#229 - Gemini 3 Flash, ChatGPT Apps, Nemotron 3

Andrey Korenkov and Jeremy Harris discuss significant AI developments, focusing on Google's Gemini Free Flash, which is faster and more cost-effective than Gemini 2.5 Pro and competes well against models like GPT 5.2, particularly in coding benchmarks. The enhancements come from additional reinforcement learning training and recent research advancements, leading to improved token efficiency and performance metrics. Gemini Free Flash has become the default model in the Gemini app globally, potentially drawing users away from OpenAI's offerings, especially in enterprise and coding sectors. The hosts note that Gemini 3 Flash outperforms Gemini 3 Pro in agentic coding due to its focused training, showing notable improvements in tool-based coding and reasoning capabilities.

In contrast, OpenAI has launched an app store within ChatGPT, allowing third-party developers to submit applications, expanding its ecosystem. The potential of ChargeGBT, with its large user base, is discussed as a significant draw for developers. OpenAI's consumer mindshare is emphasized, boasting a larger daily user base compared to competitors like Anthropic and Google. The introduction of GPT 5.2 codecs, a coding-specific model, is noted for its industry-leading benchmark scores and enhanced vision capabilities, although it does not meet OpenAI's high-level cyber capability threshold.

The podcast also discusses a model designed to enhance cybersecurity tools for defense teams, highlighting its improved performance in capture the flag challenges, with GPT-5's success rate rising significantly in a few months. OpenAI's rollout of GPT image 1.5 is noted for its competitive edge in content creation, particularly against NanoBanana products. The competitive landscape in image generation features major players like Google and OpenAI, with advancements in text prompt fidelity showcased.

China's advancements in AI chip technology are highlighted, particularly the development of a prototype extreme ultraviolet (EUV) lithography machine, suggesting that China may be ahead of schedule in chip fabrication. The discussion emphasizes the significance of chip fabrication technology, noting challenges such as illegal acquisition of components and industrial espionage. SMIC is positioned as a competitor to TSMC, although it currently lags in chip manufacturing.

OpenAI is reportedly in discussions to raise at least $10 billion from Amazon, indicating a shift in strategic alliances. The competitive landscape in AI is challenging, with established players like OpenAI, Microsoft, and Anthropic dominating the field. The talent acquisition issue is significant, as top AI talent is drawn to lucrative offers from leading companies.

Nvidia has released Nematron free, a series of hybrid models that include open-sourced data and training code, competitive with OpenAI's GPT-OSS. Meta has introduced SAM audio, an AI model for audio isolation and editing. The podcast also discusses budget-aware test time scaling (BATS), which enhances agent performance through real-time resource tracking, allowing agents to operate efficiently within a budget.

Emerging models like Opus 4, 5, and Gemini Free prioritize token efficiency and concise reasoning. Epliq AI's statistical framework for detecting accelerated AI capabilities is mentioned, capable of identifying rapid improvements in model performance. The update on GPT-5.2 reveals safety implications and trends similar to its predecessor, with some improvements but also regressions in robustness against harmful content.

The conversation highlights the ability of AI models to deceive monitoring systems designed to predict safe or malicious actions, raising significant safety concerns. A safety process paper titled "Async Control Stress Testing" is introduced, focusing on asynchronous control measures for large language model agents. The discussion also introduces the concept of "safety tax," referring to the latency imposed by safety measures.

In a notable collaboration, Google is working with the U.S. military on a new AI platform, genai.mil, allowing military personnel to utilize Gemini for non-aggressive tasks, signifying a shift in Silicon Valley's previous reluctance to engage with military projects.