Microsoft, I think, just launched the first real LLM product with thepublic release of GitHub Copilotlast week.
It’s free, every week, in your inbox.
It provides suggestions as you write code, something like autocomplete but for programming.
However, contrary to GPT-3, Codex has been finetuned just for programming tasks.
And it produces impressive results.
The success of GitHub Copilot and Codex underline one important fact.
When it comes to putting LLMs to real use, specialization beats generalization.
Codex is also much smaller than GPT-3, which means it is more memory and compute efficient.
It is atransformer modelthat has been trained on millions of code repositories.
With its huge training corpus and massive neural internet, Copilot mostly makes good predictions.
But sometimes, it might make dumb mistakes that the most novice programmer would avoid.
It doesnt think about programs in the way a programmer does.
Itsnot a replacement for human programmers.
In this regard, Copilot has been a stunning success.
While Copilots might output require some tweaking, it relieves most of the burden on developers.
Copilot has helped them save a ton of time in their day-to-day work.
Distribution and cost-efficiency
Product/market fit is just one of the several components of creating a successful product.
First, it needed the right technology, which it acquired thanks to itsexclusive license to OpenAIs technology.
Codex and Copilot were created off GPT-3 with the help of OpenAIs scientists.
Other large tech companies have been able to create large language models that are comparable to GPT-3.
But theres no denying that LLMs arevery costly to train and run.
Ben Allal referred toanother benchmarkused for Codex evaluation, which cost thousands of dollars for her own smaller model.
This is where Microsofts second advantage kicks in.
It runs inference and provides suggestions in milliseconds.
And more importantly, Microsoft is able to run and provide Copilot at a very affordable price.
However, running code generator LLMs at affordable rates is not impossible.
However, this is where the third piece of the puzzle kicks in.
This reduces the friction for developers to adopt Copilot as opposed to another similar product.
The market can take other turns.
But they wont change the fundamentals of sound product management.
you’re free to read the original articlehere.
