Early Days, NeuroHacking, Klarman, Fast

Early Days

We are still in the earliest days of generative AI tools. Much of what has been released to date is a proof of concept. One neat emergent behavior of generative AI tools is that they seem to behave more like people and less like computers or calculators. Generative AI, so far, is lacking precision and accuracy, but it thrives when it comes to learning, creativity and best-efforts attempts at new tasks.

For example, see DeepMind’s latest robotics blog post. DeepMind’s new robotic agent model RoboCat allows a robotic arm to “watch” any robot or even human hand performing a task and then attempt to recreate the desired end state. In the testing environment, it was able to arrange gears, blocks and fruit with increasing levels of accuracy after each attempt. You can watch the video here of the various examples.

It’s important to note that the model improved with each training, just like an employee, student or athlete improves with practice. Coding robotics to do specific tasks is expensive and time-consuming. An industrial-grade robotic arm called a ‘cobot’ from Universal Robotics costs between $10-50k. When any business owner can buy and train a cobot by showing them the end state goals and letting it learn to improve over time, adding a robot to a business process becomes a reasonably inexpensive and simple task. Google owns Deepmind, and hopefully, they have plans to release these models for public use via their Google Cloud services. I’m sure many customers would pay for access to the models.

Most businesses today have AI goals and strategies, but few have even started implementing their vision. We are entering a phase where business owners and executives will spend more time supervising and training AI agents than doing the actual work themselves.

Neuro-Hacking

Using neural data and machine learning, researchers were able to build a prediction model that can identify a hit song with fairly high accuracy. Thanks to human biology hacks, we already have many examples of foods that are unnatural and addictive, like McDonalds, Doritos, Diet Coke, Oreos, and many others. TikTok’s success is also based on known hacks that are highly effective at capturing and retaining attention.

TikTok has already impacted music in a major way by promoting shorter songs with 30-second hooks that resonate. Fortunately for the arts in general, these music models won’t be able to adjust for niche preferences or tastes that change over time. There is still plenty of room for organic creativity.

If models can eventually predict movie reactions and preferences, that could save movie studios from expensive mistakes, as marketing is roughly 50% of the average film budget. Some studies are working on predicting video engagement. There are historical examples of algorithms that have tried to predict music success, like Polyphonic HMI, which is the subject of a relatively well-known case study. It’s unclear how much success can really be attributed to Polyphonic’s older algorithms. Expect more examples of novel AI/machine learning models used to uncover new addictive patterns.

Seth Klarman on CNBC

Seth Klarman made a rare appearance on CNBC to promote the 7th edition of Security Analysis, considered one of the essential investment books. The two clips can be viewed here and here. The book has been updated to include comments on emergent financial trends like cryptocurrencies, the dominance of FAANGs and technology and private credit, a relatively new asset class compared to the history of markets. The previous version was published in 2008.

Building Fast Is Possible

Packy McCormick had a great write-up of how Philly rebuilt the collapsed section of its I-95 bridge in just 12 days. Construction was live-streamed from the X stadium and was treated almost like a sporting event. Stripe co-founder Patrick Collison has a well-known blog post titled, Fast, that covers large-scale accomplishments that were completed quickly. This contrasts with many painfully slow and extremely over-budget public projects, like the BART train in SF and many others. The 12-day I-95 repair reminds us that we can complete public projects quickly when we want to.