My fine-tuning journey begins today

I am training llama 3 8b to create high quality datasets for financial Q&A.

Fine-tuning approach:
• create high quality datasets via gpt-4o
• fine-tune llama 3 on datasets
• evaluate performance

I am using my financial-datasets library to

Human Eval Is No Longer Reliable Because It Can Be Gamed - Mystery Behind GPT2-Chatbot Revealed.

A few months ago, people claimed that LLM benchmarks like MMLU weren't reliable because they can be gamed.

You can memorize the answers and get a high score when the model is bad.

This week, Google announced a doubling of Gemini Pro 1.5's input context window from 1 million to 2 million tokens, and OpenAI released GPT-4o, which generates tokens 2x faster and 50% cheaper than GPT-4 Turbo and natively accepts and generates multimodal tokens. I view these

It’s only been 2 days since OpenAI revealed GPT-4o.

Users are uncovering incredible capabilities that completely change how we use and interact with AI.

The 12 most impressive use cases so far:

MrDee@SOG 🫡 This is a surprisingly good response. I highly recommend you look at DQN. They are small, fast, and learn over time. They are great for small tasks where information is scarce. E.g., should you evict an element in a cache, or should you switch lanes in a car?

Now combine Q learning + A* + LLMs

Conceptual Integration

-Defining States and Actions:
States: Each token or a sequence of tokens during the generation of text.

-Actions: The choice of the next token or sequence of tokens.

-Applying Q-Learning:

Develop a reward function

Now I asked it to explain A star search.

Alright, imagine you're playing a video game where you're on a quest to find treasure. You're in a big maze with lots of paths, and you want to find the quickest way to the treasure without wandering around too much.

The A* (pronounced

Then I used this nifty plugin on chatgpt called 'Show Me' to ask it to explain it to me visually.

- The Player explores the maze and chooses actions.

- The Maze (Environment) provides the current state and reward to the player.

- The Player checks the Q-Table (Guidebook) for

Openai actually quietly published GPT4-o's capabilities in their blog openai.com/index/hello-gp…

If you go through each example and understand the implications of each, you'll realize this is a bigger deal than they painted it to be.

People who aren't impressed by GPT4-o are clouded by their judgment of the present and not what the future beholds.

I'll try to break down why GPT4-o is impressive--but not in the benchmark sense but because it shows a glimpse that scaling multimodality actually works!

We all

Imagen 3 can generate high-quality visuals in a wide range of styles- from photorealistic landscapes to richly textured oil paintings or whimsical claymation scenes. 🎨

🧵 dpmd.ai/imagen3

We’re partnering with Reddit to bring its content to ChatGPT and new products: openai.com/index/openai-a…

Today we describe a method for generating differentially private synthetic training data, which reflects the overall characteristics of the original data and enables the training of robust models while maintaining strong privacy protections. Learn more at goo.gle/3QMoJrT

Introducing LearnLM: our new family of models based on Gemini and fine-tuned for learning. LearnLM applies educational research to make our products — like Search, Gemini and YouTube — more personal, active and engaging for learners. #GoogleIO

After almost a decade, I have made the decision to leave OpenAI.  The company’s trajectory has been nothing short of miraculous, and I’m confident that OpenAI will build AGI that is both safe and beneficial under the leadership of Sam Altman, Greg Brockman, Mira Murati and now, under the

Introducing Med-Gemini, our new family of AI research models for medicine, building on Gemini's advanced capabilities. We've achieved state-of-the-art performance on a variety of benchmarks and unlocked novel applications. goo.gle/3UK7Oax #MedGemini #MedicalAI

OMG! GPT-4o is really really really good with scientific writing! This is a game changer.

I used my Deep Writer (a massive agent system which writes anything of any length, aiming for professional human quality).

This is the closest that I or anyone may have come. Results:👇🧵

TED Talks Andrej Karpathy Agrim Gupta Kyle Sargent ~10 years is a small blip in history, but a giant leap forward for the field of AI, and Computer Vision. This was my TED talk in 2015, at the dawn of Modern AI. What a decade we have had since then! If you watch both these two talks, it will give you a pretty good understanding

TED Talks Andrej Karpathy Agrim Gupta Kyle Sargent One day, Spatially Intelligent robots can even hope to help people to do tasks that they need help with. 9/ x.com/drfeifei/statu…

TED Talks Andrej Karpathy Agrim Gupta Kyle Sargent Empowering embodied intelligence with spatial intelligence and language intelligence will unlock exciting possibilities. 8/ x.com/drfeifei/statu…