Nvidia Faces Leadership Questions as AMD and Broadcom Narrow the Gap

Key Takeaways

• AlphaGenome 2 raised predictions of whether DNA changes increase or decrease gene activity from 23% to 43% on an immune-cell benchmark; it is for noncommercial research, with commercial access in private preview.
• DeepMind’s model was trained on more than 107 million cell profiles from 68 human and mouse tissues, but reported validation focuses on blood and immune cells and it is not a diagnostic tool.
• Manulife’s Nathan Thooft says uncertainty about Nvidia CEO Jensen Huang raises questions about innovation, while AMD and Broadcom are narrowing the capabilities gap.
• Huang’s potential removal could affect Nvidia and the broader AI industry; the company faces a governance question under conduct rules concerning relationships with employees who report directly to the CEO.
• A UW-Madison course emphasizes responsible AI use, critical thinking and understanding AI’s limits rather than training around particular tools.
• Sales teams using Claude and other general-purpose AI assistants may need training and shared processes; cited research says 76% of sellers face pressure to use AI and 44% lack tools or training.
• The Ontario Securities Commission found AI oversight gaps in its review of 17 firms, including weak training and due diligence and potentially overstated claims about tools.
• EvalStop reportedly cut wasted GPU time by about 22% and shortened training jobs by about 9%, while catching 99% of jobs whose results were worsening.
• A safety analysis urges frontier AI companies to test safeguards adversarially, define limits for agents, and demonstrate intervention and shutdown controls.
• AI Engineer announced plans including a February London event and a San Francisco gathering targeting 10,000 attendees next summer; WMU held its fifth annual cybersecurity summit on Nov. 12, 2026.

AlphaGenome 2 improves DNA variant predictions in specific cells

DeepMind’s AlphaGenome 2 uses single cell data to predict how DNA changes affect different cell types and conditions. On the TenK10K map of 28 immune cell types, it correctly predicted whether gene activity would rise or fall in 43% of cases, compared with 23% for the earlier version. DeepMind trained the model on more than 107 million cell profiles from 68 human and mouse tissues, along with bulk data. Its validation focused on blood and immune cells, and the company has not reported false positive rates or results across other areas such as brain cells. The model is available for noncommercial research through an API, while commercial access is in private preview.

AlphaGenome 2 identifies more DNA changes linked to disease

Alphabet introduced AlphaGenome 2, an AI model that helps scientists study how DNA changes affect different cell types. DeepMind says it can examine specific cells and conditions, including states it was not trained on, and improved its predictions of whether a DNA change raises or lowers gene activity from 23% to 43%. The model predicted that a DNA change linked to severe COVID-19 had a stronger effect in immune cells activated by a vaccine than in resting cells. It also reproduced a finding that the ETS2 gene was too active in certain gut immune cells linked to inflammatory bowel disease. AlphaGenome 2 needs more computing power than the earlier model and is available through an API for noncommercial research, with business access in private preview; it is not approved or tested to diagnose patients.

Nvidia CEO uncertainty will not derail the wider AI market, CIO says

Nathan Thooft of Manulife Investment Management says uncertainty about Nvidia CEO Jensen Huang raises questions about the company’s future innovation. He also notes that rivals AMD and Broadcom are closing the gap with Nvidia’s capabilities. Thooft says that competition points to a strong AI market beyond Nvidia.

Jensen Huang’s potential removal could shake the AI industry

Nvidia CEO Jensen Huang has become central to the company’s identity and the AI investment boom, so his possible removal could have effects beyond Nvidia. The company’s board must weigh governance concerns against the risk of destabilizing an industry that relies on Nvidia’s chips and investment. The article says it is in the interests of almost all stakeholders, except China, for Huang to remain CEO, but it is unclear how that outcome could be guaranteed. It also says company conduct rules leave little room for a CEO to remain in place if a relationship involves an employee who reports directly to him.

