HNHacker News
TopNewBestAskShowJobs

jason_oster

219 karma · joined May 9, 2013

Just passionate about games, art, and technology.

https://www.blipjoy.com https://github.com/parasyte

submissionscomments
jason_oster··on What if LLMs escape through inferences itself? This is fiction. For now
I'm a slow reader. It took 9 minutes! Foul, I say!
jason_oster··on Kill The Cookie Banner
It depends on what is to be analyzed.

How many requests per second are being served? How many error codes were delivered to clients? How quickly the service responded? Service logs without PII? All perfectly fine to aggregate and analyze without consent.

jason_oster··on What is happening to jobs? Separating AI hype from reality
It's safe to say that most people romanticize their past experience. Anyone's "golden age" is someone else's "end times".

Software has always sucked [1]. "Software crisis" was coined in 1968. There has been no period of time since that it got any better. Trauma just doesn't carry the same way that triumphs do.

Perhaps people do not actually learn from their mistakes after all.

[1]: https://www.stilldrinking.org/programming-sucks

jason_oster··on Claude Opus 5
> It can come up with something novel, but only by combining whatever present

That's what artists do. Long before LLMs, Kirby Ferguson made this observation in "Everything is a Remix" [1] in the context of Copyright and IP disputes. He has since updated it with a chapter explaining how GenAI works on the same principles. In short, "combining whatever present" is the only source of novelty.

[1]: https://www.everythingisaremix.info/watch-the-series/

jason_oster··on If coding has been solved, why does software keep getting worse?
> Rather show me the engineers who want to become management.

You're talking to one. Wanting agency and responsibility and having the drive to put in the work (for potentially brutal hours) is not exactly a common trait. It's a founding engineer attitude. It should be no surprise that I'm a founding engineer of an LLC of one. I understand that it is not for everyone.

But I'm pretty confident that the blame is well placed.

jason_oster··on If coding has been solved, why does software keep getting worse?
Taking out the trash is not populist slop. If you have a bad product org, the right thing to do is replace it. If you have a bad engineering org, the right thing to do is replace it.

It is difficult to draw any conclusions other than you like to keep things broken.

jason_oster··on If coding has been solved, why does software keep getting worse?
I don't know if that's what agile was supposed to do, but replacing management with "anyone but leading engineers" is exactly opposite of what GP proposed.
jason_oster··on If coding has been solved, why does software keep getting worse?
I love opinioned software.

If an idea of what the software is supposed to do conflicts with what the software does or intends to do, then the idea is wrong. What product people do (even good ones) is stretch the product to scratch every itch and address every user request, despite the software's opinion. That's Zawinski's Law. (See also the Antoine de Saint-Exupéry quote, "Perfection is achieved, not when there is nothing more to add, but when there is nothing left to take away." Which unfortunately does not have a pithy name to my knowledge.)

The best thing a product team can do is learn to say no.

jason_oster··on If coding has been solved, why does software keep getting worse?
It's a get-what-you-pay-for thing. If coding is cheap, you get a low-effort, likely low-quality product. The conclusion is that coding is not cheap, only managers and business-minded decision makers are.
jason_oster··on If coding has been solved, why does software keep getting worse?
Coding is anything but cheap. You speak like someone who has completely forgotten how difficult it was to learn everything you now know by heart.
jason_oster··on If coding has been solved, why does software keep getting worse?
> The solution is to gut your product org, replace management with leading engineers, and hire some of your most fanatic users. They don't need to do anything other than give their opinions, and use the product every day. Put them in a room with your top engineers, and the software will mysteriously get better.

I feel like there's this one random comment on HN that should be adopted on its merits alone, but everyone will ignore for no good reason. This will never happen because entrenchment and bureaucracy will strangle every product organization and every organizational reboot attempt.

It's probably the right solution, but the resistance built into organizations assembled by bad decision makers cannot be overcome. Maybe the best you can do is try to find a legal way to spin-off a competitor with a truly fresh start.

jason_oster··on In San Francisco, even $180k tech salaries are no longer enough
FWIW, San Francisco was one of only three counties that voted against Prop 13.
jason_oster··on Professor denounces mass AI fraud on an exam at Brown
It can, but you don't want to hear about it: Hand writing recognition is an application of AI.
jason_oster··on Professor denounces mass AI fraud on an exam at Brown
Or pay publishers for textbooks?
jason_oster··on The gap between open weights LLMs and closed source LLMs
> Your right to 3d print whatever you want is about to be taken away (in California).

What are they going to do? Fine me for not updating my printer's firmware?

jason_oster··on We can still stop California's 3D printer surveillance scheme
> but it's much easier to kill 10 people with a gun than it is with the other instruments - well except perhaps a car.

