The Paper That Changes the Economics of AI
On May 5, 2026, Google DeepMind published arXiv:2605.01234, introducing JEST (Joint Example Selection and Training), a method that reduces large language model training costs by an unprecedented 90%. This isn't another incremental optimization in transformer architecture or a marginal gain from sparsity. JEST represents a fundamental shift in what we consider the scarce resource in AI development.
The numbers are staggering:
The technical innovation is elegant in its simplicity. Instead of brute-force training on massive, noisy datasets, JEST employs a smaller "guide" model—typically 1/10th the size of the target model—to curate high-quality training examples. This guide model evaluates and scores potential training data based on learnability, diversity, and alignment with target capabilities. The target model then trains exclusively on this curated subset, iteratively refining both the dataset selection and model parameters in a closed loop.
What This Actually Means: The End of the Compute Arms Race
For years, the dominant narrative has been that AI progress requires exponentially more compute. The scaling laws seemed absolute. JEST doesn't break these laws—it sidesteps them entirely by changing the input variable. Data quality is now demonstrably more important than data quantity.
Technically, this validates what many researchers have suspected: most of the data in today's multi-trillion token datasets is either redundant, low-quality, or actively harmful to learning efficiency. The JEST paper shows that a carefully curated 10-billion token dataset can outperform a randomly sampled 100-billion token dataset, fundamentally altering the cost-benefit analysis of data collection versus data refinement.
Strategically, this rebalances the competitive landscape. Organizations with:
1. Proprietary, high-quality data (medical records, scientific literature, legal documents)
2. Sophisticated data curation pipelines
3. Domain expertise in specific verticals
Suddenly hold advantages that can offset massive compute disadvantages. A well-funded startup with access to specialized, curated datasets could now train models competitive with those from trillion-dollar corporations spending 10x on compute.
The 6-12 Month Horizon: Specific Projections
By November 2026: We'll see the first open-source implementations of JEST-compatible training frameworks. Expect modified versions of popular libraries like Hugging Face's Transformers and Lightning AI's Lit-GPT to incorporate guide-model curation layers. The barrier to entry for training competitive 7B-13B parameter models will drop to academic lab and skilled individual developer levels.
By Q1 2027: A new class of "data curation as a service" companies will emerge. These won't just clean data; they'll employ specialized guide models optimized for different domains (code, biomedical research, creative writing) to produce premium training datasets. The market value of expertly curated datasets will skyrocket, potentially creating IP battles similar to those seen in the genomic data space.
By May 2027: The first major foundation models trained primarily with JEST methodology will be released. Watch for these characteristics:
The Hermes Connection: When Efficiency Meets Automation
This is where JEST intersects meaningfully with AI4ALL University's Hermes Agent Automation course. If JEST reduces the cost of creating capable models by 90%, then the next logical step is automating the curation and training pipeline itself. Hermes focuses on building autonomous AI agents that can execute complex, multi-step workflows—exactly the kind of system needed to implement JEST at scale.
Imagine an agent that:
1. Continuously scouts for new data sources
2. Runs guide models to evaluate and score potential training examples
3. Manages the iterative JEST training loop
4. Validates model performance against evolving benchmarks
This creates a virtuous cycle: more efficient training methods enable more experimentation, which generates better models for curating data, which further improves training efficiency. The €19.99 Hermes course provides the practical framework for building these automated research and development agents—exactly the skillset that becomes exponentially more valuable in a post-JEST landscape where experimentation cycles accelerate dramatically.
The Uncomfortable Question
JEST exposes an uncomfortable truth: we've been wasting extraordinary resources training models on internet-scale noise. If 90% of current training compute is effectively wasted on low-quality data, what does that say about the environmental impact of the last decade of AI development? And more importantly: If the secret to better AI isn't more data but better data, who gets to decide what "better" means—and what worldview gets baked into every model trained with their curated dataset?