• The Signal
  • Posts
  • Sunday Signal: Anthropic announces Claude-3, lessons from Peter Thiel and mental drag

Sunday Signal: Anthropic announces Claude-3, lessons from Peter Thiel and mental drag

Author

Alex Banks
March 10, 2024

Hey friends 👋 Happy Sunday.

Here’s your weekly dose of AI and introspection.

AI Highlights

On Monday, Anthropic announced Claude-3, their next generation of AI models. The family includes Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus, with each successive model offering increasingly higher performance. The most powerful model ‘Opus’ outperforms GPT-4 across industry benchmarks.

Alex’s take: I was impressed with Claude 3 Opus’ recall capability with the 'Needle In A Haystack' (NIAH) evaluation. This involves putting a target sentence (needle) inside a large corpus of text (haystack) and asking the model a question only the ‘needle’ answers. Not only did Opus achieve near-perfect recall (surpassing 99% ), but it recognised the ‘needle’ was so out of place it had to be artificially inserted by a human to test its recall ability.

This tweet by Alex Albert, a prompt engineer at Anthropic, breaks it down really nicely.

Elon Musk sued OpenAI, claiming the company breached their founding agreement to develop and share Artificial General Intelligence (AGI) technology for the benefit of humanity, not for profit.

Alex’s take: As Elon highlights in this video, over the last 8 years OpenAI has moved from an open-source non-profit AI lab to a closed-source for-profit corporation. I’m curious to see what’s next, especially after OpenAI’s response intending to dismiss all of Elon’s claims.

Inflection-2.5 approaches GPT-4’s performance, but used only 40% of the amount of compute for training. The new model is now incorporated inside their ‘Pi’ chatbot, an AI designed to be personal and empathetic.

Alex’s take: At the moment I don’t think it’s clear whether the gap that remains to GPT-4’s performance was worth only using 40% of the compute power. Regardless, it’s great to see the GPT-4 competition heat up this week alongside the ranks of Claude-3.

1 Article I Enjoyed

Joe Lonsdale is the managing partner at 8VC, a US-based venture capital firm that manages over $6 billion in capital.

On 11th April 2010, he posted a blog article titled ‘Lessons from Peter Thiel’. It captures the key learnings Joe gained from working alongside Peter Thiel, dating back to co-founding Palantir together in 2003.

My favourite takeaways:

  • Don’t divide your attention: focusing on one thing yields increasing returns for each unit of effort. An extra hour of focus on the current project has a much higher return than an hour on something new. Understand the opportunity cost of your effort.

  • The key in hiring is to value potential skills rather than currently existing skills — and potential skills are based on intelligence rather than training.

  • Return to first principles and act quickly on new conclusions. Just like questioning your assumptions, returning to first principles and building the argument up from scratch is a very powerful intellectual device that helps uncover options and opportunities you’d otherwise miss.

Another I think worth highlighting is the idea of seeking out intelligent people who disagree with you and being open about what you might learn from them.

Over time it’s all too easy to become set with processes and principles that you think work well.

Joe mentions that Peter constantly takes the time to listen to smart people whom he disagrees with—and this makes him extremely effective.

1 Idea I Learned

Mental drag.

I first came across the idea of ‘mental drag’ in James Clear’s article “Never Check Email Before Noon (And Other Thoughts on Doing Your Best Work)”.

The idea is simple. If your brain is constantly filled with secondary tasks like those 12 overhanging emails sitting in your inbox, how much memory do you have left to do meaningful, creative work?

There are a few ideas James highlights that I feel are worth mentioning to help nail your creative work.

Firstly, manage your energy, not your time. Energy is often highest in the mornings so prioritise proactive meaningful work then. In the afternoons, especially after the post-lunch slump, I find myself better suited to reactive secondary tasks like email and message processing.

Secondly, leave your phone in another room. It’s much easier to do focused work when you don’t have the pings of messages and notifications dragging you away. A close substitute for this is the do-not-disturb + airplane mode combination. (I’ve recently found setting a timer with a label of [insert task name here] to be a great visual cue when doing a block of focused work. If ever your mind wanders, you can cast your gaze to remind yourself of what really matters during the next 60 minutes.)

Thirdly, learn to eat the frog. Do the most important thing first each day before the busyness of the world encroaches on your focus.

Something I’d add to this list: use one monitor. In my past life working in finance, I was rolling with 2-3 monitors. After months of attention competition, the simple truth emerged. Humans can only focus on one thing at a time.

The ‘Ctrl + tab’, ‘Ctrl + arrow’ or famous 3-finger swipe on the Mac allows us to quickly switch between multiple windows and solve the flawed adage that ‘more monitors = more productivity’. In fact, I believe the number of monitors you have is inversely correlated to your productivity (and your ability to beat the market).

1 Quote to Share

Max Levchin on becoming a better leader:

“For me, leadership comes down to believing so strongly in the future you want that your enthusiasm, confidence, and drive in getting there infects those around you and the reality of today transforms into that vision of tomorrow through shared work and sacrifice… People ultimately want to feel a sense of agency: a common goal, a common result, and an epic effort to achieve something. If they are learning and the challenges are interesting, chances are that they're not going to leave.”

1 Question to Ponder

Who do you think will build the best chatbot by the end of 2024?

I like to keep my eye on the Chatbot arena leaderboard, a crowdsourced open platform that’s collected over 300,000 human preference votes to evaluate LLMs with the Elo ranking system.

💡 If you enjoyed this issue, share it with a friend.

See you next week,

Alex Banks

P.S. A digital archive of all Buffett's letters to shareholders.