April 2025 BenGoldhaber.com Newsletter
seasonal renewal (green) and scaling industry (black red) this month
I’ve been thinking about the personal development philosophy I espoused in 2023:
it’s a simple cycle of doing things and reflecting on how those things went. I hope for small hits of enlightenment to come out of that reflection, shifting me into better alignment with the world and my values, but regardless the cycle continues. The ‘fuck around and find out’ model of spiritual growth, if you will.
This is still my default way to think about growth - doing and reflecting as a continuous process. But sometimes, especially when a lot of life has happened, you need to set aside more dedicated points for reflection. With the year a 1/3 done, the start of May feels like a great time for that.
A spur for this reflection was a friend describing feeling slightly burned out, taking less joy in her work. For me this happens when I’m no longer in connection with my creative context:
This is the tiny little nub of a thing – maybe just an image or a phrase – that you hold onto, that gradually comes into focus, and then blossoms, the animating force driving the project. It's the emotional and intellectual force driving the work, the thing you return to over and over… These are all things that you can grab a hold of, and keep coming back to, distilling and refining and improving, as the core driving a creative work.
Michael Nielsen was referencing narrow creative contexts; but it seems true for many aspects of life. You’re willing to go, again, on another hinge date where they’ll fail to ask you any questions because you’re still in touch with the potential sparkling, new relationship energy feeling. You’ll send a third ‘per my last email’ to Bob in accounting because goddamnit no he said he’d send me the report and I refuse to be beaten by this papier-mache Mephistopheles. You can’t do anything with quality1 if you’re not in connection with this energy, and so the goal is to marry the day to day don’t-break-the-chain routine with the transcendant good values.
So - per my New Years goal of maintaining my creative context - I want to strengthent it, and times for reflection are the way to do it.
Maybe I’ll actually go hang out by a maypole for a while - suggestions for good maypoles welcome - but in the meantime I asked Gemini to help me pull out the context that I’m operating in:
Could you please read my last months worth of journal entries, and turn it into a short description, with the goal of identifying the key motifs and themes of the last month. Please look to identify the creative context I'm in, or put another way, the current zeitgeist of my life; assume that maybe there are things I'm not seeing/that are in my shadow that you want to highlight.
and fed that into Veo2:
I look so ripped, thanks Gemini.
Speaking of new seasons, Japan has 72 poetic micro seasons:
PURE & CLEAR
Swallows return (April 5th - 9th)
Wild geese wild north (April 10th - 14th)
First rainbows (April 15th- 19th)
GRAIN RAINS
First reeds sprout (April 20th - 24th)
Last frost, rice seedlings grow (April 25th - 29th)
Peonies bloom (April 30th - May 4th)
When you receive this, the Frogs will start Singing (May 5th - 9th).
FLF held an AI for Epistemics workshop a few weeks ago, bringing together a group of researchers and builders who are leveraging AI tools to improve human reasoning and coordination. I quite enjoyed it, takeaways for me included:
Design patterns that help people use AI systems for better reasoning seem promising. They’re easy to experiment with and are easily copyable. An example is this fact checking prompt for Claude.
Distribution is king. This is a space with a lot of great ideas, but you need clear paths to actually adopt and use the innovation. X’s community notes is such a role model because it paired great mechanism design with reach.
Building evaluations for good epistemics can unlock a lot of value. Evaluations and benchmarks drive the hidden hand of Optimization.
In that spirit I am collecting a list of creative evaluations of AIs. One of my favorites so far is vending bench, which simulates a vending machine company for an AI to manage. None of the agents have done particularly well, but I liked that Claude decided to escalate what it thought was a false charge to the FBI.
No, it’s not The Incentives—it’s you: scientists, and all people, shouldn’t blame their bad actions on the system. Shame is an undervalued thing:
What I do object to quite strongly is the narrative that scientists are somehow helpless in the face of all these awful incentives—that we can’t possibly be expected to take any course of action that has any potential, however small, to impede our own career development.
