Aurornis

I’ve been spinning up VMs or Docker containers for development. Run the harness inside the VM or container. Note I do this on Mac so the Docker containers are actually in a VM, but standard disclaimers about Docker not being meant for security apply.

Popular harnesses have mature enough remote access features that I can use the same interfaces and even mobile apps.

If you’re not familiar with Docker, you can even have an LLM set up the environment for you and help you start and stop the container. If you do this, be very sure that you switch to having a harness run inside the VM or container, not having the harness on the host machine run commands inside the container. If you do the latter style, it will be running commands on the host which is what you want to avoid.

> Real ones are oncall for the systems they build.

That’s what DevOps meant where I worked.

I remember the separation of developers from deployment being one of the things DevOps pushers wanted to reduce.

In fact it was the beginning of my career when the older senior people were already trying to get away from the assembly line style of software where each person worked on their part and then tossed it over to someone else to become their problem.

More recently the ride sharing Bubble.

For years the consensus on Hacker News was that the ride share companies were a bubble and it was all going to collapse. Nobody would pay profitable rates once the early discounted rides ended, everyone said.

Then the rideshare companies became sustainable and they’re accepted as a normal part of life.

> Essentially there's an asymmetry, because companies going bankrupt have no liability for debts, but they gain all profits if things go well. This has always encouraged risk

Companies are liable for debts. I don’t know why you would conclude they’re not. If a company can’t or won’t pay their debts, they go into bankruptcy, which has significant consequences up to the company losing its assets to the creditors.

Lenders are not dumb. They collect interest in exchange for the risk of lending. This is the symmetry that you’re not seeing. The lenders can also structure their lending in ways that if the company can’t pay, the lenders get assets from the company in exchange.

LLCs apply at a lower level (not AI lab scale), and shielding individuals from downside risk is a feature, not a bug. It’s by design. There are a lot of countries where the only way to get startup funding involves signing your personal finances away in the event of failure. Financial ruin if the company doesn’t work out, aside from some minimal protections. Do you know what the small business and startup environments are like in those countries? Abysmal. Nobody can afford to take the risk, so the big establish companies are the place to work.

> Nobody depends on AI for core operations.

Financial failure of an AI lab wouldn’t make that AI disappear.

If people depend on them (aka demand) the IP of the labs would be valuable, someone would purchase it in bankruptcy, and they would continue serving the model.

There is basically no comparison to the credit market collapse, even though all of the simplified comparisons try to draw a straight line between the two.

The real concern is different: The theory is that if the frontier labs start collapsing, they will argue that their failure is equivalent to the US losing AI supremacy, which is equivalent to losing the future if you believe AI is winner take all. The argument is that they’ll be financially reckless and then ask the government to backstop losses if it falters, both to retain their lead and to protect the overall economy.

The other part that of this is the increasingly widespread belief that AI is a total bubble without demand and it’s all going to collapse because there’s no profits to be found at all because it’s all financial theater. That argument is harder to make if you’ve looked at the actual unit economics of these companies (hint: Anthropic isn’t subsidizing inference like so many believed). AI is not going away and the demand for compute will only increase. The question is how accurately these scale ups track it and if they’re going to dip below the insolvency threshold from getting it wrong.

But if the companies blow up, the technology doesn’t go away. The model weights aren’t deleted. The ML engineers don’t retire. Companies get restructured, IP assets can be sold to new operators, and engineers move to new companies with new business models.

> IMO, the crux of the issue is that higher education hasn't expanded availability with population growth

I don’t know where you're getting this. The complaint is that too many people are going into higher education where they shouldn’t be, getting expensive educations for careers that don’t need it.

An oversupply of education, not undersupply.

> It's access to the upper crust of society that you're paying for, not an education.

Generic college degree does not get into the upper crust. It’s a basic signal that someone is able to follow a program, stick with something for a few years, and complete some work. Sounds basic but when you’re hiring people it’s a valuable signal.

For most people trying to send that signal, getting scholarships plus in-state tuition is the right answer. A good enough college education doesn’t have to be expensive.

To get into “upper crust of society” range you can’t treat college like a checkbox where any college will do. The education does actually matter. The cynical hot takes about college being only about networking and a piece of paper are wrong at the upper end. The educational programs at the upper end are hard. They’re demanding. You’re going to have to learn something to graduate. People do actually fail out or give up, which is rare at the generic programs.

Even if you don’t use much of the education of one of those top programs (like if you study computer science and then take a finance job) you’ve learned how to learn, how to get hard things done, how to manage yourself and your time, and how to follow through on something hard that most people cannot do. That is the real signal of a top education program if you’re going to discount the education part.

There’s always sneering about how someone knew an Ivy League grad who was actually the dumbest person they know, or anecdotes about how the smartest programmer I ever met had no college degree, or other anecdotes they try to disprove this. They’re missing the point that it’s not a 100% perfect signal nor the only way to be good at something, but it is a strong heuristic that someone is at least equipped to set goals and accomplish hard things.

> If my kids have a 1/3 chance of interacting with someone that doesn’t think college is worth it,

Many people believe college is not worth it for everyone, but believe a college degree is important for the job they’re hiring for.

This statistic also includes people who think college is too expensive for what it delivers, but that the value of a college degree is still nonzero.

There is yet another level of people who are frustrated with how academically weak and unchallenging common colleges have become, or how many students are going to 4-year degree programs to study topics with no job applicability. The answers would change completely if you asked people if a history degree from a $60K per year private college was worth it, compared to a Computer Science degree from Stanford. The question “is college worth it” is too generic to mean anything now that it could mean anything from a $200K piece of paper without real job skills, or a golden ticket into the recruiting funnel at FAANG.

