Find The Right Questions

Hey everyone,

This is your Sunday Space, where I serve up the best ideas, tools and resources I’ve found each week as we explore the technology shaping the future.

If you find something thought-provoking, forward it to a friend.

IDEAS
Find The Right Questions

The common belief is that the smartest and most successful people on Earth have all the answers.

This couldn’t be further from the truth.

Successful people know the pursuit of answers alone is folly. Instead, they optimise for asking the right questions.

As more and more of the world’s data gets sucked up into LLMs, making the dissemination of information even more instantaneous, one’s edge lies in their ability to formulate the questions that will move them toward the truth.

Therefore, don’t seek the answers for how to solve the world’s problems.

Focus on asking the right questions.

(Click here to share this idea on X/Twitter.)

RESEARCH
Llama 3.1

Meta released Llama 3.1 405B last week, an open-source model trained on over 15T tokens using more than 16K H100 GPUs.

Amazingly, this is the first open-source foundational model that is competitive with leading closed-source models like GPT-4o and Claude 3.5 Sonnet.

It’s clear that companies like OpenAI, Anthropic, and Meta building these models are exhausting new and novel internet data sources to train on. So, it’s no wonder the performance gap between open- and closed-source models continues to narrow.

Llama 3.1 405B can be accessed for as low as $3 per 1M output tokens—5x cheaper than GPT-4o and Claude 3.5 Sonnet.

And, when we compare LLMs’ quality versus price, Llama 3.1 405B rapidly approaches the most attractive quadrant.

Luca’s take: Meta’s Llama models are a boon for developers, who can now build GenAI apps using open models that are cheaper, more transparent, customisable, and as capable as the top closed-source models.

Llama 3.1 405B may even threaten the viability of companies like OpenAI, Anthropic, and xAI. As the quality of transformer models plateaus, these companies must develop new and innovative architectures to stay in the game.

(Click here to share this on X/Twitter.)

AI Word of the Day

Mixture of Experts (MoE)

This is an advanced machine learning technique in which multiple specialized neural networks (experts) are combined to solve complex tasks. Each expert focuses on a specific aspect of the problem, and a gating mechanism determines which expert(s) to use for a given input.

MoE models are becoming increasingly popular in large language models, like Mistral’s Mixtral, as they can improve efficiency and performance.

INSIGHTS
1 Article

China Is Closing the A.I. Gap With the United States

Chinese AI companies are rapidly closing the gap with the United States by releasing advanced, open-source models and applications (from video models to open-weight LLMs) accessible to consumers and developers. This is accelerating AI development and deployment despite regulatory and resource challenges.

Luca’s take: Companies like 01.AI, Kuaishou, and Alibaba's approach could potentially position China as a leader in AI, especially if US regulations stifle American open-source projects and allow Chinese innovation to set global standards. If we aren’t careful, the world may be forced to build the next generation of AI applications using China’s technology.

1 Post

1 Video

THOUGHTS
Quote I’m Pondering

h/t to Tim Ferriss for this incredible passage on not losing sight of the bigger picture:

“The more frequently you look at data, the more noise you are disproportionally likely to get (rather than the valuable part, called the signal); hence the higher the noise-to-signal ratio.​

And there is a confusion which is not psychological at all, but inherent in the data itself. Say you look at information on a yearly basis, for stock prices, or the fertilizer sales of your father-in-law’s factory, or inflation numbers in Vladivostok. Assume further that for what you are observing, at a yearly frequency, the ratio of signal to noise is about one to one (half noise, half signal)—this means that about half the changes are real improvements or degradations, the other half come from randomness. This ratio is what you get from yearly observations.

But if you look at the very same data on a daily basis, the composition would change to 95 percent noise, 5 percent signal. And if you observe data on an hourly basis, as people immersed in the news and market price variations do, the split becomes 99.5 percent noise to 0.5 percent signal. That is two hundred times more noise than signal—which is why anyone who listens to news (except when very, very significant events take place) is one step below sucker.”

​— Nassim Nicholas Taleb, Antifragile

(Click here to share this on X/Twitter.)

Was this email forwarded to you? If you liked it, sign up here.

If you loved it, forward it along and share the love.

Thanks for reading,

— Luca

P.S. Life is hard…