The proliferation of AI models in recent times has reached an overwhelming point, with approximately 10 new models being introduced each week. However, it has become increasingly difficult to compare and differentiate between these models. This raises the question of their purpose and relevance.
The AI landscape is currently experiencing a peculiar phase, characterized by the emergence of models from both small, niche developers and well-funded organizations. Let’s examine some of the notable models that have been introduced recently:
1. LLaMa-3: Meta’s latest flagship language model, referred to as “open” despite ongoing disputes over its openness. This model is widely used within the community.
2. Mistral 8×22: Developed by a French outfit, this “mixture of experts” model leans towards larger sizes. However, the organization has deviated from its previous commitment to openness.
3. Stable Diffusion 3 Turbo: An upgraded version of SD3 designed to complement the open-ish Stability’s new API. The borrowing of the term “turbo” from OpenAI’s model nomenclature is somewhat peculiar.
4. Adobe Acrobat AI Assistant: This model allows users to interact with their documents and is likely a wrapper for ChatGPT, a well-known language model.
5. Reka Core: Developed by a small team that was previously employed by a major AI company, this multimodal model competes with larger models in terms of performance.
6. Idefics2: Built on top of recent Mistral and Google models, this multimodal model boasts a more open approach.
7. OLMo-1.7-7B: An expanded version of AI2’s LLM, known for its openness and potential for future advancements.
8. Pile-T5: A fine-tuned version of the reliable T5 model, optimized for code databases.
9. Cohere Compass: An embedding model focused on incorporating multiple data types to cover diverse use cases.
10. Imagine Flash: Meta’s latest image generation model, which employs a new distillation method to balance diffusion speed and quality.
11. Limitless: A personalized AI that adapts to user preferences and experiences across various platforms.
The number of models introduced or previewed in a week extends beyond the listed ones. Fine-tuned existing models, experimental models, and niche models are continuously being developed. Additionally, new tools for building and battling generative AI have also been released recently.
Given this constant influx of AI models, it becomes impossible to review each one individually. However, it is important to understand that not all models require equal attention. Models like ChatGPT and Gemini have evolved into comprehensive platforms that cater to various use cases. On the other hand, models such as LLaMa or OLMo, although they share a basic architecture, serve different purposes. They are designed to function as background services or components rather than standalone name brands.
Developers often create intentional confusion around their models, hoping to generate the same level of excitement as major AI platform releases like GPT-4V or Gemini Ultra. However, it is crucial to recognize that the significance of these releases may not directly impact most individuals.
A useful analogy is the automotive industry. Initially, there were only a few types of cars available. However, as time progressed, the market became flooded with numerous car models. Despite this abundance, individuals do not need to keep track of every new car release because not all of them are relevant or necessary. Similarly, AI is transitioning from a period of limited options to one of widespread proliferation. Even AI specialists struggle to keep up with the constant influx of new models.
It is worth noting that the AI field has always been active and evolving, even before the emergence of models like ChatGPT. Earlier on, fewer people paid attention to AI developments, but the technology was consistently advancing. Conferences like SIGGRAPH and NeurIPS served as platforms for machine learning engineers to collaborate and build upon each other’s work.
While the AI industry continues to progress, the advancements made by newer models are incremental rather than revolutionary. OpenAI’s groundbreaking change to machine learning architecture, which other companies have adopted, remains unparalleled. Therefore, the improvements seen in newer models are often marginal, such as achieving slightly higher benchmarks or generating more convincing language and imagery.
However, this does not imply that these models are insignificant. Each incremental step contributes to the overall advancement of AI technology. Some models address critical shortcomings or reveal unforeseen vulnerabilities. Although it is impossible to cover every model released, efforts are made to highlight the most noteworthy ones. A curated list of models that the ML-curious should be aware of is currently being compiled.
Rest assured, when a truly groundbreaking model emerges, it will become evident to everyone, not just to those following TechCrunch. The impact of such a model will be impossible to ignore, signaling a significant leap forward in AI technology.