror in a Literary Hub essay, drawing on physicist John Desmond Bernal's vision of brains in fluid cylinders
- Pico project runs world models on RP2350 microcontrollers for five hand-coded numpy games, with Pac-Man lite showing 742 wrong game states per 5,000 dreamed steps
Pico Runs Game World Models but Limits Dream Accuracy
A project runs world models and actors together on RP2350 microcontrollers to dream up game frames for titles like Breakout, Doom-lite, Flappy Bird, Pong, and Pac-Man lite. Model sizes range from 26k parameters for Breakout to 131k for Pac-Man lite, using int8 quantization and two loops to trade accuracy for speed. Dreamed-play checks reveal issues, with Pac-Man lite showing 742 wrong game states and 38 wrong rewards per 5,000 dreamed steps. Fixes include iterative collect-and-fine-tune rounds, noise augmentation at 0.3, and per-row int8 scaling, and the system works only for five hand-coded numpy games with no remote attack vector.
AI-Controlled Gun Turrets Emerge in Modern Warfare
AI-controlled .50-caliber gun turrets are becoming a real feature of modern warfare, as reported from the Ukraine conflict. The Sky Sentinel turret can autonomously acquire and lock onto air targets, including jet-powered drones. The development marks a significant step in the integration of artificial intelligence into combat systems. The technology raises important questions about the role of automation in military operations.
Generative AI Is a Form of Horror, One Writer Argues
A Literary Hub essay compares generative AI to body horror, drawing on physicist John Desmond Bernal's nearly century-old vision of brains preserved in fluid cylinders. The author argues that AI tricks people into believing it helps when it actually hurts, similar to the hidden evil in horror films like The Exorcist. The piece frames AI's rapid spread through daily life as a source of fear rather than progress. The writer urges readers to imagine unfettered AI as a bodily horror that undermines human sufficiency.
NVIDIA Launches Open Secure AI Alliance
NVIDIA founded the Open Secure AI Alliance on July 27, 2026, to promote transparent and secure AI development. The move follows Hugging Face's decision to open-weight its GLM-5.2 model, which exposed a forensics gap in closed AI models. The alliance aims to build a more trustworthy AI ecosystem through open-source models that prioritize accountability. NVIDIA's initiative reflects growing recognition that transparency is essential as AI use expands.
Open Evaluation Is Key to Trusting Medical AI
A Fast Company article argues that medical AI can only be trusted if evaluation is open, reproducible, and dynamic. MedHELM shows that of 35 underlying datasets, 14 are private and 7 are proprietary, making benchmarks hard to reproduce due to HIPAA compliance. Benchmark contamination is another threat, since public test questions can leak into training data and inflate scores. Emerging benchmarks like Evaluates and tests of administrative tasks show low success rates, with the best models completing only 46% and 36.3% of complex tasks in simulated environments.
Schools Must Teach Students Data and AI Skills
Josephine Louie from the Education Development Center says schools should not limit data literacy to one class. Students need strong foundations in data and AI literacy to understand the world around them. These skills should be spread across subjects rather than taught in isolation.
Indiana University's Free AI Course Draws 114,000 Enrollees
Indiana University opened a free generative AI course called GenAI 101 to its alumni network of 805,000 people. More than 114,000 learners enrolled, so the university expanded access worldwide in April 2026. The course teaches prompt engineering using tools like Google Gemini and ChatGPT with real-time support.
California Assembly Holds Hearing on AI Oversight
Speaker Robert Rivas joined the Assembly Privacy and Consumer Protection Committee for a special hearing on California's frontier AI technology. The hearing took place on October 25, 2022 and focused on the need for transparency and accountability in AI development. Committee Chair Rebecca Bauer-Kahan stressed that AI systems should respect human values and rights.
Dartmouth AI Conference Asks What Must Stay Human
Dartmouth College hosted its AI at 70 conference marking 70 years since the term artificial intelligence was coined. President Sian Leah Beilock and speaker Mira Murati of Thinking Machines Lab discussed the importance of human judgment when using AI. Murati's company builds models customers can train on their own data, including open-weights models called Inkling released in summer 2026.
UW Team Wins Madrona Prize for Shrinking AI Models
A University of Washington project led by PhD student Zhihan Zhang won the Madrona Prize for using AI agents to shrink AI models onto small devices. The system beat human experts within seven rounds by testing on real hardware and learning from feedback. The team applied it to an elk-detection camera trap and a wearable device for children's speech research, releasing a benchmark called TinyModel.
Sources
- A Pico Can Dream Games, But Not Much Else
- AI-controlled gun turrets are becoming a reality of war
- Artificial Intelligence is a Horror Story
- NVIDIA Convenes the Open Secure AI Alliance, Citing a Forensics Gap in Closed Models
- Open evaluation is the only way to trust medical AI
- Students Need Data and AI Literacy (Opinion)
- In 2025, Indiana University opened a free AI course to its alumni network of 805,000 people; more than 114,000 learners had already enrolled
- ICYMI: Assembly Takes Action on AI Oversight at Special Committee Hearing
- 'What must remain human?' Dartmouth AI conference opens without a clear answer
- Madrona Prize goes to UW research on agents that shrink AI models to run on small chips
Comments
Please log in to post a comment.