elgertam

There were also certain IMO low-value questions that really excited the SO hive-mind. I asked a question about the peculiarities of Python assignment syntax, and earned several dozen points for the question, even though no one really should have written code the way I presented it.

I liked StackOverflow for the first ten years or so of its existence, but I gradually stopped using it then suddenly quit altogether when valid questions were being closed unreasonably. At this point, LLMs with documentation in the context, issue trackers and eve the source code (if available) have surpassed SO. Now my main issue is telling the LLM to crap on my idea rather than wishing it were kinder.

As a Virginian, this is good information to have. I see a lot of ludicrous objections to data centers here (the most ludicrous being water consumption, when most of our data centers have closed-loop systems and regardless the humidity here isn't evaporating water).

I've suspected that the energy regulations and the ruling party's close connection with Dominion Energy (the Governor recently attempted to fire the chair of Virginia Tech's board and replace him with the CEO of Dominion) have had an impact on power use more than data centers themselves.

Let's just say this organization is very, very large and doesn't necessarily have the budget for everyone to have all of the tokens.

I ran into a problem at work recently: we are given access to a bunch of models up to a full Claude Opus 4.8, but a monthly budget of 100k tokens. We are also given access to Gemini 3.5 Flash & 3.1 Pro with a daily budget of 50M tokens, but no tool calling. I'd love to hook Claude Code (or Pi) into the Gemini model, but the lack of tool-calling makes it quite difficult. I've been planning out how an intelligent router might be able to use a token-efficient tool-calling model (including a small local open-weights model) to handle the basic tools like reading from the file system or interfacing with MCP servers such that context is gathered, but then send the built up context to the Gemini model where I have a nearly unlimited (for my use cases) token budget.

Could your router handle this?

My first thought when looking at it is that I doubt the vehicle could pass US safety regulations. Maybe I'm wrong.

The design of transformers (including LLMs and multi-modal transformer-based models such as OpenAI's image generators) is to attend to relevant details. OpenAI did this at first without guardrails. In response to public backlash, they bolted on "content filtering," which IMO seems like a very GOFAI approach, and regardless doesn't work very well. It routinely flags innocent prompts, then with crafty prompt hacking will generate these kinds of images.

The design of the model is literally to find patterns and attend to them. The infrastructure and process around an OpenAI model is intended to filter "bad" things (in this case, I agree that the outputs are bad), but is designed to stop some enumerated-ish list of things that aren't allowed, perhaps with some limited "reasoning" about them.

I don't exactly appreciate words being put in my mouth. When did I say it was working perfectly? And we're comparing you, a human with common sense and real intelligence, to a multi-mode LLM?

The transformer was designed to attend to relevant pieces of context and generate new ones that match the pattern. OpenAI in particular was doing that work without guardrails, then attempted to bolt on "content filters," which in my opinion just can't work in a rigorous way. (I think Anthropic's "constitutional" approach is much better though not flawless. And regardless, Claude models don't generate images.)

So, yeah, working as designed. Maybe not as intended, because these things are somewhat resistant to the host's intent when the prompter is hostile.

The spontaneity isn't that ChapGPT woke up and sent this to the author. The spontaneity is that ChatGPT was asked to restore an image that was attached without filtering it, and when no image was attached, instead of generating an error message, it cobbled together random outputs, some of which included graphic, disturbing imagery.

But that's not what happened. The missing image was described as "graphic" or "violent." If I were to receive an email with that request and a missing attachment, my imagination certainly would not conjure images of butterflies & unicorns. Seems the model is working as designed.

"If you aren't paying for a taco, you are the taco." --Future AI, probably

I've been using the OpenSCAD version of this for a while. This new release is a big upgrade! I wish it worked with my preferred CAD, FreeCAD. But this is neat!

Take AI for instance. The US grid is struggling to keep up with demand, while Chinese one has a lot of headway [1]. Usually, this could be solved by an increase in spending lasting a few years which would make the debt tick up, but that would've been an absolutely fine use of debt since it buys some shiny new infra that will pay dividends for the next 20ish years.

I object. The CCP is much more deeply indebted than the US when taking into account provincial and local governments as well as state-owned enterprises.[0] And of course the US debt is financed in its own currency while Chinese foreign debt is financed in dollars or other currencies.

