Ranges vs. Narratives

[candide]

I wrote this piece in a cafe, where I didn't connect to wifi, as I wanted to focus on my thinking. I wanted to positively put forward what I wanted to see more of, rather than complain about what I didn’t want to see. Even if some lines look like they are referencing one person in particular, they are not. I see broad trends here, and I am treating this as a mission statement for what I will try to do instead.

======

"I would rather have questions that can't be answered than answers that can't be questioned"

-Richard Feynman

To make accurate predictions, best practice includes thinking in ranges and being able to explicate, question and adjust premises. Also, we (those trying to understand the future) get to adjust our ranges when new information comes in.

As the march of the algo recommended content marches on, "research" is going to be more and more infotainment, and that tends to rely on narrative, binaries (including ingroup/outgroup), and confidence (you know, like how "con men" show confidence). The medium is the message, my friends, and I think too many of you are letting this seep in as an acceptable methodology, if not for your own claims, then the claims of others.

While still keeping this at the level of principles, let's remember the fallacy fallacy: just because a methodology is bad doesn't mean it can't accidentally be correct (someone can win betting on red three times in a row). But even then, even if bad practice leads to a "good" answer, what is more likely is that something will be directionally correct, but fail in terms of magnitude.

And my point is that magnitude, and other aspects we can be precise on, matter. One of the great weaknesses we must guard against as humans is that adjectives over-compress reality. This is also where bubbles tend to come from: it's not that the underlying asset that is made into a bubble has no value; rather, it is that it does not have the precise amount of value implied in the narratives-- also, price starts to be an answer that cannot be questioned.

...

To try to get the ball rolling on what I am talking about, I have a research question I want to start with. How many organization are playing with the kind of compute to be in the running for AGI? I think figuring out how to know that, and how to monitor that would be interesting to uncover and useful for the kind of thinking in ranges I am talking about.

[daylen]

My p-doom is about 50%.. plus or minus 50%.

[jacob]

To try to get the ball rolling on what I am talking about, I have a research question I want to start with. How many organization are playing with the kind of compute to be in the running for AGI? I think figuring out how to know that, and how to monitor that would be interesting to uncover and useful for the kind of thinking in ranges I am talking about.

I'm not sure what you are talking about, but my answer for my best guess [in terms of ranges vs narratives vs other] would be viewtopic.php?p=303130#p303130 (considering a range for a/one variable corresponds to formal analysis)... Are these the droids you're looking for?

I wouldn't know how to begin to evaluate AGI (much less know who is pursuing it and who is not). ChatGPT3.0 already passed the Turing test for me relative to substantial fraction of the human population. It has changed how I think about intelligence. I'm no longer sure that humans are intelligent per se but rather just have brains capable of generating language to various degrees. "Intelligence" would then be an emergent quality arising from a few actually intelligent, alternatively random, events that then get elevated/selected for in the language-generating conversation much self-selected evolution. (I realize I'm simply repeating the habit of using the latest and greatest innovation to understand human consciousness. We've come some way of thinking of the brain as little steam pistons and electric wires.)

[philipreal]

To try to get the ball rolling on what I am talking about, I have a research question I want to start with. How many organization are playing with the kind of compute to be in the running for AGI? I think figuring out how to know that, and how to monitor that would be interesting to uncover and useful for the kind of thinking in ranges I am talking about.

Here's a bit of a jumping off point in terms of compute: https://epoch.ai/data-insights/computing-capacity. Generally it seems from the AI-related news that I've been reading that OpenAI (ChatGPT), Google Deepmind (Gemini) and Anthropic (Claude) are the leading AI labs, with xAI (Grok) and Deepseek a bit behind, and Meta somewhere after that.

[candide]

@Jacob. Fair points. Also, when you factor in breadth of knowledge, speed, and tirelessness, we are already dealing with super-intelligence. But to be a little more explicit, it seems like progress is capability is made through huge runs that use a ton of compute. And that just makes a small group of organizations that we can track.

Here's a bit of a jumping off point in terms of compute: https://epoch.ai/data-insights/computing-capacity. Generally it seems from the AI-related news that I've been reading that OpenAI (ChatGPT), Google Deepmind (Gemini) and Anthropic (Claude) are the leading AI labs, with xAI (Grok) and Deepseek a bit behind, and Meta somewhere after that.

Yes, that's exactly what I was looking for. Thanks. That seems like a great organization to check on articles from time to time.

This piece seems to have what for a lack of a better place to anchor, a place to put my priors:

https://epoch.ai/blog/announcing-gate

I'm just saying that if we are going to do futurism, let's try to do good futurism.

