Inflection AI launches new model for Pi chatbot, which almost matches GPT-4
Join leaders In Boston on March 27 For A exclusive night of networking, knowledge, And conversation. Request A invite here.
Today, Inflection AI, THE Palo Based on the viola to start up based by deep mind co-founder Mustafa Soliman And LinkedIn co-founder Reid Hoffmann, announcement A new foundation model called Inflection-2.5.
Built on THE work do SO far, Inflection-2.5 surpasses THE the company original Inflection-1 significantly And almost matches OpenAI GPT-4 model, especially through STEM topics. He NOW powers THE the company Pi assistant, designed has take on ChatGPT And Gemini, And can be tested via mobile And Web.
THE move Brands THE last effort In THE quickly evolution AI space has take on THE dominance of OpenAI, which keep on going has clarify It is approach has development AI For humanity. Just recently, Anthropic released Claude 3 Opus, which became THE First of all model has beat GPT-4.
Carried out better but always delays behind GPT-4From It is creation, Inflection AI has has been building A “empathetic, useful And on" AI that actions more personally And familiarly that other models, including THE Google Tag series. THE business used unique empathetic fine tuning has give THE model behind Pi A signature personality And A exceptional equalizer (emotional quotient).
V.B. EventTHE AI Impact Tour – Boston
Were excited For THE following stop on THE AI Impact Tour In Boston on March 27. This exclusive, invite only event, In Partnership with Microsoft, will functionality discussions on best practices For data integrity In 2024 And beyond. Space East limit, SO request A invite today.
Request A inviteWith THE introduction of THE upgraded Inflection 2.5, THE to start up, which raised A $1.3 billion round In June 2023, East building up THE IQ appearance, covering areas as physics And mathematics. In A Blog job published Today, THE business said users talk with Pi, supported by Inflection 2.5, can discuss A range of the subjects, RIGHT Since discuss A hobby has coding, verification answers has A biology paper Or writing A business plan.
In terms of performance In marks, THE upgraded model watch substantial improvements on Inflection 1 through THE advice And closes on GPT-4 – although he always delays.
For example, on THE MMLU reference, measure performance through Tasks varying Since high school has Professional level difficulty, Inflection-2.5 brand 85.5, session just behind GPT-4 87.3. In the same way, In STEM exams, THE model carried out almost as GOOD as THE OpenAI model, rating 63 In THE Hungarian Mathematics exam (against 68 of GPT4) And 85th percentile In Physics GRE, against GPT-4 97th percentile.
In THE GSM8K reference, consisting of 8.5K high quality grade school mathematics problems, THE Inflection model brand 86.3, against GPT-4 92. In 0-shot HumanEval, designed has assess THE coded generation abilities, he brand 73.8 against GPT4 79.3.
While THE performance East not better that Google Tag 4, Inflection AI did indicate out that This “94% GPT-4 level performance" has has been reached with a lot more effective training that that do For THE OpenAI big language model (LLM).
According to has THE business, Inflection-2.5 took only 40% of THE training FLOP (calculate) of GPT-4 has get these results.
In addition...
Join leaders In Boston on March 27 For A exclusive night of networking, knowledge, And conversation. Request A invite here.
Today, Inflection AI, THE Palo Based on the viola to start up based by deep mind co-founder Mustafa Soliman And LinkedIn co-founder Reid Hoffmann, announcement A new foundation model called Inflection-2.5.
Built on THE work do SO far, Inflection-2.5 surpasses THE the company original Inflection-1 significantly And almost matches OpenAI GPT-4 model, especially through STEM topics. He NOW powers THE the company Pi assistant, designed has take on ChatGPT And Gemini, And can be tested via mobile And Web.
THE move Brands THE last effort In THE quickly evolution AI space has take on THE dominance of OpenAI, which keep on going has clarify It is approach has development AI For humanity. Just recently, Anthropic released Claude 3 Opus, which became THE First of all model has beat GPT-4.
Carried out better but always delays behind GPT-4From It is creation, Inflection AI has has been building A “empathetic, useful And on" AI that actions more personally And familiarly that other models, including THE Google Tag series. THE business used unique empathetic fine tuning has give THE model behind Pi A signature personality And A exceptional equalizer (emotional quotient).
V.B. EventTHE AI Impact Tour – Boston
Were excited For THE following stop on THE AI Impact Tour In Boston on March 27. This exclusive, invite only event, In Partnership with Microsoft, will functionality discussions on best practices For data integrity In 2024 And beyond. Space East limit, SO request A invite today.
Request A inviteWith THE introduction of THE upgraded Inflection 2.5, THE to start up, which raised A $1.3 billion round In June 2023, East building up THE IQ appearance, covering areas as physics And mathematics. In A Blog job published Today, THE business said users talk with Pi, supported by Inflection 2.5, can discuss A range of the subjects, RIGHT Since discuss A hobby has coding, verification answers has A biology paper Or writing A business plan.
In terms of performance In marks, THE upgraded model watch substantial improvements on Inflection 1 through THE advice And closes on GPT-4 – although he always delays.
For example, on THE MMLU reference, measure performance through Tasks varying Since high school has Professional level difficulty, Inflection-2.5 brand 85.5, session just behind GPT-4 87.3. In the same way, In STEM exams, THE model carried out almost as GOOD as THE OpenAI model, rating 63 In THE Hungarian Mathematics exam (against 68 of GPT4) And 85th percentile In Physics GRE, against GPT-4 97th percentile.
In THE GSM8K reference, consisting of 8.5K high quality grade school mathematics problems, THE Inflection model brand 86.3, against GPT-4 92. In 0-shot HumanEval, designed has assess THE coded generation abilities, he brand 73.8 against GPT4 79.3.
While THE performance East not better that Google Tag 4, Inflection AI did indicate out that This “94% GPT-4 level performance" has has been reached with a lot more effective training that that do For THE OpenAI big language model (LLM).
According to has THE business, Inflection-2.5 took only 40% of THE training FLOP (calculate) of GPT-4 has get these results.
In addition...
What's Your Reaction?
