5 stories that caught my eye this week and what I took from them.
Which AI should I use? Superpowers and the State of Play – Ethan Mollick
Which AI should I use? Superpowers and the State of Play (oneusefulthing.org)
Perfect primer on the state of play in AI models
3 different but similar models: GPT-4, Claude 3, Gemini. Most likely next change to this assessment: release of GPT-5 later this year
What is happening at the frontier, what to expect next:
Bigger context windows. This should reduce hallucinations and give better answers with the ability to reason across giant sets of data. RAG would then become less important.
Agents. Devin, the AI software engineer, is a first glimpse at this. Imagine having a junior analyst work on a project for you. It will work out a plan how to tackle the project and then execute each step. You then give feedback and direction to refine the output.
Top tip: spend the 10 hours you need with any frontier model to learn what they do and how they can help. ‘Use AI for everything’ you legally and ethically can, to see what it can do.
Bottom line: The current frontier models are all pretty similar. Pick one and use it. Larger context windows and agents are the next frontier…exciting!
Biden 2.0 – Ezra Klein (podcast)
- Opinion | What a Second Joe Biden Term Would Look Like - The New York Times (nytimes.com)
- It’s fair to focus on the challenger, but it’s also important to understand what Biden 2025 could look like from a market perspective.
- The theme seems to be ‘Finish the Job’, which sounds reassuringly boring from a market perspective.
- The IRA, Chips & Science Act and Infrastructure Bill have passed, but there is a lot of work to do on the implementation side, to make them successful and to prevent significant roll-back under future administrations.
- Plenty of culture war issues to work on (e.g. abortion, immigration), but these will have little impact on markets.
- No significant new legislation likely to be passed. 1) this is typical in second terms and 2) it would take a big blue wave to give Democrats control of both the House and the Senate. Without that, expect the same gridlock of the past 2 years.
- This also means that Biden’s tax ideas are unlikely to become reality.
- More important will be the negotiations around the tax cuts in 2017’s Tax Cuts and Jobs Act, which are due to sunset at the end of next year. The election will have an impact on balance of power in those negotiations, but neither side have shown any appetite for fiscal austerity in the past.
- Big picture, the focus would gradually shift towards Biden’s legacy
- It’s likely that a second term would be characterised not by the policy agenda, but by macro events: geo-politics or recession.
- Bottom line: Biden 2.0 = gridlock = politics are a sideshow for markets = good news
Farmland During Recessions – Artem Milinchuk
- Farmland During Recessions | Portfolio for the Future | CAIA
- As an asset allocator you’re always looking for ways to add diversification and to find uncorrelated returns. Farmland has been one of those assets.
- The numbers speak for themselves (1992-2021): 10.7% annual return, only 6.8% volatility and low correlation with many other assets.
- Even in a soft landing there is little debate that the economy is ‘late cycle’. Late cycle can last several years and has historically been a good time for equities, but it’s also a time to think about what’s next. Eventually the recession forecasts will be correct.
- Farmland has done well in the last four recessions, but this paper gives a lot of colour on farmland’s behaviour in those recessions.
- One thing that stood out was how differently it behaved from real estate. There is a connection, but farmland has many fundamental drivers that are different.
- Caveat: Are 4 recessions a big enough sample to have confidence for the next recession?
- Also interesting: Farmland acreage has been in a steady decline in the US, down over 20% since the mid-1960s.
- Everything has a Tech angle. Deere has a partnership with SpaceX, using Starlink satellites for ‘precision agriculture’, autonomous tractors and autonomous drones.
- Bottom line: Farmland has been a great place to be, especially in the last 4 recessions
Nvidia reveals Blackwell, the ‘world’s most powerful chip’ for AI – The Verge
- Nvidia reveals Blackwell B200 GPU, the ‘world’s most powerful chip’ for AI - The Verge
- 3 interesting things from Nvidia’s presentation:
- Chip progress is not slowing down. Nvidia claims Blackwell will be up to 30x faster than Hopper when it comes to AI inference tasks, and with slightly mod=re modest gains on AI model training speeds.
- Efficiency gains got a lot of attention. Nvidia claims that training a 1.8 trillion parameter model would have previously taken 15 megawatts of power. With Blackwell it would take just 4 megawatts. Dig deeper here.
- Priced in. Nvidia’s share price did not budge on the reveal and Nvidia’s shares lagged behind the Nasdaq over the past few days. A data point that suggests the much raised expectations and reality are reasonably closely matched.
ChatGPT for Music Is Here – Rolling Stone
Inside Suno AI, the Start-up Creating a ChatGPT for Music (rollingstone.com)
You’ve seen Sora, try Suno.
A new text-to-music AI with impressive results. With a bit of prompt engineering I got some acoustic country songs about the start of the baseball season that could easily have been on the radio. I can only imagine what future versions can do.
Sign up for a free account and make some music this weekend!
There will be lawsuits as elsewhere in generative AI, but by the time the lawsuits are settled, the technology will have long moved on.
Favourite (familiar) quote: “just last year, AI-music experts told Rolling Stone that a service as capable as Suno’s might take years to arrive”. Humans are and will be bad at understanding exponential growth!
Bottom line: Another week, another AI frontier pushed out.