The problem in the US is regulation. An environmental impact study takes 54 months in the US.[1] The CCP, which has no problem poisoning its people or even launching rockets over inhabited villages, doesn't delay itself at all.[2] I'm glad we don't poison our people or place dangerous industry in places that could harm populated areas, or even perform some prophylactic measures to protect nature, but I'm confident that we could do this in less then a year (less than six months?) and make much faster progress. Even for something like nuclear, the ten years (mostly caused by red tape) are really onerous.

China is the only one that can run if it comes down to it (unless of course the numbers coming out of China are mega bogus, but for that I don't know enough to have an opinion).

Yes, the common opinion among China watchers is that any number the CCP touches is "mega bogus." They're actually in the midst of something of a financial crisis at the moment because of the high debt.

[0]https://www.statista.com/topics/11662/debt-in-china/

[1]https://www.rff.org/publications/reports/how-long-does-it-ta...

[2]https://arstechnica.com/science/2019/11/china-keeps-dropping...

Vaporwear*

Having read the blog post and then the comments here, I'm rather astonished. Do we understand our craft so little that our only realistic option is to ban LLMs (so-called AI)? Has everyone forgotten we've been in a software crisis for almost sixty years?[0] Have we so internalized the sweat-of-the-brow we've accumulated for decades that it's now part of the identity of being a programmer, and the only reliable signal of whether a contribution is beneficial?

As far as I can tell, architecture, i.e. sound, precise definitions of exactly what a software artifact must do, is now critical. And with LLMs, it's now feasible to begin implementing such things, though many brownfield projects may be intrinsically unsound in ways that their creators are unaware of. In such a world, contributions simply require a modified proof that the software does what it must do, with perhaps additional claims that the maintainers provide.

[0]https://en.wikipedia.org/wiki/Software_crisis

"I only allow robots with stainless steel tools to prepare and serve my food."

it's more about America transitioning from a high-trust society to a low-trust one.

We're talking about Princeton, here. Trust among elites remains persistently high. In fact, it's likely higher than ever due to assortative mating & geographic sorting. Elites, even students in the Ivies, still have trust of government and elite institutions, which the elite stratum itself runs. Trust between elites and lower strata has declined, where elites and middle- and lower-classes have significant mistrust between each other, and the latter have lower trust within their own strata than in the past.

What's more likely IMO is that 1) the cost of cheating (i.e. the cost of assembling a ripped off assignment multiplied by the risk of being caught) has declined precipitously due to LLMs and 2) elite institutions remain the most ruthlessly competitive in the country and even the world.

I set up a VM on Hetzner a few weeks ago. I've been quite impressed so far, and was able to orchestrate everything with Terraform without a problem.

Math, physics, and chemistry RLHF freelancing is typically north of $40/hr. Even competence at simply reading & writing English prose earns at least $20/hr. I've never seen an offer for less than that, and I lived off of that kind of work for a month after a layoff in 2024.

That seems like a fair trade considering the freelancer takes on none of the risk and has very little required capital.

I have a nearly total opposite take. I can't tell you how many times I've read a book, a paper or something else and been confused by some ambiguity in the author's prose. Being able to drop the paper (or even the book!) into an LLM to dig into the precise meaning has been an unbelievable boost for me.

Now I can actually get beyond conceptual misunderstanding or even ignorance and get to practice, which is how skills actually develop, in a much more streamlined way.

The key is to use the tool with discipline, by going into it with a few inviolable rules. I have a couple in my list, now: embrace Popperian falsifiability; embrace Bertrand Russell's statement: “Everything is vague to a degree you do not realize till you have tried to make it precise.”

LLMs have become excellent teachers for me as a result.

I built an MCP server to speak WHOIS/RDAP so I could have Claude give me better domain name suggestions that weren't already taken. It can also be used in LLM-enabled applications (provided that the model is "tool calling" and that there's an orchestrator).

In principle, MCP servers can be created for just about any OAuth-protected API. However, you still need to create the server, and this is where the usage I'm talking about shines: when working on the MCP server, an LLM can be quite helpful in getting the right APIs integrated.

The same goes for other development that doesn't need an LLM context built-in. If I wanted to sync two calendars, for instance, I wouldn't build an MCP that speaks CalDav and Exchange and then let it loose (though this so-called agentic workflow is becoming more popular); I'd want to build software with an LLM's help that can speak both protocols by having it generate code to handle whatever OAuth tokens and scopes are necessary and then help me deploy the thing.

