From: aidotengineer
Over the past few years, a significant shift has occurred in the development of reliable AI solutions, moving towards more advanced agentic workflows [00:00:10]. Initially, many companies built AI wrappers, but the landscape changed dramatically as models got better at coding and AI adoption skyrocketed [01:11:39].
The Evolution and Limits of AI Models
The success of AI-powered tools like Cursor AI, which achieved $100 million ARR in 12 months, highlights not just improved models, but also new techniques for orchestrating these models to work effectively in production [01:01:00]. This orchestration is crucial because there are clear limits to model performance; issues like hallucinations and overfitting persist, and developers require more structured outputs [01:38:00].
For years, making models bigger and fitting them more data led to smarter AI [02:01:83]. However, this approach eventually “hit a wall,” with improvements slowing down despite additional data, as models reached their limits on existing tests [02:06:00]. This led to the exploration of new avenues and new training methods [02:24:26].
New Training Methods Pushing the Field Forward
In recent months, new training methods have emerged that are advancing the field [02:38:48]:
- Deep Seek R1 Model The Deep Seek R1 model is notable as the first model trained without using any labeled data [02:45:30]. This method is referred to as “real reinforcement learning,” meaning the model learned on its own [02:49:50]. This approach was reportedly used by OpenAI to train their reasoning models like 01 and 03 [02:57:97].
- Chain of Thought Thinking Modern reasoning models utilize Chain of Thought thinking at inference or response time to generate answers [03:03:00]. This allows models to “think” before providing an answer, enabling them to solve more complex reasoning problems [03:12:00].
Additionally, model providers are enhancing their models with more capabilities, such as tool use, improved research capabilities, and near-perfect OCR accuracy (e.g., Gemini 2.0 Flash) [03:24:00].
The Need for New Benchmarks
As models get better and the field progresses, traditional benchmarks have become saturated [03:41:20]. Consequently, new benchmarks are being introduced to capture the performance of these new reasoning models [03:44:80]. For example, the Humanities Last Last Exam measures performance on truly difficult tasks, where even the latest smart models still struggle [03:52:04].
Beyond Models: The Importance of Building Around AI
Ultimately, for an AI product to work effectively in production, success extends beyond just the models themselves; it hinges on “how you build around it” [04:10:95]. This involves evolving parallel to model training by learning to prompt models better and developing advanced techniques like Chain of Thought [04:20:00]. Other crucial advancements include:
- Retrieval-Augmented Generation (RAG): Grounding model responses with proprietary data [04:31:00].
- Memory: Essential for multi-threaded conversations and capturing context in long interactions [04:42:00].
- Long Context Models: Enabling new use cases with extended context windows [04:47:00].
- Graph RAG: Experimenting with hierarchical responses [04:52:00].
- Agentic RAG: Making workflows more powerful and autonomous [05:12:00].
While the field is still evolving, these techniques are critical. Deep understanding of the problem and a test-driven development approach are essential to find the right combination of techniques, models, and logic for a specific use case [05:22:00]. This approach involves continuous experimentation, evaluation, deployment, and monitoring to continuously improve the product for customers [05:52:00].