This creates the discrepancy between polls like this and actual revealed hiring preferences. Many will tell you they don’t think college is worth it but will prefer applicants will college degrees.

> If I ask them a question about which library is best for a given task, they'll download 4 different candidates, read the source, and write test programs. Meanwhile Opus will look at a couple of web pages and give me glib bullshit answer. Opus is often like a brilliant coworker who can't be bothered to do their job.

I’ve been using Qwen3.8FN and DeepSeek. I would not trust them to evaluate 4 different libraries from scratch unless I gave them extremely specific benchmarks to run. They’re very good at consuming tokens for long stretches and then giving you what looks like thoroughly researched output, but their intelligence just isn’t at a level I trust.

But really, the difference you’re describing is easily shaped by prompting. If you want Opus to download and compare the libraries, just tell it to. It will. All of the models benefit from being clear in your intentions. If you try to find models that will read your mind where spending 20 extra seconds to type out your intentions would accomplish the goal, you’re ignoring the very important intelligence and efficiency differences between models.

> Look, all I actually need out of a model is things like "tell me where this distributed algorithm might break", "get this ancient C code to build on modern Debian with modern dependencies", or "write a plan to migrate this database from MySQL to PostgreSQL and to test that the app still works." The modern open models can do all that, no problem.

I actually read the code output for projects that matter. There is a huge difference between a model that can eventually accomplish a shallow goal like “fix this bug” and a model that does it cleanly. Even with Opus or Astra I will spot cleanup issues frequently.

When using models down to Qwen3.8-FN or DS-Flash size, the code can feel like the kind of spaghetti mess that you get from an overly caffeinated and confident junior developer who accomplished the goal via brute force and staying up all night. Yeah it compiles and does the thing you asked, but all of that spaghetti code compounds with each change until a couple months later your harness is spending half a day and countless tokens doing simple tasks because it has to navigate the mess of a codebase.

That problem is much less of an issue at the Opus level and I can trust Opus to clean up better when directed to do so. It’s those differences that compound.

But if I’m doing a throwaway task in a sandbox that just needs to work for a day? I’ll throw it to the open models and see what they can do. They are interesting to play with.

> Time and time again, when I try some newly released open weights model I am disappointed

The article is correct that the major labs are not benchmaxxing, but I think it overstates their focus being on something else.

They don’t play the benchmark manipulation game because they’re comfortably in the lead and they have their own teams producing internal benchmarks and targets that are representative of what their users expect.

Your experience with the open weights models is not missing anything. Nearly all of them are disappointing if you set your expectations according to where they fall in the benchmark charts. They can be useful for some tasks once you’ve developed an intuition for their capabilities, but you have to start learning their relative cost per task and their limits, which takes a lot of observation and honestly might not be worth it.

One of my unpopular open weight model hot takes is that, right now, OpenAI’s Luna model is often cheaper per task than even DeepSeek and others (paying API prices, not based on the fluctuating subscription prices).

The real advantage of open models has always been that you can run them on your own hardware or a rented server and have more control over where your data goes. Some of the hype about them being 3 months behind frontier models or people making claims that 27B local models have replaced their Claude subscription become extremely questionable if you actually use both open weight and frontier models frequently.

> For starters, the limits are dreadfully low. Five messages on medium is already enough to trigger a 4 hour cooldown.

Is that chat messages that involve tool calls (web search, file reads, etc) or actually just messages?

When I'm on the $20/month OpenAI plan, I can get about 3-5 turns out of Astra before the 5 hour limit is up.

When Fable was on the $20/month Anthropic plan, I remember getting about the same.

3-4 turns out of the $20/month plan would seem about par with other providers for their best model, but really Carbon should occupy that slot, not Argon.

If Argon's limits are that low, it wouldn't make sense to use that plan for anything when models of similar intelligence are available with much higher limits on other $20 plans.

There are situations where this makes sense.

I gift meals to people in other cities when they're in need (sick, childbirth, death in the family) and I'm not nearby to help.

I've had intense working situations where an extra hour of work was worth paying someone $30 to deliver my $100 of groceries.

We've frequently used DoorDash and other services for small business meetings where we wanted to provide lunch, or in some situations where we had to stay late (security incidents, etc) and wanted to provide the team with something better than pizza delivery.

However it's also scary to watch some people who will DoorDash whenever they feel like it without regard to the cost. I watched some eye-opening conversations among college students comparing how much they had spent on DoorDash in the past year, even though they were not in a financial demographic that should have been indulging in delivery.

Municipal services like garbage pickup are generally funded by a mix of per-household fees and property taxes. Property taxes loosely correlate with the size of the household (and yard and construction materials and everything else that scales with price) while the per-household fees can scale depending on how big of a collection can you request.

It's a good system for aligning the costs with the people who use the service.

If you thought all tax money went into a big general bucket and the government decided what services to offer out of that, municipal garbage pickup does not work that way.

There couldn't be a free delivery service because the usage would be all over the place. Someone DoorDashing burritos for breakfast lunch and dinner 7 days a week would use it 20X more than someone doing their once-weekly grocery order. Subsidizing it would incentivize a lot of waste, traffic, vehicle usage, and other things that we actually don't want to incentivize people to use more of.

The only way this could work is with a substantial per-delivery fee, which gets us right back to what we already have with DoorDash.

Every time it gets posted most assume that the forest was supposed to be protected and they decided to cut it down anyway.

The details are much less dramatic: The company doing the work messed up on a part of the approval process that was supposed to be completed first. It’s technically possible that the environmental impact review would come back with some obscure exception that required them to revise their plans, but it’s unlikely that anything was going to stop those trees from being cut down.

The 250K trees headline appears dramatic if you don’t know that Finland is covered in trees and cutting down a lot of them is part of building anything.