AGI Reality Check, Gemini 2.5 Update, Are Your AI Chats Safe & Fun with Veo3 - EP99.07-06-05

Chris and Larry discuss the impact of lighting on their on-camera appearance and the costs associated with using Google's new video model, VO3, which charges $3.75 for five seconds of video. They compare this to Replicate's pricing and reflect on their experimentation costs, estimating nearly $100 spent on multiple iterations. The conversation shifts to recent allegations against the Chinese lab DeepSeek, which updated its R1 reasoning model, with speculation about its training sources, including Google Japan. The hosts agree that competition among models often revolves around bragging rights about training sources.

They discuss the release of Gemini 2.5 Pro, which claims to excel in various benchmarks and has improved creativity and response formatting. User feedback suggests that the latest version may not be as effective as a previous iteration from March. The new model features a "thinking budget" of up to 32,000 tokens, enhancing its processing capabilities. The hosts note Google's steady update strategy, contrasting it with OpenAI's more abrupt announcements, and highlight the importance of server resources in maintaining model performance.

Concerns about the unpredictability of model availability and the confusing naming conventions used by labs are expressed. They mention that Gemini 2.5 Pro leads in tech-related tasks, with other versions closely following. One speaker prefers Flux context over GPT image for image generation due to better realism, while another shares their experience with Claude Sonnet, which effectively executed multiple queries simultaneously.

The conversation highlights the extensive effort behind AI models and their potential, with speculation on whether they could outperform current evaluations when equipped with the right tools. Anthropic's announcement of their four series model, designed for agentic capabilities, is noted, alongside the effectiveness of the 3.5 Sonnet model, which has increased user reliance. A comparison between Google’s Gemini 2.5 and Claude reveals that Gemini can handle 3000 files in a single request, enhancing context utilization.

Concerns about data ownership arise, particularly regarding the implications of sharing information with large model providers. The concept of "compaction" is introduced, where chat contexts are synthesized into manageable forms as they reach limits. Speculation about the competitive edge of larger context windows suggests Google may be developing an infinite context model. The shift in developer preferences from Claude to Gemini Pro for better output, especially in creative coding and 3D visualizations, is noted.

Skepticism surrounds the hype of GPT 4.5, perceived as a failed training run for GPT 5, while OpenAI models are acknowledged for their historical excellence in vision tasks. Speaker 1 discusses the limitations of raw AI models, particularly in dynamic environments, emphasizing the need for quick updates. He expresses concerns about Windsurf's decision-making and critiques the potential damage to developer goodwill stemming from fears about OpenAI training on their outputs.

The conversation shifts to Yann LeCun's skepticism regarding large language models (LLMs) achieving artificial general intelligence (AGI). LeCun argues that merely scaling LLMs will not yield human-level AI, asserting they lack true intelligence and agency. Participants reflect on this perspective, acknowledging that while LLMs can perform useful tasks, they do not possess the inventive capabilities of human experts. Despite their limitations, LLMs are recognized for their rapid improvement and impact across various industries.

Discussion also touches on the potential peak of LLM capabilities, with one speaker expressing confidence that LLMs will eventually be viewed as essential productivity tools. Concerns about the pace of technological change are raised, with a call for proactive policymaking to address potential inequalities. Criticism is directed at alarmist rhetoric from some industry leaders, suggesting it may be a fundraising tactic.

A court order involving OpenAI requires the retention of all ChatGPT logs, including deleted chats, raising significant privacy concerns. The New York Times is suing OpenAI for allegedly using their data for training, highlighting the potential depth of personal information shared with AI. The speaker emphasizes that consumer protections regarding data usage should be a fundamental right, advocating for a business model prioritizing user privacy over ad revenue.

The anticipation of AI becoming a central interface in personal and professional lives is expressed, predicting a new "model stack" for productivity. The conversation compares current technological shifts to past transitions, highlighting the slow societal adaptation to new technologies. The belief is voiced that most individuals who learn and adopt AI will experience increased productivity and satisfaction.

Speaker 1 prefers using Claude over Opus due to reliability issues with token capacity, while Sonnet meets their needs effectively. The models, particularly the four series, are designed for long-form tasks, and there's an intriguing concept of an internal "clock" that influences decision-making. Speaker 2 mentions that this episode was their first in-person recording, creating a different dynamic. The conversation wraps up with Speaker 1 rating the episode as average or below average, while Speaker 2 invites listener engagement.