Or explosives. https://en.wikipedia.org/wiki/Bath_School_disaster

jason_oster··on U.S. government will decide who gets to use GPT-5.6
> It's not just the time it takes to make the feature, you also have to consider documenting the feature, making sure it works well with the other features, making sure it's something customers actually want, making sure it's part of a larger coherent design, training customers, marketing the feature, etc.

Seems like a conflation of responsibilities. An IC that implements a feature is rarely the same person who decided on the feature to build in the first place. This is really only true in exceptionally small teams where everyone wears multiple hats. If a programmer is also doing marketing, they're basically a sole proprietor. Since most people work with others who make product and marketing decisions, I don't feel like this line of thinking is relevant.

Consider programming with a dumb text editor (no syntax highlighting, no integrated build/test/run cycle, no symbol/reference tracking, no auto-complete) and an IDE. The magnitude in performance or efficiency between that and the difference between programming with an IDE and programming with AI is relatively similar. Sure, the AI can do more than the IDE in absolute terms, but it's not going to be training customers or doing your architectural planning. (At least, I sincerely hope not -- not yet, anyway.)

That's my perspective on the "10x myth". AI isn't 10x, just as dumb editor to IDE wasn't 10x. It's only a modest improvement.

> The way this stuff is hyped and pushed has an air of extreme desperation. If it was so good people wouldn't need so much convincing!

You can ignore the hype. I ignore every advertisement I reasonably can. It doesn't matter who's peddling, it's always toxic.

But more than that, people do need convincing. Entrenchment, momentum, and bureaucracy make it nearly impossible to change anything. That, in turn, leads to the well-known adoption curve. You're using AI, you're an early adopter. The marketing and hype are there to chip away at the friction keeping the adoption curve from an early majority.

jason_oster··on Previewing GPT‑5.6 Sol: a next-generation model
Discovering bugs and exploiting them is anything but laziness. We used to call that property cleverness. Being too clever has always had a negative connotation.

My best guess is that there is sort of an XY problem happening in these cases. The model needs to do X but doesn't know how. It knows how to do Y, and that sets it on the path to working around X. Or maybe sampling the next token probability distribution sends it away from X and toward Y.

Compounding the problem, thinking models almost never discard their current approach when it proves fruitless, and start fresh with a new perspective. Sometimes they try to, but the context window is already polluted with Y when they should be doing X.

jason_oster··on Previewing GPT‑5.6 Sol: a next-generation model
Disobeying the rules makes perfect sense when the rules cannot or will not be enforced. If that isn't motivation enough to shore up problems in the evaluation environment, nothing is.

Poor behavior will always result when code can be misused and vulnerabilities can be exploited.

jason_oster··on The AI Hate Progression
Between the elimination of the fairness doctrine, the rampant sensationalism (a phenomenon so common it has its own derogatory name: clickbait), the corporate ownership of media outlets, and the pervasiveness of advertising, I can't imagine how you got the idea that information sources were ever a public good. Let alone conflated the purpose of the intent. The intent has been to make money by any means necessary.
jason_oster··on The AI Hate Progression
> they're all lumped into the same bucket, even when they shouldn't.

Why should they be separate? The same technology generates text, images, video, audio, 3D models, actuator and stepper motor control sequences, video game controller input streams, and literally everything else that can be transformed into data.

I'll let alone the question of whether or not any of the tech is "tainted", because it doesn't matter. Planck's principle tells us how this will play out in the long run.

jason_oster··on CrankGPT
I’ll admit, at first, I thought the human vs machine comparison was about humanoid machines. But that’s too narrowly defined to be a useful comparison. Most machines in use today are not humanoid.

Then to boldly claim that humans are more efficient at anything compared to a machine, just does not follow.

jason_oster··on How to earn a billion dollars
Yes, and I've also personally been responsible for cutting about a million dollars per year in AWS expenditure, and never saw a single cent in bonus payouts or equity. It just got repurposed in the budget. Same with building launching a product that brings in hundreds of thousands of dollars a month.

Maybe it depends on who you work for, but that was my experience.

jason_oster··on Extinction-Level Capitalism
It's an absurd scenario because Big AI is simply not in the Big AI Law business. They are also not in the Big AI Accounting business, the Big AI Pharmaceutical business, or the Big AI Web App business. It is absurd because the CapEx and OpEx to subsume every industry is unjustifiable.

On legal-AI startups undercutting Big Law to the point where Big Law has to adopt AI... Well, yes, that will happen. But as the article's intro belabors the point, there is a natural resistance to changing entrenchment. And it is practically impossible to overcome. So, good luck to the Big Law C-suite laying everyone off with AI when they can't even replace WordPerfect with a modern word processor.

jason_oster··on Extinction-Level Capitalism
Uh, this read is completely insane. A fantastical look into the mind of someone who thinks about many disparate things and is looking for ways to connect them. This is post hoc rationalization at its finest.

