Last week, Meta said it hoped to launch Llama 3 within the next month, the latest iteration of its large language model (LLM) that powers generative AI assistants.
Meta’s president of global affairs, Nick Clegg, signalled it was set to be a busy year. “There will be a number of different models with different capabilities, different versatilities [released] during the course of this year, starting really very soon.”
Joelle Pineau, its vice president of AI research, added. “Our goal over time is to make a Llama-powered Meta AI be the most useful assistant in the world. There’s quite a bit of work remaining to get there.”
Currently, Llama 2 models come in 7 billion, 13 billion, and 70 billion parameter sizes, and while the team didn’t not talk about the sizes for Llama 3, it’s rumoured to have around 140 billion parameters.
In order to increase AI infrastructure, Meta has accumulated 350,000 highly sought-after H100 GPUs over the past year — a quantity that far outpaces that of its rivals.
All this signals that Meta could be a serious contender in the open AI race, where currently Claude 3, GPT-4, Bard, Command R+ and Mistral dominate.
Just a couple of months ago, Meta’s research teams released a guide called ‘Prompt engineering with Llama 2’ on Github, meaning developers, researchers and AI enthusiasts are keeping a keen eye on the platform for a fresh drop.
Still, there’s plenty to learn from the guide, even if the model is being updated soon.
What to do
The guide advises that detailed, explicit instructions produce better results than open-ended prompts, and it helps to provide a persona from the off.
For example:
- Explain this to me like a topic on a children’s educational network show teaching elementary students.
- I’m a software engineer using large language models for summarisation. Summarise the following text in under 250 words:
- Give your answer like an old school private investigator hunting down a case step by step.
Formatting also matters, so try bullet points and use return as a JSON object.
The guide also provides advice on including restrictions for improving accuracy. For example:
- Only use academic papers.
- Never give sources older than 2020.
- If you don’t know the answer, say that you don’t know.
Adding encouraging step-by-step thinking also significantly improves the ability of LLMs to perform complex reasoning. So the below is more likely to lead to the correct answer, than simply just the question.
“Who lived longer, Elvis Presley or Mozart? Let’s think through this carefully, step by step.”
This is called a CoT or Chain-of-Thought prompt.
To receive an answer without a cheery, extraneous prefix, like “Sure, here’s information on…” and to get a more usable JSON format, you need to be specific, and ideally provide a sample answer. The guide provides the following sample prompt:
You are a robot that only outputs JSON.
You reply in JSON format with the field ‘zip_code’.
Example question: What is the zip code of the Empire State Building? Example answer: {‘zip_code’: 10118}
Now here is my question: What is the zip code of Menlo Park?
“””,
model = LLAMA2_70B_CHAT,
)
# “{‘zip_code’: 94025}”
What not to do
The guide also advises against seeking very specific facts as it may hallucinate, aka confidently giving the wrong answer.
So asking for a list of capital cities is fine, but asking for the temperature in a specific place on a specific date or time will not produce accurate results.
It also cannot retrieve private information, of course, and it isn’t great at performing calculations.
Want to delve further? Meta has released a short course, available for free on DeepLearning.AI, taught by Amit Sangani.
Ready to use your refined prompt skills in a new environment? Check out the Alternative Credit Investor Job Board to see thousands of jobs all across the UK, like the three below.
AI Leader, Financial Services, Oracle, United Kingdom
Some of the world’s leading AI companies like NVIDIA, Uber, xA, Zoom and Microsoft have selected Oracle AI infrastructure to deliver innovative services to their customers. Now Oracle is seeking an AI Sales Leader for Enterprises, who will play a critical role in identifying, developing, and closing AI business customers in financial services. You’ll need at least ten years’ experience selling tech platforms and infrastructure solutions in a cloud provider, and five years’ experience selling cloud or software services in the financial and pharma sectors. Check out the job spec here.
AI Content Writer, DataAnnotation, Remote
This AI Content Writer position is a full-time or part-time remote role, where you’ll be able to choose which projects you want to work on, and you can work on your own schedule. You should be curious, detail-oriented and eager to teach AI chatbots. You will have conversations with chatbots in order to measure their progress, as well as write novel conversations in order to teach them what to say. You will need to come up with diverse conversations, write high-quality answers, compare the performance of different AI models, as well as research and fact-check AI responses. Find out more here.
Generative AI Designer, Fanatics Inc., Manchester
Do you have experience prompting gen AI to create visuals? Fanatics, based in Manchester, is hiring a Generative AI Designer to leverage the capabilities of AI to elevate its design process, and to produce captivating visuals, characters, stories and videos. In this role, you’ll utilise a variety of tools, including Midjourney, Adobe and other software to create finished images, and you’ll combine traditional design techniques with gen AI approaches, and incorporate CGI to push the limits. A degree in graphic design/illustration, AI or a related field is required, as is excellent prompting skills, and proven experience working with gen AI tools. Apply here.
This article was written by Amanda Kavanagh at Amply.
