GPT-4.5 = Big Model Energy | YC Decoded

TL;DR

GPT-4.5 stands out for more natural conversation, emotional intelligence, creative work, and fewer hallucinations, but it trails specialized reasoning models and costs much more than GPT-4o. It achieved 61.9% accuracy on SimpleQA and reduced hallucinations to roughly 37%. Its combination of stronger subjective performance and clear practical limitations makes the details worth reading.

Transcript

GPT 4.5 is finally here and it's open ai's largest and most humanlike model to date it's really the next step in scaling up unsupervised learning it has a lot more deeper understanding of the world and of Human Experience 4.5 excels at natural conversation creative tasks and complex planning it also hallucinates far less than previous models howeve... Read More

Key Insights

• 💋 GPT 4.5 marks a milestone in AI development, offering deeper emotional understanding and conversation skills compared to its predecessors.
• 🧑‍🏭 The accuracy improvements in answering fact-based questions highlight its enhanced performance on benchmarks, making it a more reliable assistant.
• 💗 Its ability to produce humorous and ironic content indicates a growing sophistication in understanding human context and creativity.
• 😒 The high operational costs of GPT 4.5 create barriers for widespread use, suggesting it may not be ideal for all deployment scenarios.
• ❓ The upcoming integration of reasoning capabilities with broad understanding signifies a promising direction for future AI models.
• 🤨 OpenAI's emphasis on human feedback in refining models raises concerns about subjective evaluations but aims to enhance user experience.
• 🎰 The ongoing evolution of AI models underscores the potential for profound changes in how machines interact with human emotions and creativity.

Questions & Answers

Q: What sets GPT-4.5 apart from GPT-4o and OpenAI's reasoning models?

GPT-4.5 offers deeper understanding of human intent, more natural conversation, stronger creative output, and fewer hallucinations than GPT-4o. However, specialized reasoning models such as o1 perform better on structured reasoning, advanced math, complex STEM tasks, and difficult coding challenges.

Q: How accurate is GPT-4.5 compared with GPT-4o?

GPT-4.5 achieved around 61.9% accuracy on SimpleQA, a benchmark for single-relation factoid questions. GPT-4o scored 38.4% on the same benchmark, suggesting GPT-4.5 is more effective for general factual inquiries.

Q: Does GPT-4.5 hallucinate less than GPT-4o?

Yes. Its hallucination rate was roughly 37%, compared with 61.2% for GPT-4o, making GPT-4.5 more trustworthy for general inquiries according to the transcript.

Q: What creative tasks does GPT-4.5 handle well?

GPT-4.5 performs well when drafting emails, generating imaginative stories, telling jokes, and brainstorming ideas. Testers also found its prose more humanlike and noted that it could be funny and understand irony better than other models.

Q: Why were reactions to GPT-4.5 relatively muted?

The transcript says reactions were muted because GPT-4.5's improvements over GPT-4o appeared incremental. Although it improved accuracy, emotional intelligence, creativity, and hallucination rates, it was not presented as a frontier reasoning model.

Q: How expensive is GPT-4.5 compared with GPT-4o?

GPT-4.5 is 30 times more expensive per input token and 15 times more expensive per output token than GPT-4o. Those costs make it unlikely to be suitable yet for deployments that need to operate at scale.

Q: How was GPT-4.5 evaluated for emotional intelligence and writing quality?

Researchers relied partly on human feedback and “vibes testing” because qualities such as emotional intelligence, writing quality, and model feel are subjective. Human trainers compared its outputs with GPT-4 and provided feedback about where the model was better or worse.

Q: What does GPT-4.5 suggest about the future of OpenAI models?

GPT-4.5 shows that scaling unsupervised learning can still improve accuracy, emotional intelligence, and creativity, though the gains may be increasingly incremental. Sam Altman suggested that broad knowledge and intuition like GPT-4.5's could converge with the reasoning capabilities of models such as o3 in a unified architecture that might appear in GPT-5.

Summary & Key Takeaways

• GPT 4.5 is OpenAI's most advanced model to date, with significant improvements in emotional intelligence and conversational abilities, achieving 61.9% accuracy in benchmarks.

• Despite being more capable in creative tasks and exhibiting less hallucination, its incremental advancements over GPT 4 have led to mixed reactions in the tech community.

• OpenAI anticipates future models will combine the broad understanding of GPT 4.5 with cutting-edge reasoning capabilities, hinting at a transformative direction in AI development.