This strategy cannot be abused
Posted: Sat Dec 07, 2024 5:42 am
More than one example is considered n-shot. N-shot does not change the model, unlike fine tuning. You are simply giving examples right before asking for an answer, every time you ask a question.
: llm models have a maximum context size and their price depends on the size of the message. Fine tuning can eliminate the need for n-shot examples, but it takes longer to get it right.
Other rapid engineering techniques
there are other prompt engineering techniques, such as chain of thought, that force models to think out loud before coming up with an answer.
This increases the quality of the response, but at the cost of duration, cost and speed.
My recommendation
although each project will have its portugal mobile phone number own needs, i'm going to give my opinion on a solid approach.
A good starting point is to use a standard model that balances speed and quality, such as gpt-4o mini. Start by looking at the quality of the responses, the speed of response, the cost, the needs of the context window and decide what needs to be improved from there.
Next, with a limited use case, you can try simple notice engineering, followed by rag, and finally fine tuning. All models that go through these processes will improve their performance, so it can be difficult to decide which to use.
Privacy considerations
in an ideal world, every llm would be 100% under your own control, and nothing would be exposed anywhere.
Unfortunately, this is not what we see in practice, and for very good reasons.
The first is simple: it requires engineering to host and maintain a custom model, which is very expensive. When the hosted model experiences downtime, business metrics are affected, so the deployment must be very robust.
Another reason is that industry leaders - such as openai, google and anthropic - are constantly releasing newer, more capable and cheaper models that make any fine-tuning work superfluous. This has been true since the release of chatgpt 3.5 and it doesn't seem like it will change.
If your use case has extremely sensitive data, it makes sense to use a model and optimize it for your use case. If gdpr is a priority, there are many gdpr-compliant models available.
Building after selecting your llm
once you've selected llm, you can start figuring out how you'll build and maintain your ai project. As an example, i'll take the type of project i'm most familiar with: an ai agent or ai chatbot.
You can answer the following questions to narrow the scope of your project:
where would i like my ai agent to live? (slack, whatsapp, a website widget, etc.)
what knowledge should you have, where is that knowledge?
What other functions should it have, apart from answering knowledge questions?
Should it be activated when something happens somewhere in the company?
Download engineering to save money
maintaining a tight budget is essential to making your project a reality. One of the ways to achieve this is to reduce engineering time by decoupling requirements.
: llm models have a maximum context size and their price depends on the size of the message. Fine tuning can eliminate the need for n-shot examples, but it takes longer to get it right.
Other rapid engineering techniques
there are other prompt engineering techniques, such as chain of thought, that force models to think out loud before coming up with an answer.
This increases the quality of the response, but at the cost of duration, cost and speed.
My recommendation
although each project will have its portugal mobile phone number own needs, i'm going to give my opinion on a solid approach.
A good starting point is to use a standard model that balances speed and quality, such as gpt-4o mini. Start by looking at the quality of the responses, the speed of response, the cost, the needs of the context window and decide what needs to be improved from there.
Next, with a limited use case, you can try simple notice engineering, followed by rag, and finally fine tuning. All models that go through these processes will improve their performance, so it can be difficult to decide which to use.
Privacy considerations
in an ideal world, every llm would be 100% under your own control, and nothing would be exposed anywhere.
Unfortunately, this is not what we see in practice, and for very good reasons.
The first is simple: it requires engineering to host and maintain a custom model, which is very expensive. When the hosted model experiences downtime, business metrics are affected, so the deployment must be very robust.
Another reason is that industry leaders - such as openai, google and anthropic - are constantly releasing newer, more capable and cheaper models that make any fine-tuning work superfluous. This has been true since the release of chatgpt 3.5 and it doesn't seem like it will change.
If your use case has extremely sensitive data, it makes sense to use a model and optimize it for your use case. If gdpr is a priority, there are many gdpr-compliant models available.
Building after selecting your llm
once you've selected llm, you can start figuring out how you'll build and maintain your ai project. As an example, i'll take the type of project i'm most familiar with: an ai agent or ai chatbot.
You can answer the following questions to narrow the scope of your project:
where would i like my ai agent to live? (slack, whatsapp, a website widget, etc.)
what knowledge should you have, where is that knowledge?
What other functions should it have, apart from answering knowledge questions?
Should it be activated when something happens somewhere in the company?
Download engineering to save money
maintaining a tight budget is essential to making your project a reality. One of the ways to achieve this is to reduce engineering time by decoupling requirements.