HNHacker News
TopNewBestAskShowJobs

mediaman

6,586 karma · joined April 20, 2009

submissionscomments
mediaman··on The Folk Economics of Housing
This analogy seems confused.

If someone in Chicago moves to Seattle, then our policy options are (1) no new condo in Seattle or (2) new condo in Seattle.

Under policy 1, the new buyer from Chicago must outbid locals for the fixed housing supply; they will wind up buying older housing stock, which otherwise would have gone to existing local residents. Prices go up.

With policy 2, the new entrant buys the new condo and does not compete for pre-existing housing stock.

In this scenario, whatever house is freed up in Chicago is irrelevant to the housing stock in Seattle. I'm not sure why you included it.

The entire question can be contained by the assumption that "there is someone new coming to Seattle" and whether it would be better to have a new condo unit to sell to them or have them compete for existing fixed stock. The whole bit about the Chicago housing market is a distractor, because it stays the same under either policy.

mediaman··on Rerank-2.5 and rerank-2.5-lite: instruction-following rerankers
Keyword search (or something similar in concept, like bm25) would typically be first stage, rather than second, since it can be done with an inverted index.
mediaman··on GPT-5 for Developers
I actually don't agree. Tool use is the key to successful enterprise product integration and they have done some very good work here. This is much more important to commercialization than, for example, creative writing quality (which it reportedly is not good at).
mediaman··on GPT-5 for Developers
It works on Codex CLI, install it with npm.

That's been out for a while and used their 'codex' model, but they updated it today to default to gpt-5 instead.

mediaman··on Study mode
This is endlessly brought up as if the human operating the tool is an idiot.

I agree that if the user is incompetent, cannot learn, and cannot learn to use a tool, then they're going to make a lot of mistakes from using GPTs.

Yes, there are limitations to using GPTs. They are pre-trained, so of course they're not going to know about some easter egg in DDG. They are not an oracle. There is indeed skill to using them.

They are not magic, so if that is the bar we expect them to hit, we will be disappointed.

But neither are they useless, and it seems we constantly talk past one another because one side insists they're magic silicon gods, while the other says they're worthless because they are far short of that bar.

mediaman··on Section 174 is reversed, mostly
He's not saying it's only true when salaries are equal. It makes a significant difference at the margin: if you're an employer debating whether to employ US talent or offshore talent, weighing time zones, skill and other factors against cost, then introducing a tax advantage for domestic hiring will cause those employers who were otherwise indifferent between the two options to now prefer US labor.
mediaman··on GitHub CEO: manual coding remains key despite AI boom
It's a tool. It's not a human. A line number works great for humans. Today, they're terrible for LLMs.

I can choose to use a screwdriver to hammer in a nail and complain about how useless screwdrivers are. Or I can realize when and how to use it.

We (including marketing & execs) have made a huge mistake in anthropomorphizing these things, because then we stop treating them like tools that have specific use cases to be used in certain ways, and more like smart humans that don't need that.

Maybe one day they'll be there, but today they are screwdrivers. That doesn't make them useless.

mediaman··on GitHub CEO: manual coding remains key despite AI boom
AI is less useful when I'm picking through specific, careful business logic.

But we still have lots of boilerplate code that's required, and it's really good at that. If I need a reasonably good-looking front-end UI, Claude will put that together quickly for me. Or if I need it to use a common algorithm against some data, it's good at applying that.

What I like about AI is that it lets me slide along the automation-specificity spectrum as I go through my work. Sometimes it feels like I've opened up a big stretch of open highway (some standard issue UI work with structures already well-defined) and I can "goose" it. Sometimes there's some tricky logic, and I just know the AI won't do well; it takes just as long to explain the logic as to write it in code. So I hit the brakes, slow down, navigate those tricky parts.

Previously, in the prior attempt to remove the need for engineers, no-code solutions instead trapped people. They pretended like everything is open highway. It's not, which is why they fail. But here we can apply our own judgment and skill in deciding when a lot of "common-pattern" code is required versus the times where specificity and business context is required.

That's also why these LLMs get over-hyped: people over-hype them for the same reason they over-hyped no-code; they don't understand the job to be done, and because of that, they actually don't understand why this is much better than the fanciful capabilities they pretend it has ("it's like an employee and automates everything!"). The tools are much more powerful when an engineer is in the driver's seat, exercising technical judgment on where and how much to use it for every given problem to be solved.

mediaman··on Building Effective AI Agents
What do you recommend for this? I've actually had good luck having them create XML, even though you're "supposed" to use the native tool calling in a JSON schema. There seems to be far fewer issues with getting JSON syntax correct.
mediaman··on Builder.ai did not "fake AI with 700 engineers"
The original story doesn't make any sense. How would you fake an "AI" agent coding by using people on the other side? Woudn't it be...obvious? People cannot type code that fast.

