Host Christina Stathopoulos, founder of Dare to Data and a former data scientist at Google and Waze, returned to This Week in AI with developments that stretched from Claude testing ways to improve model safety to Chinese open weight models gaining developer traffic and new systems learning to model physics. She also examined what Anthropic and OpenAI’s business moves, workforce forecasts, and debates over access reveal about how quickly the AI landscape is broadening.
Claude is taking on more of the research process
Anthropic provided an early example of AI helping improve future AI systems. In research Christina highlighted, Claude searched existing work, proposed methods, generated training data, and repeatedly tested and refined its approaches to reduce unwanted model behaviors. The experiments covered 10 such behaviors, including deception, hallucination, prompt injection, privacy violations, and reward hacking. Anthropic reported improvements across all 10 without degrading the model’s broader capabilities.
For deception, Claude tested more than 150 methods and eventually closed 85% of the measured safety gap. Human safety researchers closed only 20% in Anthropic’s comparison. Christina emphasized that this wasn’t a direct contest because Claude could run and refine experiments much faster and at a much greater scale. She also cautioned that the work didn’t amount to full recursive self-improvement.
This showed how AI could increasingly handle experimentation in model development, changing the pace and scale of research while humans still set the goals and evaluate the results.
Model performance is only one part of the frontier race
Competition among AI labs increasingly involves business performance, infrastructure, and deployment options alongside model quality. Anthropic estimates that the market for its systems could eventually reach $30 trillion, a long-term estimate that Christina treated skeptically because it approaches the size of the entire US economy. She also highlighted more concrete evidence of momentum in how Anthropic’s annualized revenue run rate rose from less than half of OpenAI’s at the start of the year to surpassing it within several months. Both companies are preparing for possible public offerings.
OpenAI faces a different set of pressures, and Christina highlighted its 14 executive departures this year. That sustained leadership turnover could raise questions about the company’s ability to execute consistently. OpenAI is also trying to gain more control over its infrastructure. Its Jalapeño inference chip, developed with Broadcom, delivered up to 1.9 times more AI work per watt and up to 3.6 times lower latency than comparable NVIDIA systems in OpenAI’s own testing.
Chinese open weight models are widening the field further. Christina cited an AI gateway where open weight models recently reached as much as 62% of developer traffic on a single day, compared with an average of roughly 10% in April. The competition now spans benchmark performance, capital, infrastructure, cost, deployment flexibility, and organizational execution.
Physics models could extend AI beyond language and images
Christina closed with research aimed at helping AI systems model physics. Researchers from MIT and Tsinghua University developed a pretraining approach that learned from more than one million synthetic interactions between moving particles and complex 3D objects, then applied those patterns to simulations involving wind, water, collisions, and light. The researchers described physics as a potential “third modality” for AI alongside language and pixels.
She also covered Accelerated Understanding, a startup that recently emerged from stealth with an architecture based on neural operators rather than transformers. The company is targeting problems involving enormous physical datasets, including chip design, robotics, extreme-weather forecasting, and geological exploration.
By learning directly from physical systems, these models could become valuable for simulation, engineering, robotics, forecasting, and other work that depends on understanding complex real-world environments.
What’s next
Christina also examined who could benefit from these advances. She discussed Bill Gates’s argument that access, deployment, policy, and distribution will shape AI’s social impact, and brought in new US Bureau of Labor Statistics projections showing job growth in areas including technical services and healthcare, while office and administrative roles face greater pressure from automation.
Her larger point was that access, workforce preparation, and public policy will determine how AI’s benefits and disruptions are distributed.
Due to the Labor Day holiday, This Week in AI will return on Monday, September 14, when we’ll dive into more of the news, issues, and key developments shaping the AI era. And check back each Friday for the latest episode, or watch on YouTube, Spotify, Apple, or wherever you get your podcasts.