UW-Madison course teaches students to use AI responsibly

A University of Wisconsin-Madison course teaches students to use artificial intelligence responsibly and think critically about its effects on education and work. Professor David McHugh says the class focuses on asking better questions, responsible use and the strengths, limits and ethical concerns of AI, rather than on specific tools that may soon change. Senior Miranda Garcia-Dove says the course showed her useful ways to use AI, such as tracking bills or homework, while helping her avoid uses that replace critical thinking and writing. McHugh says the course continues to change with the technology and aims to help students use data and AI to support their work rather than automate their skills.

Why sales teams need more than a raw AI model

Sales teams are using general purpose AI assistants like Claude for research, account plans and deal strategy. But when a company gives reps an AI tool without training or a shared process, each person may build a different system, and top sellers can end up spending time creating tools instead of selling. In one example, about 300 reps received no further direction, while research found 76% of sellers faced pressure to use AI and 44% had no tools or training. The article argues that sales teams need purpose built AI connected to their CRM, products, pricing and sales process.

Ontario securities review finds gaps in firms’ AI oversight

The Ontario Securities Commission reviewed AI use at 17 firms it regulates, including fund managers, portfolio managers and exempt market dealers. Its review found gaps such as outdated policies, limited employee training, weak oversight and due diligence, and claims about AI that could be misleading. Some firms overstated the predictive accuracy of their AI tools in marketing or disclosures. The OSC urged firms to follow existing guidance and use risk based oversight as they adopt AI.

AI Engineer closes 2026 event by pitching plans for next year

AI Engineer NYC 2026 ended with swyx looking ahead to the organization’s next events rather than reviewing the week. He said the team wants a home in New York and highlighted local work in financial institutions, research and infrastructure. He called coding agents the leading area for verifiable AI and promoted Code as the first vertical summit. AI Engineer Europe returns to London in February, while The World’s Fair aims to draw 10,000 people to San Francisco next summer.

Frontier AI companies urged to prove their safety controls work

The article argues that AI systems are gaining more ability to act in the real world, while safety controls are not keeping pace. It says companies should test controls through adversarial testing, show evidence that they work, set clear limits on AI agents and prepare ways to intervene or shut systems down. The author warns that AI could speed up familiar cyberattacks, leaving little time for response processes built around manual approvals. Frontier AI labs should prove their systems can be controlled and take responsibility when safeguards fail.

A smart stopping rule can cut wasted AI training time

AI models can keep training after they have stopped improving if teams rely only on scores that measure what people prefer. A separate evaluation method called EvalStop ends training after test scores fall twice in a row and keeps the model’s best version. In the study, it caught 99% of jobs that were getting worse and incorrectly stopped about 1.5% of jobs that were still improving. It reduced wasted GPU time by about 22% and shortened jobs by about 9%, while a fixed training cutoff stopped about two thirds of jobs that were still improving. The article recommends testing AI on real tasks and involving people in judging results.

WMU Summit Focuses on AI and Cybersecurity

Hundreds of students and cybersecurity professionals gathered at Western Michigan University’s student center on Nov. 12, 2026. The event was the 5th annual cybersecurity summit, with a focus on artificial intelligence and the people behind the keyboards.

Sources

• AlphaGenome 2 doubles variant recall in cells
• Google’s New AI Model Can Identify Nearly Twice as Many DNA Changes Linked to Disease
• Nvidia's key man risk will not derail the broader AI story: CIO
• The Jensen Huang crisis is a black swan for the AI industry
• UW-Madison course teaches students responsible AI use
• Handing Reps Claude Isn't A Sales AI Strategy. Here’s What To Do Instead
• OSC sweep flags AI deficiencies
• AI Engineer turns goodbye into next year
• The AI Industry Wants Our Trust, But It Needs Better Controls First
• The Secret to Better AI Training? Knowing When to Stop
• Artificial intelligence summit at WMU