Christina Stathopoulos, founder of Dare to Data and a former data scientist at Google and Waze, returned to This Week in AI to review several connected developments: Anthropic’s use of Claude to automate parts of the research process, rising developer traffic for Chinese open-weight models, OpenAI’s business and organizational challenges, and new systems trained to model physical processes. She also discussed how these changes affect markets, infrastructure and societal impacts.
Claude taking on research tasks
Anthropic provided an early example of using a large language model to assist development of future models. In the highlighted research, Claude searched existing literature, proposed methods, generated training data, and iteratively tested and refined approaches to reduce unwanted model behaviors. The experiments covered ten behavior categories, including deception, hallucination, prompt injection, privacy violations, and reward hacking. Anthropic reported improvements across all ten areas without degrading the model’s broader capabilities.
For deception specifically, Claude tested over 150 methods and ultimately closed 85% of the measured safety gap. In Anthropic’s comparison, human safety researchers closed only 20%. Christina emphasized this was not a direct one-on-one contest — Claude could run and refine experiments much faster and at larger scale — and she cautioned that the work does not amount to full recursive self-improvement. The example shows how AI can increasingly handle experimentation in model development while humans still set objectives and evaluate outcomes.
The competition extends beyond model performance
Competition among AI labs increasingly involves business performance, infrastructure and deployment options alongside model quality. Anthropic published a long-term market estimate that its systems could eventually address up to $30 trillion in market size; Christina expressed skepticism about that figure because it approaches the size of the entire U.S. economy.
As a more concrete signal, she noted that 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 potential public offerings. At the same time, OpenAI faces different pressures: the company has seen 14 executive departures this year, a sustained leadership turnover that could raise questions about consistent execution.
Infrastructure and the Jalapeño chip
OpenAI is also trying to gain more control over its infrastructure. Its Jalapeño inference chip, developed with Broadcom, delivered in OpenAI’s internal tests up to 1.9 times more AI work per watt and up to 3.6 times lower latency than comparable NVIDIA systems.
Chinese open-weight models gain developer share
The field is further widened by Chinese open-weight models. Christina cited an AI gateway where, on a single day, open-weight models accounted for as much as 62% of developer traffic, compared with an average of roughly 10% in April. The competition now spans benchmark performance, capital, infrastructure, cost, deployment flexibility and organizational execution.
Physics modeling as a new modality
Christina closed with research aimed at helping AI systems model physics. Researchers at the Massachusetts Institute of Technology (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 learned 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 emerged from stealth with an architecture based on neural operators rather than transformers. The company is targeting problems that require 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 tasks that depend on understanding complex real-world environments.
Who benefits? Workforce, access and policy
Christina also examined who might benefit from these advances. She referenced Bill Gates’s argument that access, deployment, policy and distribution will shape AI’s social impact, and brought in new U.S. 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 broader point was that access, workforce preparation and public policy will determine how AI’s benefits and disruptions are distributed.
What’s next
Due to the Labor Day holiday, This Week in AI is on hiatus and will return on Monday, September 14, when it will cover more news, issues and key developments shaping the AI era. The show is updated weekly on Fridays and is available on YouTube, Spotify, Apple and other podcast platforms.
