🔝 GPT-4-Turbo has taken the top spot in the ranking of LLMs. Let's explain how it's compiled

Last week, OpenAI released an updated version of the GPT-4-Turbo model. According to reviews, it excels at generating code and produces texts with less "fluff." After its release, GPT-4-Turbo claimed the first position in the ranking of large language models (LLMs) — the Arena Leaderboard. This ranking is based on user voting and pairwise comparisons.

❓ What do the parameters in the Arena ranking mean?

⚪️ Arena Elo — the rating of LLMs, similar to the one used in chess. It indicates how well a model performs in competitions against others. In these competitions, users choose the best model. The comparison happens in a blind testing mode: users are presented with two windows to interact with unknown models, and they can converse with them until they determine the best LLM based on their personal preferences.

⚪️ CI, 95% — the confidence interval of the model's rating. What does this mean: for example, if the GPT-4-Turbo model has an Elo rating of 1260 and a CI of 95% of +5/-5; it means that with a 95% probability, the actual rating of the model will be within the range of 1255–1265 (+5 and -5 from the rating of 1260). The CI helps track the objectivity of the LLM's position in the table, which is why three models immediately occupy the 1st place: their ratings "overlap" in their confidence intervals.

⚪️ Votes — the number of user votes in Arena. As of today, Arena operates with 670 thousand votes.

⚪️ Organization — the company that owns the LLM.

⚪️ Knowledge Cutoff — the period the model's data is trained; for example, GPT-4-Turbo has world knowledge until December 2023.

👍 How can I test it?

In our @GPT4Telegrambot, you can work directly with four leading models: GPT-4-Turbo, GPT-4-0125, and Claude 3 — versions of Opus and Sonnet.

📕 More stories:

⚡️ Claude 3: a new AI model from the main competitor of OpenAI

📎 @GPT4Telegrambot Now Features File Recognition

#OpenAI #Claude @hiaimediaen