Yeah, semantics is extremely important. Just no way around it.

Well that's another use I have for LLMs: asking questions about these informational or architectural impedance mismatches. LLMs get it wrong sometimes, but with proper guidance (channel your inner Karl Popper), they can be quite helpful. But this doesn't really speed me up that much, though it makes me more confident that my deliverable is correct.

Every time I request the wrong OAuth scope that doesn't have the authorization to do what I need, then make a failing request, I hear Jim Gaffigan affecting a funny authoritative voice saying, "No." I can't be the only one who defensively requests too much authority beyond what I need with extra OAuth scopes, hoping one of them will give me the correct access. I've had much better luck with LLMs telling me exactly which scopes to select.

MCP connects the LLM to the APIs, which can be consulted with "tool calls." I'm talking about integrating the software I produce (with LLM assistance) to APIs. Traditionally, this is a nightmare given poor documentation. LLMs have helped me cut through the noise.

You know where LLMs boost me the most? When I need to integrate a bunch of systems together, each with their own sets of documentation. Instead of spending hours getting two or three systems to integrate with mine with the proper OAuth scopes or SAML and so on, an LLM can get me working integrations in a short time. None of that is ever going to be innovative; it's purely an exercise in perseverance as an engineer to read through the docs and make guesses about the missing gaps. LLMs are just better at that.

I spend the other time talking through my thoughts with AI, kind of like the proverbial rubber duck used for debugging, but it tends to give pretty thoughtful responses. In those cases, I'm writing less code but wanting to capture the invariants, expected failure modes and find leaky abstractions before they happen. Then I can write code or give it good instructions about what I want to see, and it makes it happen.

I'm honestly not sure how a non-practitioner could have these kinds of conversations beyond a certain level of complexity.

We spend a ton of money per capita and get bad outcomes.

I mean isn't that why they ask for name and DoB each time? I get that they want to ensure they're working with the correct patient. I just don't love trying to recall on command the litany of symptoms I've experienced and remedies I've tried before seeing a physician.

Assuming uniquely American selfishness got us in this mess, I don't buy that rugged individualism is the route out. You'll just get that classic pattern of those with enough resources to manage criticizing the resource management of those with too few resources to learn to manage. That just further corrodes solidarity.

I chose my words carefully so as not to imply "rugged individualism" was the solution. I mean I even explicitly invoked "families" as an institution greater than the individual. And I used the word "control" rather than "earned" or something.

Weirdly, "rugged individualism" is kind of why American healthcare is so bad: we tie it to employment for those under age 65. Before the ACA, this was simply because healthcare is exempt from income tax; with the ACA, it's literally mandated. In order to get health insurance, you need to pull yourself up by the bootstraps and get a job OR pull yourself up by the bootstraps and pay extremely high premiums.

Disentangling healthcare from employment and allowing for other institutions like churches or schools to play a part in helping fund it would do a lot. And for the truly catastrophic conditions like cancer, we should just have a universal policy that kicks in.

I see a lot of the comments operating from an empirical framing. This is valid analysis and is good; we should want to understand the waste in the system as it stands.

However, that isn't enough. US healthcare is wildly inefficient because the paying customer is different than the serves customer. This has been known for sixty years, since Arrow published his paper (he identified four reasons, three of which are not exclusive to healthcare and seem to be mitigated well in other industries). I'm surprised people posting can't quite see this: when you go to the doctor, would you call the experience efficient? You check in, then wait, then are called back, tell the nurse or PA why you're there, wait, see the provider who asks you again why you're there, has a short exam, wait, finally get all the paperwork and sign out.

If you have labs or tests, you then wait again. And of course if you need a specialist, you wait again, sometimes for months. If you need any sort of "specialty" medication or equipment, then you REALLY wait, as specialty pharmacies, DMEs and the like jump in.

The whole system is woefully inefficient, and overhead is only a part of the explanation. No one knows what anything costs, and the people who pay (insurance providers, the largest of which is the US Government) want to believe they're not getting scammed - they still are, but at an acceptable level.

The question we ought to ask is how we can buy better health outcomes for people. And I think part of the answer is that in most cases, individuals and families themselves must allocate resources they control to make this happen.

Montana is both cold and sparely populated, so I figure data centers would be good there. Also, I figure Zefram Cochrane could use all that compute for his warp theories in a few decades.

I think one can only hope to reveive such an offer