o1 "think through" problems before providing solutions. o1 is not GPT. It excels in tasks requiring planning and iteration. o1 doesn't surpass GPT-4o in writing creativity. o1 successfully solved the "river crossing" riddle and a crossword puzzle. This unlocks a new paradigm of model pre-training. o1 is great in many ways but it isn't superior in all areas.

NVIDIA developed a method to create a smaller and accurate LLM known as Llama-3.1-Minitron 4B using structured weight pruning and knowledge distillation.

This is huge! Prompt caching enables you to load vast amounts of data into the context window. This will unlock a wide range of new use cases. I'm so pumped.

Meta flagship model, Llama 3.1 405B is the first open weights model that is competitive with the top foundation models across a range of tasks, including GPT-4, GPT-4o, and Claude 3.5 Sonnet. The smaller models, 8B and 70B are competitive with Gemma 2 9B, Mistral 7B, Mixtral 8x22B, and GPT 3.5. The upgraded versions are more capable, have a longer context length of 128k, multilingual, and advanced tool use.

8B gets a big bump across the board, 70B instruct shows minor improvements, and 405B is the SoTA open model. But 405B still lags behind flagship models.

📚 I recently finished "Co-Intelligence" by Ethan Mollick. Is it the Gen AI guidebook you've been waiting for? 🚀 It's all about the sociological perspective of living and working with AI. A solid primer to AI for most readers, but not suited for those with advanced AI knowledge.

This evaluation provides insights into the current capabilities of leading AI models in solving complex coding problems. While Claude 3.5 Sonnet showed superior performance in this specific task, all models demonstrated the ability to produce correct solutions with varying degrees of assistance. These findings underscore the importance of conducting independent evaluations to verify public benchmarks and understand the nuanced strengths and limitations of different AI models.

Anthropic Claude 3.5 Sonnet takes the top spot on the leaderboards. It surpasses GPT-4o.

For this round of book therapy, I will be reading two books, "Experts vs. Imitators" and "Just Enough Software Architecture".

The importance of focusing on foundational principles and high-impact work in technology.

I argue for the importance of learning the C programming language despite its lack of trendiness in modern software development.

It sucks to be afraid of the tools that you use in your work every day.

A chapter by chapter detailed summary of the book.

AI research shows that leveraging computation through general methods like search and learning is far more effective than incorporating human knowledge.

Vibe checking the latest Google Gemini models by asking it about fine-tuning Transformer tools.

Signs that we're living in the assembly language period of LLMs and Generative AI.

Tech dive into GPT-4o covering techniques and data, system, and tricks.

The new advanced voice and video mode is groundbreaking! Audio AI models are having their Stable Diffusion moment too.

Make computers talk like human. My attempts and some notes about them, along the way.

Augmentation is composed of smaller automations. AI helps generate levels of abstraction.