666: GPT-4

Episode Summary

Executive Summary: This episode celebrates GPT-4 as a major leap over GPT-3.5 in safety, accuracy, reasoning, multimodality, and multilingual performance. The host explains how larger context windows, better training data, and reinforcement learning from human feedback contributed to its gains, highlights real-world uses, and warns that open-source versions of similarly powerful models could create significant policy and misuse risks.

Main Topics: GPT-4 as a major upgrade over GPT-3.5: The host argues GPT-4 dramatically improves reasoning, consistency, and natural-language generation compared with GPT-3.5, based on personal testing and benchmark results. Safety, hallucinations, and alignment improvements: OpenAI’s safety work reportedly reduced disallowed-content responses and hallucinations, making GPT-4 less risky and more reliable than prior versions. Multimodal capabilities: GPT-4 can process visual inputs in addition to text, enabling tasks like interpreting fridge photos, turning drawings into websites, and improving performance on visual exams. Benchmark performance and multilingual ability: The episode emphasizes GPT-4’s strong results on professional exams and its ability to outperform competing models across many languages, including low-resource ones. How GPT-4 was improved technically: The host attributes progress to larger model scale, more carefully curated training data, larger context windows, and RLHF fine-tuning. Practical access and real-world applications: GPT-4 is available via ChatGPT Plus and API access, and it is already being used by products such as Duolingo, Be My Eyes, and Icelandic language preservation efforts. Future risks and policy concerns: The episode closes with concern that open-source versions of GPT-4-like models may lack OpenAI’s safety alignment, potentially enabling serious misuse.

Key Arguments: GPT-4 is materially safer than GPT-3.5, with OpenAI reporting fewer disallowed responses and fewer hallucinations. The model’s improved reasoning and long-context consistency make it useful for advanced writing, summarization, and professional problem-solving. Multimodal input is a breakthrough because it extends the model beyond text into images, unlocking new applications. GPT-4’s benchmark gains suggest performance approaching or exceeding human-level proficiency in some exam settings. Longer context windows and RLHF are key drivers of the model’s quality and instruction-following improvements. The model is already enabling real products and creative workflows, especially in coding, product development, accessibility, and language learning. Even if GPT-4 itself is relatively safe, future open-source clones could pose serious societal and policy risks if released without comparable safeguards.

Data Points: Reduction in disallowed-content responses: 82% less likely - OpenAI investigation comparing GPT-4 to GPT-3 Improvement in factual correctness: 40% more likely - GPT-4 relative to GPT-3.5 Uniform Bar exam percentile, GPT-3.5: 10th percentile - Benchmark cited to show legal exam performance Uniform Bar exam percentile, GPT-4: 90th percentile - Benchmark cited to show legal exam performance Biology Olympiad percentile, GPT-3.5: 31st percentile - Visual-capability benchmark comparison Biology Olympiad percentile, GPT-4: 99th percentile - GPT-4 performance with multimodal input Languages studied where GPT-4 beat rivals: 24 out of 26 - Compared with GPT-3.5, DeepMind’s Chinchilla, and Google’s PaLM GPT-4 context window: 8,000 tokens - Standard GPT-4 version mentioned in the episode GPT-4 extended context window: 32,000 tokens (~25,000 words / ~50 pages) - Larger GPT-4 version for long-context tasks GPT-3 context capacity: about 4,000 words - Used as a comparison for GPT-4’s longer context ChatGPT Plus subscription: about $20/month (US) - Access path to GPT-4 through the ChatGPT interface Episode references: 660, 626, 646, 667, 668 - Related episodes mentioned for tips, token explanation, and upcoming GPT-4 discussions

Pivotal Quotes: "GPT-4 is 82% less likely to provide responses on disallowed content than its predecessor, GPT-3." — Host: Discussing safety improvements and reduced harmful outputs "GPT-4 scored in the 99th percentile." — Host: Explaining the Biology Olympiad benchmark result for GPT-4 with visual capabilities "This GPT-4 really is incredible. If you haven't used it, I highly encourage you to." — Host: Personal endorsement of GPT-4 after testing it extensively

Implications: GPT-4 raises the bar for AI usefulness in coding, writing, translation, and multimodal tasks, but it also intensifies concerns about misuse if similar capabilities spread without comparable safety work.