Roundup #15:
GLP-1, some Trump stuff, an important parenting hack, managing expectations, AI reasoning and alignment, and the details of AI use in 'judgement rich' domains
What caught my eye:
GLP-1 Agonists (Ozempic) are one of the most revolutionary drugs ever created. Highly effective, safe and affordable. Good news about it keeps flowing. For the first time in 40 years, the number of obese people in the US has declined, and it seems that households where one person is on the drug are spending significantly less on unhealthy snacks. This drug has the potential to seriously impact the future positively: Lower demand for shitty foods, less obese people, and likely lower future healthcare costs.
It’s very sad to me that even in the face of such overwhelming evidence that we have created something that seems to drive revolutionary health outcomes there is opposition to Ozempic. It’s coming mostly from the ‘anti-vaxxer’ corner, sure. But still it’s influential.
Dawkins’ Law of the Conservation of Difficulty:
The easier an academic field, the more it will try to preserve its difficulty by using complex jargon. Physicists use simple terms if possible, while postmodern theorists try to complexify their discipline by writing like this:
“The move from a structuralist account in which capital is understood to structure social relations in relatively homologous ways to a view of hegemony in which power relations are subject to repetition, convergence, and rearticulation brought the question …..{headache starts}”
— Judith Butler, Further Reflections on the Conversations of Our Time (1997)
From Gurwinder
Definitely applies to business topics and politics as well.
I suppose people are more and more tired of Elon Musk. He’s in the news way too often. Even before he went full MAGA I thought he had started losing the plot. But I’m still mourning the loss of one of my heroes. What he has done with SpaceX and Tesla is deeply impressive, and I really like some of his thinking around engineering and company building.
So, I’ve been hate-reading some things analyzing or simply opining on his antics. Here is a short roundup of the ones I thought were good:
A good step-back analysis of what we can probably agree on that Musk believes
The pathetic billionaires club, a piece from Paul Krugman about the weird and humiliating way Musk, Bezos, and Zuckerberg ‘bend the knee’ for Trump
A very interesting telling of the falling out of Sam Harris with Elon Musk, by Harris. The pivotal point seems to be that Musk lost a bet with Harris about the number of Covid cases. Elon bet a 1,000 to 1 that there wouldn’t be more than 35k cases. Then when there were 600k cases and 35k deaths, Harris texted Musk: Is (35,000 deaths + 600,000 cases) > 35,000 cases? After that, Elon broke contact and started attacking Harris on Twitter.
And one last one about the unbearable cynicism of Trump and MAGA
Original thoughts and ideas
AI models have an internal monologue now. It’s weird and kind of charming to see the new Chinese ‘reasoning’ AI model Deepseek R1 think out loud before answering.
After playing around with it for stuff like this, and trying it for some coding problems over the weekend, I’ve had the following thoughts:
So, you mean to tell me that the breakthrough to get from GPT-4 level to the next level1 is to give LLMs an internal monologue? That’s kind of cute, no?
5-10% of people apparently do not have an internal monologue, as someone with an internal monologue I’ve always found that hard to imagine.
As someone with an internal monologue, I can understand that this ‘reasoning’ step would work improve the answer. It’s also evident that these reasoning models are not as ‘chatty’ and fun as the regular ones.
Prediction: Reasoning will just become a ‘tool call’ that the chat model uses for certain, difficult questions.
If we get Artificial Super Intelligence by following this path, one big benefit is that our super intelligence will be very human in the way it thinks. That should make alignment a bit less daunting at least.
Completely unqualified opinion on AI alignment. When I read Superintelligence, by Nick Bostrom back when it came out I got quite worried about the prospect of AI that is very different in nature to human intelligence. It seemed impossible to control, or even monitor a system that is completely foreign in the way it works. As humans, our reasoning is mostly language based, at least it seems to be. Language is the ‘framework’ we use to express and manipulate complex ideas. What if AI would use a completely different framework that is completely foreign to us. Let’s say it makes decisions about complex ideas through some quantum physics trick, or only using geometry. Something that is completely alien to us. I thought we would have basically no hope of aligning an AI like that.
