I think we're like in the 1996 era of Llms, and in a couple of months, the pace, things are changing. I think we're all going to kind of be looking at these days and looking at these pods and being like, man, remember how crazy those things were at the beginning and how bad they were?View on YouTube
Friedberg expected that within a few months of February 2024, LLM quality would advance so fast that early‑2024 models would soon look primitive and notably bad, like 1996 web search. In reality, the main releases in the next few months—Google’s Gemini 1.5 (announced Feb 15, 2024), Anthropic’s Claude 3 family on March 4, 2024, and OpenAI’s GPT‑4o on May 13, 2024—were meaningful but incremental improvements over GPT‑4, Claude 2.x, and Gemini 1.0, not an order‑of‑magnitude jump. Claude 3 and later 3.5 Sonnet slightly outperform GPT‑4 on many benchmarks, while GPT‑4o raises MMLU from about 86.5 to 88.7 and improves multimodal and voice capabilities, but these are refinements rather than a shift that makes earlier models look ‘bad’.(en.wikipedia.org) Even by late 2025, industry coverage characterizes ChatGPT’s evolution through GPT‑4o, GPT‑4.5, and GPT‑4.1 as steady, incremental progress with ongoing issues like hallucinations, rather than a single transformative break point.(techradar.com) Older models such as GPT‑4 and Claude 3 remain in active use and are still viewed as strong systems, not as obviously obsolete in the way mid‑1990s web search is compared with later generations.(help.openai.com) Thus, both the short timeline (‘a couple of months’) and the implied scale of quality change did not materialize as predicted.