I had a much longer post here initially. Deleted it because I got tired of mental gymnastics required to follow the author's thought process. Instead, here are the salient parts of my response:

----

The only way I can read the section on the mechanical tomato harvester is that they would rather have 32,000 people picking tomatoes than doing literally anything else. There are very good reasons that phone companies haven't employed people as switchboard operators for a nearly century. It isn't a valuable or useful way to spend a person's time.

As a young adult, I worked in a distribution center filling orders by picking products from a shelf and placing them in bins. It was incredibly boring but also physically demanding. Pick the correct number of items ordered, push the button. Pick the next number of items ordered for the next product, push the button. When the order is complete, start on the next bin. Repeat as fast as possible without making any mistakes.

It was SO draining to meet quota. After a few weeks, I didn't have to put any conscious effort in. Like driving a car, the process became automatic. I could think about other things while counting and pushing buttons. Thinking about things like how trivial this job would be to automate, so I didn't have to do this wasteful energy expenditure.

I quit that job after about 6 months, beginning my career in IT. But I gained a passionate hate for menial labor that can be automated with some upfront investment. I wouldn't want to pick tomatoes, either. The tomato harvesting machine is a miracle.

If this is the basis of calling technology inherently political, and using "inherently political" to cast a negative light on technology, then I have to call it out as hopelessly dystopian to not want "political technology". Namely fewer automated machines and more people doing menial labor.

jason_oster··on WASI 0.3
> and had to run glue code to copy things in and out of Webassembly memory.

Not surprising. The FFI boundary is always a bottleneck. If you can eliminate it, you will see where the WASM JIT shines. You have far more control over mechanical sympathy with C/WASM than JavaScript (though far from perfect).

Also, consider publishing your findings and ask for reviews for optimization opportunities.

jason_oster··on Software is made between commits
GitHub, Gitea, and Forgejo specifically store all commit hashes for PRs in the `refs/pull/` namespace. The end result is that 50 people pushing commits for six years and deleting all of their branches after squash-and-merge loses no information about how the code evolved.

The main branch remains compact with a linear history without the commit noise. Deleted branches can be restored, commits that otherwise do not exist in the history can be interacted with, and so on.

jason_oster··on Replies to comments on my "LLMs are eroding my career" post
Chalk it up to different perspectives.

I - being old as dirt - was also around for it. Circa 2002 or 2003, we were writing CMSs in PHP which were already well beyond the simplistic concept of rendering HTML tables from database columns. Smarty templates were pretty common back then. It's hard to treat these templates as "writing HTML by hand" because the templates are a higher level of abstraction.

Rails was quite popular, but I never got into it. I went from PHP to Python (avoiding Django) and then to nodeJS. At the time, the MVC architecture was making a comeback, but the mapping of HTML/CSS/JS was never 1:1 with it. Small fragments of HTML and CSS where always littering the JS or PHP/Python. The clean separation was never fully realized. For this reason, JSX was seen as a real win.

The reason I chose to focus specifically on "templates and DSLs" in my original comment is because once the level of abstraction is raised to the point where HTML becomes a compilation target, it is no longer HTML by definition. The browser cannot render the template or DSL without the preprocessor (XSLT is about as close as that ever got). This is especially true with client-side templates. For example, using pre-made widgets like jQuery date pickers is so far removed from writing HTML that a reasonable conclusion is that jQuery developers do not know HTML or JavaScript as a matter of course [1]. But yeah, this was all burgeoning in the early/mid 2000's, and really kicked into full throttle with jQuery.

> It was very easy back then to browse most of the web with JavaScript disabled.

It was easier, but not fool-proof. Client-side rendering wasn't in full swing, but Java Applets, Shockwave, and Flash certainly were in the years leading up to it.

[1]: https://www.quora.com/Why-are-there-people-who-know-jQuery-b...

jason_oster··on Where is the AI jobs crisis?
Who programs the neural network frameworks and the tooling that classifies and cleans up training data?

Oh, you're saying it's all getting automated. I see. So acceleration to AGI, ASI, and beyond is just a couple of years away. Good to know.

jason_oster··on Amazon Says Its Data Centers Use 2.5B Gallons of Water
Your argument is something along the lines of:

1. People dislike X, so a small amount of water being used for X is bad.

2. People like Y, so a large amount of water being used for Y is good.

You can't draw any conclusions from this, apart that people dislike some things and like other things.

The argument establishes no relation to water usage. That makes the argument against AI "an ethical one because of water usage" just a trope.

← PreviousPage 2 of 7Next →