I built this over 6 months, almost entirely with AI, mostly Opus 4.6 in Claude Code. SF weather made no sense to me (Barely any seasons? September is the warmest month?) and I wanted to understand it better myself. This is a polished version of the app I'd want for myself, adding physics layer by layer to isolate the impact of each piece, and using an LLM to analyze and explain the data.

The models know more about math, physics, and software than I do — but especially on the physics side, they have terrible intuition. Claude can "get the error relative to observations down to 4 °C" just fine, except it'll totally hack and overfit the physics along the way. Subagents to subjectively verify "the physics is sound, no overfitting" didn't really work either. So I had to review the physics code manually.

The entire model is first principles; no machine learning or using observed data at all, except fundamental constants like the radiation of the sun and an elevation map. But after a while, it started to feel like "machine learning in slow motion": instead of an ML model training its parameters, Claude and I were choosing parameters by hand. Some amount of tuning parameters (within a physical range of uncertainty) to match observations is inevitable.

The in-app LLM layer has a tool to evaluate arbitrary math expressions over the simulated data using an AST, which was also pretty fun to build.

Repo: https://github.com/crackalamoo/building-earth

Comments URL: https://news.ycombinator.com/item?id=47691681

Points: 1

# Comments: 0

Has anyone done this for larger neural nets? Is there a way to extract some kind of pattern or is the image just noise no matter how you construct it? I'd be curious to see something like that

I think the limiting factor is that the AI still isn't good enough to be fully autonomous, so it needs your input. That's why it's still in copilot form

Maybe it makes sense if you're selling a product or service, but I don't see the appeal of AEO as the new SEO. Maybe I'm missing something?

After all, many of your readers are also human beings with lives, maybe even lives similar to yours based on your professional content. (The rest of your readers are LLMs.) Your readers might appreciate your perspectives on random life things or just getting to see what their favorite blogger is up to.

Granted, this isn't the best UX because I can't create a fresh context chat without making it temporary. But I'd say it allows enough choice that overall having the memory feature is a big plus.

But yeah, I still don't see how an 85% average would be a 4.0.

Not that I disagree necessarily, just wondering if there's a consensus that LangChain is too opinionated/bloated/whatever for real industry applications, or if there's some other reason.

> Do they reason? No. (Before you complain, please first define reason).

Defining reasoning is the problem. No, they don't reason in the same way as humans. But they seem to be able to go through reasoning steps in some important way. It's like asking if submarines swim.

The human brain uses just 20 watts of power, so it seems to me like it is possible to reach human-level intelligence in principle by using much greater power and less of the evolutionary refinement over billions of years that the brain has.

But I can't quite articulate why I believe LLMs never step outside the circle, because they are seeded with some random noise via temperature. I could just be wrong.

I'm not sure how/when codepoints matter for wcwidth: my terminal handles many characters with more than one codepoint in UTF-8, like é and even Arabic characters, just fine.

I feel it is quite important to dispel this idea given how widespread it is, even though it does gesture at the truth of how LLMs work in a way that's convenient for laypeople.

https://www.harysdalvi.com/blog/llms-dont-predict-next-word/

Comments URL: https://news.ycombinator.com/item?id=44117032

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=43895095

Points: 2

# Comments: 2