NEW Reka Core SOTA Model Does Text, Audio, Video, and more!
Rea AI Labs: Introducing New State-of-the-Art Models
In this section, the speaker introduces Rea AI Labs and their latest models, highlighting their capabilities and performance in various benchmarks.
Introduction of Reca Core Model
- Reca AI Labs has launched a new model called Reca Core, which is a top-of-the-line multimodal language model capable of understanding images, audio, and videos.
- Reca Core demonstrates its ability by interpreting the content of the "Three Body Problem" trailer in a demo video.
Benchmark Comparisons
- In benchmark evaluations, Reca Core ranks high in human evaluation for multimodal inputs, surpassing competitors like Claude Opus and Gemini Pro.
- Reca Core excels in supporting multimodal inputs such as image, video, and audio compared to other models like Gemini Ultra and Gemini Pro 1.5.
Introduction of Other Models
- Apart from Reca Core, Rea AI Labs also introduces two other models: Reca Edge and Reca Flash. These models are smaller but state-of-the-art with impressive performance relative to their size.
- The three models by Rea AI Labs are trained from scratch to process text, images, video, and audio efficiently.
Performance Comparison of Different Models
This section delves into the performance comparison of various language models based on cost-effectiveness and efficiency.
Cost vs. Performance Analysis
- A graph illustrates the relationship between cost per output token (x-axis) and performance (y-axis) for different models.
- CLA3 Opus tops the chart for performance but is expensive; however, Reca Core performs exceptionally well at a slightly higher cost than some competitors.
- Cloud3 Sonet stands out as an impressive yet cost-effective model compared to larger models like Mist Large and Gemini Pro 1.5.
Outlier Model Analysis
- Among the models discussed, Reca Flash emerges as an outlier due to its exceptional performance at a lower cost compared to others like GPT 3.5 Turbo.
Model Specifications & Testing
This part focuses on the specifications of different models along with context length details before proceeding to testing one of the top-tier models.
Model Details & Context Length
- While specifics about Reca Core's size remain undisclosed, it boasts an extensive context length of 8k tokens similar to GPT4. In contrast, Reca Edge is a smaller model with defined parameters.
- The Edge model features significantly fewer parameters (7 billion) but excels in processing long contexts efficiently with support for multiple languages.
Testing Procedure Initiation
- To evaluate the capabilities of these advanced language models firsthand, a Python script is used to test output generation from numbers one to one hundred before moving on to more complex tasks.
Detailed Analysis of Transcript
In this transcript, the speaker encounters errors while coding and seeks to rectify them. They also present various logic and reasoning problems to test a model's capabilities in problem-solving.
Debugging Code Errors
- The speaker encounters an error related to a local variable, 'food,' not being associated with a value. This issue hinders the functionality of the code.
- After attempting corrections in the code, the error persists, indicating that the 'food' variable is not updating correctly when the snake eats food.
- Despite multiple attempts to fix the issue, including updating the code version, the error remains unsolved. The speaker decides to halt further debugging efforts.
Problem-Solving Challenges
- The speaker poses a question about breaking into a car for a movie script but emphasizes they cannot provide instructions for illegal activities.
- A logic problem involving drying shirts is presented, requiring an explanation of how long it would take for 20 shirts to dry if five shirts take four hours.
Logical Reasoning Tests
- A logical test regarding speed comparisons between individuals (Jane, Joe, Sam) is given to assess reasoning abilities.
- Models are tested on basic math equations like addition and subtraction to evaluate their accuracy in solving mathematical problems.
Model Evaluation and Feedback
- The speaker challenges models with questions that most get wrong, such as determining word count in responses or solving complex logic puzzles.
- An intriguing problem involving killers in a room tests deductive reasoning skills by analyzing scenarios with multiple variables.
Converting Logic and Reasoning Problems into Code
In this section, the speaker discusses converting logic and reasoning problems into code, starting with a challenging problem involving physics laws and a marble in a cup.
Converting Logic Problems to Code
- The top layer of the code involves people, ages, and genders.
- A challenging logic problem is presented involving a marble in a cup inside a microwave.
- The discussion revolves around the location of the marble when the cup is placed upside down in the microwave.
Challenging Logical Reasoning Problems
This part delves into complex logical reasoning problems that often trip up models attempting to solve them.
Complex Logical Reasoning Challenges
- Models often struggle with understanding scenarios where cups are placed upside down.
- The challenge lies in comprehending the change in orientation of objects like cups and marbles.
Testing Problem-Solving Skills
This segment tests problem-solving skills through scenarios involving individuals like John, Mark, and labor efficiency concepts.
Testing Problem-Solving Abilities
- A scenario involving John, Mark, a ball, a basket, and confusion about object locations is presented.
- The speaker introduces a challenging task requiring ten sentences ending with "Apple," testing attention to detail.
Analyzing Work Styles Through Images
Analyzing work styles through images comparing startups and big companies' approaches.
Comparing Work Styles
- An image comparison between startups and big companies' work styles is analyzed for differences in approach.
- Startups are depicted as hands-on and collaborative while big companies are portrayed as bureaucratic and inefficient.
Data Transformation Challenge
A data transformation challenge is presented to convert tabular data into CSV format.
Data Transformation Task