What's your non-snarky theory about how this could possibly be true?

mediaman··on Tell HN: Help restore the tax deduction for software dev in the US (Section 174)
You're bringing up the question of "what is a fair tax rate," which is reasonable.

The questions of this legislation, though, are different:

Should we incentivize companies to hire corporate executives instead of engineers?

Should we favor trillion dollar companies over startups? (It is much cheaper for Amazon to loan money to the government than three people starting a new venture from scratch, so this favors concentration.)

If you agree that we should discourage hiring engineers in favor of management, that concentration is good, and that low corporate tax rates are good, then this legislation is perfect.

I say that it's good if you believe in low corporate tax rates because this legislation was passed to pay for overall corporate tax cuts, which primarily benefits the largest companies. Amazon actually pays $11.3 billion in income taxes a year (not zero), so on net, even though they have a lot of software engineers, they benefit from this legislation, because they effectively traded having to float money to the government in exchange for lower tax rates.

Big companies care about tax rates more than liquidity, because their borrowing rates are cheap, whereas small companies care more about liquidity (they effectively cannot borrow, or it is very expensive) and their profits are low. So this effectively subsidizes big companies at the cost of small companies.

mediaman··on Tell HN: Help restore the tax deduction for software dev in the US (Section 174)
NOLs are generally not transferrable in the US (they used to be, but now the benefit can only be used if the acquirer of the 'rump' continues the existing operating business).
mediaman··on LLMs are cheap
Salary, mostly. It's useful to separate out the GPU cost of training from the salary cost of the people who design the training systems. They are expensive.

That does not mean, however, that inference is unprofitable. The unit economics of inference can be profitable even while the personnel costs of training next-generation models are extraordinary.

mediaman··on Tell HN: Help restore the tax deduction for software dev in the US (Section 174)
A lot of people don't know what this Section 174 is about, so here's a brief explainer.

Normally, when you have expenses, you deduct them off your revenue to find your taxable profit. If you have $1 million in sales, and $900k in costs, you have $100k in profit, and the government taxes you on that profit.

Section 174 says you can't do this for software engineers. If you pay a software engineer, that's not "really" an "expense", regardless of the fact that you paid them.

What you've actually done, Congress said, is bought a capital good, like a machine. And for calculating tax owed, you have to depreciate that over several years (5 in this case).

Depreciating means that if you pay an engineer $200k in a year, in tax-world you only had $40k of real expense that year, even though you paid them $200k.

So the effect is that it makes engineers much more expensive, because normally when a company hires an engineer, like they spend on any other expense, they can at least think "well, they will reduce our profit, which reduces our tax obligation," but in this case software engineers are special and aren't deductible in the same way.

In the case of the $200k engineer, you deduct the first $40k in the first year, then you can expense another $40k from that first year in the second year, the third $40k in the third year, and so on through the fifth year. So eventually you get to expense the entire first year of the engineer's pay, but only after five years.

The effect is that companies wind up using their scarce capital to loan the federal government money for five years, and so engineers become a heavy financial burden. If a company hires too many engineers, they will owe the federal government income tax even in years in which they were unprofitable.

These rules, by the way, don't apply to other personnel costs. If you hire an HR person or a corporate executive, you expense them in the year you paid them. It's a special rule for software engineers.

It was passed by Congress during the first Trump administration in order to offset the costs of other corporate tax rate cuts, due to budgeting rules.

mediaman··on My AI skeptic friends are all nuts
That's the old way of thinking about software economics, where marginal cost is zero.

Marginal cost of LLMs is not zero.

I come from manufacturing and find this kind of attitude bizarre among some software professionals. In manufacturing we care about our tools and invest in quality. If the new guy bought a micrometer from Harbor Freight, found it wasn't accurate enough for sub-.001" work, ignored everyone who told him to use Mitutoyo, and then declared that micrometers "don't work," he would not continue to have employment.

mediaman··on My AI skeptic friends are all nuts
What are you doing to get results this bad?

I tried this question three times and each time the first two products met both requirements.

Are you doing the classic thing of using the free version to complain about the competent version?

mediaman··on My AI skeptic friends are all nuts
Just like some people who wrote long sentences into Google in 2000 and complained it was a fad.