[7Wannabe5]

I think this interesting, although also fairly creepy, recent TED Talk with Eric Schmidt goes some way towards answering your first research question. He states that only organizations based in the U.S. or China currently have the necessary capital base. So, I would estimate 10. I have noticed that my bias when watching interviews such as this with Tech types is that I tend more towards believing those who are baby-faced, like Dario Amodei, than those who give off my least favorite internet-influencer/interviewee vibe which I might describe as towards "sociopathy masked by extreme civility." However, I have learned that when it comes to men (or those primarily in their masculine energy) in any circumstance, the "baby-faced" are also not to be believed. Therefore, if one isn't willing to do all the research on one's own, the character you need to find is the well-informed straight-up, self-interested, skin-in-the-game player. Humans will talk shit all the day long until they actually have to put their own money on the table.

https://www.youtube.com/watch?v=id4YRO7G0wE


In only semi-related news, scientists recently put the human NOVA-1 gene variant (believed to be responsible for humans acquiring transferable intelligence through language) in rats and this caused them to create more complex mating calls. Ergo, entirely possible that the advantage derived from the ability to sweet-talk the ladies is basis for general human intelligence.

[candide]

Mistral, though interesting, and maybe potentially trying to join in the race, is not in a competitive spot. So yeah, U.S. and China for SOTA. And realistically, only U.S. this year.

In only semi-related news, scientists recently put the human NOVA-1 gene variant (believed to be responsible for humans acquiring transferable intelligence through language) in rats and this caused them to create more complex mating calls. Ergo, entirely possible that the advantage derived from the ability to sweet-talk the ladies is basis for general human intelligence.

I've seen theories that humans did sign language before they started talking. So maybe at first it wasn't as much sweet-talk as ... dancing?

[7Wannabe5]

I've seen theories that humans did sign language before they started talking. So maybe at first it wasn't as much sweet-talk as ... dancing?

Maybe. It has been my experience that to the extent that men don't employ sweet-talk, they usually just grab you like any other animal would. However, I obviously suffer from inattentive form ADHD, so MMV significantly.

[Henry]

I think it's a false dichotomy. The first written narrative of going to the moon was in 125AD and it was a satire. So a facetious Ancient Roman motherfucker got his Apollos right if you gave his narrative an 1,844 year range. If any of his readers invested, well, as they say, being a few millenia early is being shit out of luck. It's the basic tension in investing - valuation based upon future cash flow. As far as the question of how many companies are playing with enough compute in the AGI sector, as far as a practical range is concerned, IMHO what you see is what you get. As far as narrative is concerned, you have to allow for at least the same number that are currently hidden or non-existent that will rise and eat the lunch of the ones we see.

[candide]

I think it's a false dichotomy.

Only in the sense that narratives that people refuse to discard can be expressed as ranges that they believe are 100% to 100% with a stated confidence interval of 100%.

But I'm actually drawing on a body of academic research here, and in the first post linked to a good summary of the book by Tetlock. Quote from that summary:

Superforecasters don’t rely on gut feelings or grand narratives. Instead, they use a data-driven, probabilistic approach. They avoid overconfidence and readily acknowledge uncertainty. Intuition, while valuable, is seen as a starting point, not an endpoint. Superforecasting emphasizes continuous learning and refinement through feedback.

That's just not what you're doing over in the AI-turned-Tesla thread.

I think it's a false dichotomy. The first written narrative of going to the moon was in 125AD and it was a satire. So a facetious Ancient Roman motherfucker got his Apollos right if you gave his narrative an 1,844 year range. If any of his readers invested, well, as they say, being a few millenia early is being shit out of luck. It's the basic tension in investing - valuation based upon future cash flow.

So the 125AD narrative (which, as you stated was satire, and so was only right by accident) was directionally correct, but wrong in terms of magnitude? Where have I heard that before? (I hesitate to quote myself from earlier up on the same page, but at this point I think it's the only way some people here are going to read it):

While still keeping this at the level of principles, let's remember the fallacy fallacy: just because a methodology is bad doesn't mean it can't accidentally be correct (someone can win betting on red three times in a row). But even then, even if bad practice leads to a "good" answer, what is more likely is that something will be directionally correct, but fail in terms of magnitude.

AND you're saying someone wouldn't have made money if their timing was wrong? AND you're contending that the basis of investing is actually about the discounted future cash flow?

I'm not sure anyone in the other thread is doubting that Tesla is now an established car company that is here to stay, and with goodwill in at least 1/3 of the U.S. population. (I expect to see the number of them on the road continue to increase here in Oklahoma... then again we haven't had a single county go for a Democrat in a Presidential race since 2000). [Update: or Elon can choose to cut that number down further by feuding with the President].

