The Neuron
Posts
😸 How to fine-tune GPT-4o

😸 How to fine-tune GPT-4o

PLUS: OpenAI landed its first gov client?!

Grant Harvey
August 21, 2024

Welcome, humans.

A ginger cat named Chubby has been going viral lately. From battling poverty to getting arrested for shoplifting, Chubby's heartrending stories have racked up millions of views… all to the tune of a cat-themed cover of a Billie Eilish song (call it a “meow”-mix!)

If you ask us, that orange cat looks a LITTLE too familiar…

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

OpenAI launched the ability to fine-tune GPT-4o (with free tokens!).
USAID adopted ChatGPT Enterprise to be more efficient.
NVIDIA created AIs to forecast the weather and be game NPCs.
Virginia's water usage surged due to AI-driven data centers.

You can now fine-tune OpenAI’s best model—here’s how.

Big news: OpenAI launched fine-tuning for GPT-4o, which means anyone can customize OpenAI’s best AI model for their specific use cases.

You know the old adage, “Give a man a fish, you feed him for a day; teach a man to fish, you feed him for a lifetime?”

Fine-tuning is like teaching ChatGPT how you like your fish, so it cooks it the same way every time!

Between now and September 23, OpenAI is offering 1M free training tokens per day (which means you can train GPT-4o on ~750,000 words), so it’s basically the best time for anyone to give fine-tuning a try—even you, non-coders!

Here's how to do it:

Go to the OpenAI Fine Tune Dashboard.
Select “gpt-4o-2024-08-06” as your Base Model.
Upload your training data—think of “training data” as example chats.

Now, you can’t just upload a bunch of text or PDFs for ChatGPT to use in its responses (if you wanna do that, try creating a CustomGPT instead.)

For ChatGPT to understand, you need to put your data in a format called JSON—think of this like a recipe that GPT can follow to prepare your dish.

If you’ve never coded before, this looks intimidating, but it’s actually really simple. Each example needs:

The “system” message (telling GPT its role).
The “user” message (the “prompt”).
The “assistant” message (how GPT will respond).

Here’s what it looks like all together:

{"messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "What's the capital of France?"}, {"role": "assistant", "content": "The capital of France is Paris."}]}

Make sure each example is on its own line, and follows the correct JSON format.

You need 10 examples minimum, but ~50-100 is recommended for better results. Keeping with the fish metaphor, it’s like showing ChatGPT a bunch of different fish, all cooked w/ the same “finesse.”

Now, upload your data, give your model a suffix (e.g.: Neuron-joke-bot), leave everything else on auto, and click “create.”

Once you’re done, go to the OpenAI Playground and chat with it to try it out!

P.S.: Want a video tutorial? This one makes the whole process super simple.

What’d you think—should we do more tutorials like this?

FROM OUR PARTNERS

🔒Learn how to unlock GenAI's power without risking your business.

Protecting your data shouldn't mean missing the GenAI revolution—but while you're grappling with compliance concerns, your competitors are launching super-valuable GenAI apps.

Ready to do both?

The team at OctoAI has gathered a panel of industry experts to share insider tips and help you innovate securely and efficiently.

Sign up to join the free round table session on August 27th, and you’ll learn how to navigate:

Data security.
Private deployment.
Regulatory compliance.
Quality evaluations.

…And so much more!

Sign up now to secure your place. Can't make it live? No worries—when you register, you still get the recording. Register here.

Around the Horn.

OpenAI has a new customer in the federal government—USAID, who will use ChatGPT enterprise to “reduce administrative burdens.”
NVIDIA created an AI called StormCast that can forecast thunderstorms and demoed a new AI to make video game NPCs respond more naturally to players.
Water usage in Virginia, known as “data center alley”, jumped up by 60%+ over the last 4 years (1.85B gallons!) largely due to AI.

Treats To Try.

Dell’s Precision AI-ready workstations

*Dell's Precision AI-ready workstations powered by NVIDIA RTX™ GPUs are the most efficient way to handle demanding AI workloads without costing a fortune. Check them out for yourself here.
Luma Labs launched Dream Machine 1.5 with higher quality image to video and text to video that better understands your prompts, plus custom text rendering.
Eppo accelerates and simplifies A/B testing and feature management for organizations (raised $47.5M)
STORI creates and manages personalized marketing strategies for branding, content creation, and social media tasks.
Evidently evaluates, tests, and monitors AI-products, offering tools for LLM observability, ML model assessment, and data pipeline quality control.
Molypix generates customizable designs for various occasions using AI, offering templates for posters, invitations, postcards, and cards.
See our top 51 AI Tools for Business here!

*This is sponsored content. Advertise in The Neuron here.

What gives it away that this is AI?

One look, everyone knows the rules:

What gives it away that this video is AI?

Here’s what #Neuron readers had to say about last week’s videos…

The Otter video:

C.H: “One of its paws suddenly grows a human finger towards the end.”
R.B: “Otters never travel alone. Plus the phone screen is wonky, evidenced by the left-side edge.”
C.D: “Otters are more likely to use android than apple tech.”

The party animal:

T.C: “Terrible fake shadow that follows the young man. It does not match the shadow behind him or the one on the wall.”
L.S: “the hat seems to be floating independent of the head.”
S.A: “meme is too old, timelines don’t match.” 😂