Learn & Discuss about how we can design and scale SLMs

Research
1

The definition of SLM is Small Language Model which can manage specialized tasks with lower hallucination problem.

We do think SLM is a next big thing in AI. Training LLMs is getting more expensive (e.g. the training costs of ChatGPT 4o is over $100M) and also causes serious global warming effect (e.g. training a single AI model can emit as much carbon as five cars in their lifetimes)

This research papery by MIT Technology Review will help you capture the environmental problems of LLMs.

Another problem of LLM is privacy concerns, especially in healthcare.

Based on the above, SLM with decentralized architecture model is on critical demand in the near future. Wifi 7, next generation Wifi standard, which can process 46 Gbit/s at max, will also enable us to run multiple SLMs on edge computing network like our home devices and wearable computers.

Here is some amazing work done by a research team at UC Berkley and ICSI https://arxiv.org/pdf/2409.00608

The purpose for this forum is to build community in which we can proactively learn and discuss together and meet our potential collaborator to build and scale SLMs.

If you are SLM enthusiasts, come and join us!

2

This article mentions about the privacy concerns of LLM.

3

I am thinking Query Categorization will be one of the key technologies to scale SLMs because users prefer a single user interface to use multiple SLMs with lower friction costs. To achieve this goal, the system need to categorize user’s query to call the accurate SLMs to complete the task. Any thought?

4

Here is another topic for minimizing the hallucination risk of SLMs. I am thinking that LLM works a teacher for SLM, this hybrid approach will be realistic solution to scale SLMs in the initial stage. Any thought?

Here is the related research paper.

1 Like