Meanwhile the rest of the world learned how to use it.

We have a choice. Ignore the tool or learn to use it.

(There was lots of dumb hype then, too; the sort of hype that skeptics latched on to to carry the burden of their argument that the whole thing was a fad.)

mediaman··on My AI skeptic friends are all nuts
If an LLM improves coding productivity, and it is better at one language than another, then at the margin it will affect which language you may choose.

At the margin means that both languages, or frameworks or whatever, are reasonably appropriate for the task at hand. If you are writing firmware for a robot, then the LLM will be less helpful, and a language such as Python or JS which the LLM is good at is useless.

But Thomas's point is that arguing that LLMs are not useful for all languages is not the same as saying that are not useful for any language.

If you believe that LLM competencies are not actually becoming drivers in what web frameworks people are using, for example, you need to open your eyes and recognize what is happening instead of what you think should be happening.

(I write this as someone who prefers SvelteJS over React - but LLM's React output is much better. This has become kind of an issue over the last few years.)

mediaman··on Snowflake to buy Crunchy Data for $250M
That only works if there are moats that discourage new entrants (e.g., buying a social media upstart works because of the network effect).

Not obvious serverless postgres is in that category.

mediaman··on As a developer, my most important tools are a pen and a notebook
There are no more “old school” production lines left? I’m not sure what you mean; obviously there are a lot of production lines in the US, including the manufacturing of engine blocks as shown here.

Do you mean that they use robots today and therefore aren’t “old school”? That’s true in most industrialized economies.

mediaman··on The ‘white-collar bloodbath’ is all part of the AI hype machine
The cotton gin processed short fiber cotton, but not long fiber cotton.

Did the cotton gin therefore not compete with human labor?

mediaman··on As a developer, my most important tools are a pen and a notebook
That’s funny, because if you go into any machine shop you’ll find machinists writing things down in notebooks (and probably cursing engineers who never set foot in the shop).
mediaman··on New tools and features in the Responses API
Correct, you are not seeing the reasoning chains.
mediaman··on Why does the U.S. always run a trade deficit?
The idea that the hard goods are sold for "imaginary numbers" is popular but a bit of a misnomer.

These "imaginary numbers" are actually claims on future income generated by the United States economy. It is true that they are "imaginary", in the sense that any property ownership is imaginary in a way that a toaster is not. The most common form is Treasuries, which the government then pays interest on and which must eventually be either repaid or rolled over.

So it would be more correct to say that China is trading hard work building goods in exchange for claims on future US income.

mediaman··on OpenAI Codex hands-on review
A lot of people believe that programming is the typing of odd sequences of characters into a computer.

To them, it seems LLMs are also perfectly capable of typing odd sequences of characters.

The idea that SWEs do actual structured problem solving is mostly native to industry insiders.

mediaman··on PDF to Text, a challenging problem
Corpus size doesn't mean much in the context of a PDF, given how variable that can be per page.

I've found Google's Flash to cut my OCR costs by about 95+% compared to traditional commercial offerings that support structured data extraction, and I still get tables, headers, etc from each page. Still not perfect, but per page costs were less than one tenth of a cent per page, and 100 gb collections of PDFs ran to a few hundreds of dollars.

mediaman··on Gemini 2.5 Pro Preview
I find they achieve acceptable, but not polished levels of work.

I'm not even a designer, but I care about the consistency of UI design and whether the overall experience is well-organized, aligned properly, things are placed in a logical flow for the user, and so on.

While I'm pro-AI tooling and use it heavily, and these models usually provide a good starting point, I can't imagine shipping the slop without writing/editing a line of HTML for anything that's interaction-heavy.

mediaman··on AI code is legacy code
Why are enterprises running code from 1985?

How long code needs to last is actually highly variable, and categorical absolutist statements like this tend to generally be wrong and are specifically wrong here. Some code will need to change in a year. Some will need to last for forty years. Sometimes it's hard to know which is which at the time it is written, but that's part the job of technical leadership: to calibrate effort to the longevity of the expected problem and the risks of getting it wrong.

mediaman··on How to live an intellectually rich life
Yes, it's kind of a strange slice - we get Faulkner three times and we get Joseph Conrad no fewer than four times(!), but not a single book from Dostoevsky or Tolstoy? No Bulgakov, no Turgenev? No Flaubert?
mediaman··on Port of Los Angeles says shipping volume will plummet 35% next week
If you have ever been involved in new factory construction for anything that's more than light manual assembly, this statement would be very amusing to you.
← PreviousPage 7 of 34Next →