Over the past 3 months, the open source LLM space has been buzzing with activity. Through worldwide cooperation, the open source community has achieved groundbreaking results in record time!
While closed source models such as OpenAI's GPT (or Google's PaLM) are creating headlines and are being integrated in products rapidly due to their ease of use and higher performance, open source models have quietly been catching up in terms of performance. Moreover, open source models have a few important tricks up their sleeve like their openness - no surprises there - and the fact that they are in the hands of the users who can adapt them to their specific needs.
As we saw for generative AI imagery with stable diffusion, note that open source models prevent the risk of corporatization of AI models and thus prevent a strong reliance on technology designed in the States. We expect LLM business applications to increasingly benefit from a trend towards tailored, domain-specific open source models in the future. In the end, AI technology is as valuable as the domain-specific product in which it is put to use. In the accompanying slides, we zoom in on "recent" (things are moving FAST) open source LLM developments and what they mean for you.
Open Source LLM Performance
Illustration that open source LLMs are breaking barriers lightning fast:
-
Scalable Specific LLMs: Fine-tuning a personalised LLM on your own laptop
-
LLMs on Phones: Foundation models running on a Google Pixel phone
-
Multimodality: LLaMa-Adapter allowing efficient fine-tuning of multimodal models
Data Quality Scales Better Than Model Size, Creating Opportunities
-
A large number of open source LLM projects save time & money by training on relatively small, highly curated datasets
-
These datasets are built using synthetic methods (e.g. using larger LLMs and filtering their best responses)
-
Working towards data quality and specialised models instead of larger models might be a winning strategy.
What does it mean for your business?
When choosing an LLM, note that open source models may have valuable advantages for your use case:
- ease of fine-tuning domain-specific models
- increased control and transparency
Be aware of vendor lock-in with closed source models:
more freely integrable open source alternatives may suit your needs.
Bear in mind that while tailoring a business-specific LLM was deemed unfeasible a couple of months ago, it may be a valid option now or in the near future.