But if Large Language Models are the foundation on which AGI / ASI will be built, AI thinking will follow the basic structure of language. Now of course it’s still quite possible that AI will invent an advanced language that is much more efficient than any existing one, but we can have other AIs translate that into something we can understand. It seems to me at least that, it will be easier for us to understand AIs reasoning, intents, etc if they are expressed in language and not in some totally alien way.
The devil is in the details. What’s the difference between great writing and average writing? What’s the difference between great UX and average UX? A great coach and an average coach?
There is much speculation about the next OpenAI models, GPT-o3 and o3-mini, and there’s a lot of hype around Deepseek R1 right now as well. They’ve crushed all kinds of benchmarks in getting very hard questions right. They do very well on tests. On the one hand, there is the strong suspicion that they are at least to some degree trained to do well on tests. Let’s be honest here, when I got top marks on my ‘Risk and Derivatives’ course in university, did I really understand all of the math? Or did I practice a sufficient number of prior year’s tests to generate correct answers to variations of the same questions?
Nevertheless, it is quite clear that in a year from now we will have generally available AI that is better at ‘judgement poor’ domains like math, coding and teaching (among other things) than most human experts in the same domains. But in those ‘judgement rich’ domains like creative writing, UX, and coaching it remains to be seen how these new models will do. I had been thinking about the recent onslaught of super hyped up AI news, like Sam Altman’s annouced briefing to the US government about super advanced agents.
What will they be useful for, and what will they be bad at? One attribute of current models that is likely to carry forward is their ‘jagged frontier’. They are remarkably, even mind blowingly good at some things, while failing in stupid ways at other, very simple things (e.g. the famous count the Rs in strawberry thing).
Other non-AI things
Babies are remarkably chilled out considering that:
They have no idea what’s going on
They can’t really communicate other than screaming and facial expressions
They have pretty much no control over anything that happens to them
One hack to save your back as a parent. Almost everyone gets back pain from time to time as a parent from lifting up kids in awkward angles. Many, many years ago I did one summer of Crossfit and learned to ‘brace’ when lifting weights. This means tightening the muscles in your entire core for as long as you lift a weight. This has proven for me again and again to eliminate all back pain from lifting kids. Obviously not all back pain is due to life style factors, but who knows how much actually is? Please try this if you have any back issues at all (and let me know how it goes).
Managing your own expectations. Again and again I fall into the trap of escalating expectations. This time, with Magicdoor, my idle, unrealistic hope of breakthrough success put me in a position to feel a tinge of disappointment at the 120% month-over-month growth in month two.
Then I remembered that: doubling in a month is amazing, this is only the first thing I launched, it is profitable from day one, nobody is forcing me to keep working on this idea. I can also chose to keep it running and try to grow it while launching the next idea.
I’ve never been a fan of the ‘underpromise, overdeliver’ advice, because I feel like too often it ends up being ‘underpromise, deliver’, or even ‘underpromise, underdeliver’. There are benefits to overpromising, i.e. Elon Musk got a lot of insane things done despite always vastly underdelivering compared to what he promised.
But internally, I personally seem to get consistently better results by underpromising to myself.
I got into a lasting running habit by going (much) slower so that I can listen to a podcast while running and not be so damn uncomfortable the whole time
Got into a lasting writing habit by abandoning all goals and expectations regarding its frequency and impact
Level is a bit vague here. But it seems like around every 18 months a new level is reached. From GPT 3.5 to GPT 4, and now to ‘reasoning’. Each does 3-5x ‘better’ on ‘stuff’. Another thing that seems to be true so far is that the ‘mini’ version of the new level just outperforms the ‘full’ version of the previous level by a little bit.