The Neuron
Posts
😺 Code Llama unleashed!

😺 Code Llama unleashed!

PLUS: Guide on how to make ChatGPT sound like you.

Noah Edelman & Pete Huang
August 25, 2023

Welcome, humans.

Fun announcement coming next week! Some call us the gift that keeps giving … right up there with Taylor Swift, HBO Max, and Trader Joe’s!

Here’s what you need to know about AI today:

Meta’s Code Llama rivals ChatGPT-3.5 for coding.
ChatGPT can sound like your brand using Custom Instructions.
Ideogram wants to take on Midjourney.
OpenAI is partnering with Scale AI to help companies fine-tune GPT-3.5!

Meta’s New “Code LLaMA” Codes As Well As GPT-3.5.

Meta unveiled its ChatGPT coding competitor: Code Llama (built atop Llama 2)! It’s available in three flavors and three sizes (7B, 13B, & 43B).

Like ChatGPT, Code Llama writes and edits code from text prompts. Here’s how Code Llama stacks up on the well-known coding test HumanEval Benchmark:

ChatGPT-4 — 67.0.
CodeLlama-python — 53.7
ChatGPT-3.5 — 48.1

Code Llama crushes Google’s PaLM, Codex, and other open-source models.

Some quick insights on the model (non-technical folks ... don't you dare think about skipping this!):

It’s open-source, which means anyone can see, use, and build on the model as if it were legos. Request access here.
Can run locally on the 7B model. In layman’s terms, it can run on your laptop without a cloud API & your data stays private (use llama.cpp).
Good for autocompletion and small code snippets (on par with Github Copilot).
100k token context limit, but retrieval falters after 16k tokens, which means Code Llama can manage ~10,000 lines of code.
Worse than ChatGPT-4 in most ways, but more budget-friendly.

Big day for Meta. Big day for open-source and AI. And biggest day of all for CS making their way back to university.

BTW: here’s a great prompt to use every time you code. Enjoy and build something extraordinary this weekend!

FROM OUR PARTNERS

Discover Advanced Prompting Strategies Used By AI Experts.

If you’re like 95% of professionals, you’ve merely been dabbling with ChatGPT and other AI tools.

The other 5% (ourselves included) are truly using AI to automate entire workflows at work, saving valuable time.

If you want to join us, take this Generative AI Masterclass. AI experts Jeremy and Lev will show you exactly how to integrate AI into your work routine through hands-on prototyping and proven strategies.

The course also includes advanced prompting strategies and 1:1 coaching sessions!

Enroll in this AI masterclass on Maven before signup closes on September 12th!

How To Get ChatGPT To “Talk” Like Your Brand.

Zapier has been churning out hundreds of articles every month … and they all sound exactly like Zapier.

Their secret? Uploading style guides to a chatbot.

Here’s the secret sauce for replicating Zapier’s success with everything from press releases to social posts:

First: write your brand’s style guide (tutorial). Must haves:

Company’s values and missions.
Brand’s voice and tone.
Who you're speaking to.
Must-follow grammar rules.
10+ examples of great content (for each content category).

Second: upload your style guide to ChatGPT custom instructions:

Go to ChatGPT,
Custom instructions
Answer box #1 with everything brand-related
Answer box #2 with everything writing-related.
Toggle on “Enable for new chats”.
Save it all up.

You can now instruct ChatGPT to write something fresh or feed it a content draft to finesse in your voice.

Around the Horn.

Kevin Roose: How schools can survive (and maybe even thrive) with AI this fall.
How to build a chatbot on top of GPT-3.5 with custom data sources using LLamaIndex.
OpenAI is partnering with Scale AI to assist enterprises in fine-tuning GPT-3.5.
Legal experts tested language models on 162 tasks in legal reasoning (research paper).
A cat-themed, AI-generated video created by a TV commercial director? Count us in.