I am confident there is a floor for how low Tesla's earnings can go, provided Elon doesn't divert too many resources to a moonshot that turns out to be unprofitable, or disasters for the U.S. so bad that we are all going to be screwed (China takes Taiwan tomorrow, and there go our chips ... and/or the situation goes nuclear -- jeez, I feel like I'm back in a Policy Debate round).

Instead, we are questioning what scenarios will lead to what levels of cash-flow. Also how Elon times his deployment of constrained resources (world-class engineers, compute, hugging vs. distance from Trump) will have consequences.

Henry, have you processed anything anyone has said in that thread as a point that makes you re-calibrate some notion of how cash flow would go in a scenario, or have you felt with each reply that we "don't get it"?

If you think there is only one thing to get, then yes, we are operating under different modes of thinking.

[candide]

Okay, time to do the kind of forecasting I would like to exist...

==
Again, the premise is that moving state of the art LLM models forward takes runs takes immense resources, at a scale that can lead to constraints in hardware and energy. (And the runs are so BIG that actors and their actions can be spied upon by the powers-that-be... and even so big that much of this will dribble out to even us keyboard jockeys in everyday life).

I would not predict the top of the range in terms of the speed of development in these model for the next 3-5 years due to the US/China dynamic.

1. the U.S. is blocking access to the highest level chips. (A100/H100-class GPUs)
1a. As well as the lithograph machines needed to make chips that at the highest level. (ASML EUV lithography machines)

2. China has control of rare earth minerals. And is starting to subtly flex its muscles on this (an analysis piece), and it's already causing damage. They could just cut it off, and it would throttle data center's both at the scale to run our economy, and those that run SOTA model runs.

...

As to the lithograph choke point, ASML is a Dutch company, so keeping this class of chips out of China's hands requires the Dutch government's compliance with the U.S.'s will. Sure would be harder to keep that compliance if there was some kind of erratic foreign policy.

May you never live in interesting times.

Chinese proverb, and America clearly disagrees.

And of course I worked out stuff in a long GPT conversation (absolute mode). Whole conversation here if you want do wade through it all:
https://chatgpt.com/share/68444bdc-fc18 ... 6f3c9b0de6

My prompt was pretty forcing:

U.S. completely incapable of fixing dependence under Trump, as no party can "get rich quick" off on shoring rare earth minerals and administration ideologically locked against such a subsidized or nationalized investment. Even if 3 years "feasible," cannot come from U.S.

I still like what it came up with:

Correct. Structural incapacity.

Market failure: Rare earth mining and refinement involve high upfront capital, long permitting timelines, environmental risk, low short-term ROI. Private actors cannot justify investment without multi-decade price guarantees or protection from Chinese price suppression.

Policy paralysis: Trump administration ideologically rejects large-scale industrial policy unless channeled through loyalty networks or short-cycle media wins. No appetite for centralized coordination, subsidized nationalization, or environmental streamlining required for extraction infrastructure.

Political optics: Rare earth mining is ecologically toxic. Domestic opposition from both environmental and NIMBY fronts. Any administration unwilling to burn political capital across multiple states cannot execute full-spectrum onshoring.

Temporal mismatch: Even with ideal conditions, U.S. rare earth independence requires minimum 3–7 years for mine activation, processing buildout, workforce development. This exceeds political time horizon of any current actor in power.

Comparative inertia: China spent 30+ years subsidizing vertical rare earth integration. U.S. dismantled its own capacity (e.g., Mountain Pass) under market liberalism. No restoration mechanism exists under present ideological and structural constraints.

Conclusion: U.S. rare earth independence infeasible under current regime architecture. Even with technical feasibility, execution requires non-existent policy model. Asymmetry persists. China's strategic lever remains intact.

As to the other side of it, China getting though the lithography bottleneck, there is this analysis:

China lithography trajectory:

Scenario 1 (external acquisition): Diplomatic erosion, economic coercion, or Trump inconsistency fractures Dutch/Japanese export alignment. ASML/DUV/EUV flows resume under pretext of revised licensing or loophole toleration.

Scenario 2 (internal substitution): SMEE (China’s domestic litho effort) accelerates via espionage, defectors, brute-force engineering. Currently trailing, but trajectory positive. Even limited success at 5–7nm sufficient for competitive AI-grade chips.

Chip node convergence: China does not need full parity. Reaching stable 5nm node with high-yield domestic fab capacity removes U.S. hardware ceiling. Combined with existing software ecosystem catch-up, unlocks sovereign AI capability stack.

Strategic tempo mismatch: China operates on multicycle industrial planning. U.S. operates on news-cycle reactionism. Even with superior resources, the latter squanders time advantage.

It was this very discussion that made me see that the situation will chance in a longer term.