13x Faster and 10x More Efficient : Google’s DeepMind JEST Revolutionizes AI Training

DeepMind has introduced JEST, a revolutionary AI training framework that accelerates training speeds by 13x and enhances efficiency by 10x.


In the fast-evolving realm of artificial intelligence, breakthroughs often set the pace for transformative advancements. Recently, DeepMind, the pioneering AI research lab under Google’s umbrella, introduced a game-changing innovation: JEST (Joint Example Selection Technique).

Traditionally, AI models are trained on randomly chosen data points or based on individual relevance. JEST focuses on selecting the most helpful subsets of data for training. It uses two AI models: a smaller pre-trained model to evaluate data quality and a larger model being trained. The smaller model identifies high-quality data batches that are most effective for the larger model’s learning process.

This cutting-edge technology promises to revolutionize AI training processes by achieving speeds up to 13 times faster and enhancing energy efficiency by a staggering 10-fold compared to conventional methods.

For instance, consider the implications of reducing training time for complex models like GPT-4o from months to weeks, while simultaneously decreasing the colossal energy demands of AI data centers.

This advancement is crucial not only for its potential cost savings and environmental benefits but also for its broader implications on AI’s scalability and accessibility.

Follow us on Twitter here

Overview of JEST

Traditionally, AI model training relies on processing individual data points sequentially. However, JEST introduces a paradigm shift by focusing on batch-level selection. Here’s how it works:

  1. Batch Selection: Instead of training on every available data point, JEST employs a smaller AI model to assess the quality of batches from high-quality sources.
  2. Optimized Training: The selected high-quality batches are then used to train a larger AI model, resulting in optimized training efficiency.

This method allows JEST to achieve up to 13 times faster training speeds and 10 times higher power efficiency compared to conventional approaches, as highlighted in DeepMind’s recent research publication.

Here’s a breakdown of how JEST works, along with an example to illustrate:

Traditional Training:

Imagine you’re training a dog to identify different types of toys. Traditionally, you might throw a bunch of random toys (data points) for the dog to play with (train on). The dog might eventually learn to distinguish them, but it might take a while and involve picking up some irrelevant objects (low-quality data).

JEST Approach:

  1. Two Models: JEST uses two AI models. Imagine you have a well-trained assistant dog (smaller pre-trained model) who knows the basic types of toys.
  2. Quality Check: This assistant dog first sniffs through a pile of various objects (data batches) and picks out only those that seem like toys (high-quality data).
  3. Focused Training: Then, you train your main dog (larger model being trained) using only the selection of objects chosen by the assistant. This focused training with relevant objects helps the main dog learn to identify the toys much faster.


  • Reduced Training Time: By focusing on high-quality data, the main dog learns quicker, just like you wouldn’t waste time showing your dog random objects.
  • Increased Efficiency: JEST avoids wasting resources on irrelevant data points, making the training process more energy-efficient.


Imagine training an image recognition model to tell the difference between cats and dogs. Traditionally, it might be trained on random images from the internet. JEST could involve:

  • A smaller pre-trained model that already knows basic features of cats and dogs.
  • This pre-trained model would scan large batches of images and select only those that clearly show cats or dogs, discarding blurry or irrelevant pictures.
  • The main model would only be trained on these high-quality images of cats and dogs, leading to faster and more accurate learning.

Overall, JEST acts as a data filtering system, ensuring the main model gets the most relevant and helpful information for efficient training.

Read More:iVP Semiconductor: India’s First Fabless Chip Company Eyes $70 Million Revenue – techovedas

Technical Breakdown

The technical underpinnings of JEST are detailed in DeepMind’s research paper, demonstrating:

  • Batch Grading: The smaller JEST model grades batches based on data quality.
  • Performance Comparison: Comparative analysis against state-of-the-art methods like SigLIP, showcasing superior efficiency in terms of speed and Floating Point Operations per Second (FLOPS).

Graphical representations provided in the paper illustrate the significant efficiency gains over traditional AI training methodologies.

Top 10 Technical Universities of 2024 for International Courses | by techovedas | Jul, 2024 | Medium

Implications for AI Development

The implications of JEST for the AI industry are profound:

  • Cost Savings: By reducing the number of iterations and computational requirements, JEST has the potential to lower the astronomical costs associated with training large-scale AI models. For instance, training models like GPT-4o reportedly cost millions of dollars, and JEST could mitigate such expenses.
  • Expert Data Curation: However, the success of JEST hinges on the quality of the initial training data. High-grade, curated datasets are essential for its effectiveness, posing a challenge that may limit its accessibility to expert-level researchers initially.

Environmental and Economic Impact

The environmental footprint of AI data centers has become a pressing concern.

In 2023, AI workloads consumed approximately 4.3 GW of power, nearly equivalent to Cyprus’ annual electricity consumption.

JEST’s enhanced energy efficiency could significantly reduce these demands, aligning with global efforts towards sustainable technology development.

Read More: 7 Exciting Technology Products From CES 2024 – techovedas

Future Outlook

Looking ahead, the adoption of JEST by major players in the AI landscape remains uncertain but promising. If widely implemented, JEST could pave the way for faster advancements in AI capabilities while mitigating environmental impact and lowering operational costs.

Read More: What is India’s DLI and PLI Schemes : Catalyzing a New Era in Manufacturing and Innovation – techovedas


DeepMind’s JEST represents a watershed moment in AI training methodologies, offering unprecedented gains in speed and efficiency. As the industry grapples with the dual challenges of technological advancement and environmental sustainability, JEST stands as a beacon of hope for a more efficient and responsible AI future.

Editorial Team
Editorial Team
Articles: 1925