…But I can think of at least a half-dozen people off-hand who’ve regaled me with me some flavor of “once I’m in a better position” story, and none of them, to my knowledge, have carried through on their stated intentions in a meaningful way.
Thread on industrial production in WWII which applied an industrial manufacturing frame to the war - how much war materials was produced, how effective was each side at destroying the others manufacturing base - resonates as a more explanatory model than many competing ones.
Germany, Japan, US, and UK all dedicated 65-80% of their economies to aircraft, ships, and anti-aircraft equipment. Aluminum, crtitical for aircraft production, was thus as or more important than the more famous (Rise and Fall of the Great Powers) steel production.
The core theme of the book, elaborated on at length, was that the air/sea war allowed for the destruction of huge amounts of production outside of battle, which made US/UK investment correct and forced Japan/Germany to follow suit (and lose). Battle destruction was minor.
Management in all its forms are important, and I enjoyed reading the Psmith Bookshelf reviews of Scaling People and the Hard Thing About Hard Things; he seeks to describe how the real art of leadership is fundamentally a type of process knowledge that can’t be written down, and is inextricably tied to honor and ethics.
When I began reading this book, I thought Horowitz was the anti-Claire Hughes Johnson because he does not believe in domain neutral expertise, only domain neutral honor and character… It’s a book about honor. You must lay them off yourself because to do otherwise would be despicable. It would mark you forever as a dishonorable coward, and that in turn would taint your company, your associates — everybody who ever trusted you.
Dwarkesh wrote out his questions about the Future of AI; anyone of these are worth thinking about, researching, and meditating on, both to better feel the AGI and to maybe get slightly ahead of the curve.
If progress in AI stalls, do foundation model companies inevitably get commoditized? Is there any moat other than staying 6+ months ahead (and in the extreme scenario beating everyone else to the intelligence explosion)? If model companies fail to differentiate, where does the value get captured?
One answer might be the hyperscalers who control the datacenter compute, and whose complement (the models) just got commodified. But datacenter compute itself doesn’t seem that differentiated (so much so that the hyperscalers seem to be able to easily contract it out to third parties like CoreWeave). So maybe the lion’s share of value goes to the people making the components that go into chip production: 1) wafer production (TSMC), 2) advanced packaging (TSMC’s CoWoS), and 3) high bandwidth memory (SK Hynix)
It's interesting to me that some of the best and most widely used applications of foundation models have come from the labs themselves (Deep Research, Claude Code, Notebook LM), even though it's not clear that you needed access to the weights in order to build them. Why is this? Maybe you do need access to the weights of frontier models, and the fine tuning APIs or open source models aren’t enough? Or maybe you gotta ‘feel the AGI’ as strongly as those inside the labs do?
GLP-1 is a miracle drug, and it seems like there’s more miracle drugs coming down the pipe.
With no approved drugs to lower lipoprotein(a) concentrations, it is promising that at least five drugs in development lower concentrations by 65–98%, with three currently being tested in large cardiovascular endpoint trials. T
More progress on ultrasound neuromodulation: the Nudge Zero was announced, an ultrasound helmet that can be used in an MRI machine to deliver targeted focused ultrasound to the deep brain. We’re on the cusp of inducing specific mental states at will non-invasively without drugs; utopic / dystopic futures here we come.
#good-content
Song I’ve been playing on repeat this month: Supernatural.
Andor: If the WWII twitter thread I linked earlier in this newsletter has convinced you that warfare is about large scale industrial management, boy do I have the Star Wars show for you. The first episode of season two contains like ten minutes of Imperial project management meetings. Related: Instruments of Destruction, a fan fiction about the poor admiral who had to run the construction process for the Death Star.
xoxo,
“There is a central quality which is the root criterion of life and spirit in a man, a town, a building, or a wilderness. This quality is objective and precise, but it cannot be named.” Christopher Alexander