Page 15 - AI Vol 1: Foundations of AI
P. 15
5. OPEN-SOURCE MODELS
Open-source and closed-source (also known as
proprietary) are types of software licenses which dictate OPEN-SOURCE IS SEEN AS A
who can view and build upon source code. Software WAY OF DEMOCRATIZING THE
that is open-source generally means that the source TECHNOLOGY AND PREVENTING
code is freely available to anyone who can inspect it or LOCK-IN, BUT THERE IS CAUSE
build upon it (with certain restrictions). Closed-source FOR CONCERN.
software, as the name implies, means that the software
is not freely available.
With respect to LLMs, the popular models most people
are familiar with, such as GPT-4, Claude2, and Google’s OPEN-SOURCED
Gemini family of models, are all closed source.
However, there is an active and increasingly robust Source code is available to view,
open-source LLM community. Meta has released their learn from, copy, change and
share.
LLaMA2 model as open-source, allowing startups and
hobbyists alike to use and fine tune the model without Updates are made often and
cost. Mistral_AI’s Mixtral model, discussed above, was errors are corrected rapidly.
also released as open-source.
Greater risk of harmful results.
There is a large philosophical debate regarding open-
source and closed-source LLM approaches. The open- VS
source community is concerned that as LLMs and AI
generally become more powerful, there is a risk that
a few large companies will control the technology on CLOSED-SOURCED
which society will increasingly depend. Open-source
is seen as a way of democratizing the technology and Only the original developers
preventing lock-in. The concern with open-source of the source code are able to
LLMs is that they typically do not have the same built- modify.
in safety guardrails, discussed in the next part, that Any modifications attempted by
large technology companies employ. This means there a non-developer is illegal.
is a greater danger that these unrestricted models could
be used by bad actors to produce harmful results (e.g., Built in guardrails to keep user
misinformation, malware, bullying content). information safe.
In practice, there are several advantages to open-
source models. First, due to the active startup and
hobbyist community building on these models, there
FOUNDATIONS OF AI | LOZANOSMITH.COM VOLUME 1 | 15