HNHacker News
TopNewBestAskShowJobs

mediaman

6,586 karma · joined April 20, 2009

submissionscomments
mediaman··on Generative AI is not replacing jobs or hurting wages at all, say economists
Out of curiosity, did you try this in o3?

O3's web research seems to have gotten much, much better than their earlier attempts at using the web, which I didn't like. It seems to browse in a much more human way (trying multiple searches, noticing inconsistencies, following up with more refined searches, etc).

But I wonder how it would do in a case like yours where there is conflicting information and whether it picks up on variance in information it finds.

mediaman··on A $20k American-made electric pickup with no paint, no stereo, no screen
This is mixing two questions: whether a system can be elegantly designed and do the job without major market distortion, versus the question of whether various actors will stand in the way to prevent it.

You could say the same thing about zoning. Higher density is better for affordability, but faces opposition from landowning existing residents. Does that make it wrong, or not worth pursuing? No, and that particular movement seems to be getting traction despite the political opposition.

I read "trivially fixable" as "there is an elegant solution to this," not that "it is easy to get it politically passed."

mediaman··on Large language models, small labor market effects [pdf]
Great read. One of the interesting insights from it is how difficult good application of AI is.

A lot of companies are just "deploying a chatbot" and some of the results from this study show that this doesn't work very well. My experience is similar: deploying simple chatbots to the enterprise doesn't do a lot.

For things to get better, two things are required, neither of which are easy:

- Integration into existing systems. You have to build data lakes or similar system that allow the AI to use data and information broadly across an enterprise. For example, for an AI tool to be useful in accounting, it's going to need high quality data access to the company's POs, issued invoices, receivers, GL data, vendor invoices, and so on. But many systems are old, have dodgy or nonexistent APIs, and data is held in various bureaucratic fiefdoms. This work is hard and doesn't scale that well.

- Knowledge of specific workflows. It's better when these tools are built with specific workflows in mind that are designed around specific peoples' jobs. This can start looking less like pure AI and more like a mix of traditional software with some AI capabilities. My experience is that I sell software as "AI solutions," but often I feel a lot of the value created is because it's replacing bad processes (either terrible older software, or attempting to do collaborative work via spreadsheet), and the AI tastefully sprinkled throughout may not be the primary value driver.

Knowledge of specific workflows also requires really good product design. High empathy, ability to understand what's not being said, ability to understand how to create an overall process value stream from many different peoples' narrower viewpoints, etc. This is also hard.

Moreover, this is deceiving because for some types of work (coding, ideating around marketing copy) you really don't need that much scaffolding at all because the capabilities are latent in the AI, and layering stuff on top mostly gets in the way.

My experience is that this type of work is a narrow slice of the total amount of work to be done, though, which is why I'd agree with the overall direction this study is suggesting that creating actual measurable major economic value with AI is going to be a long-term slog, and that we'll probably gradually stop calling it AI in the process as we attenuate to it and it starts being used as a tool within software processes.

mediaman··on OpenAI releases image generation in the API
It is correct, the shift from diffusion to transformers is a very, very big difference.
mediaman··on Teaching LLMs how to solid model
I've created (small, toy) transformers and also modeled injection molded parts in Solidworks.

There is a really big difference. It's obvious how programming languages can use tokens for an attention mechanism, which gives them excellent ability to have parallelized performance (versus RNNs, the prior way of relating tokens) with much broader ability to maintain coherence.

I don't know the parallel with brep. What are the tokens here? It's a fundamental architectural question.

But unlike four or five years ago for programming, when the attention mechanism was clear with transformers and the answer was, basically, just "scale it up", we don't even really know where to begin here. So assuming some magic is going to happen is optimistic.

It'd be exciting if we could figure it out. Maybe the answer is that we do away with brep and all tree based systems entirely (though I'm a little unclear how we maintain the mathematical precision of brep, especially with curves, which is necessary for machining -- your machinist is going to throw you out if you start giving them meshes, at least for precision work that has anything with a radius).

mediaman··on Ask HN: My CEO wants to go hard on AI. What do I do?
That's a good point. Lots of software companies are trying to become AI companies, not because they believe they are truly AI, but because CIOs are willing to invest in AI, so everything must become something that can fit in that budget.

This will crash in 1-2 years as CIOs realize most of this is useless, because at that point all software will be "AI Software" according to the marketing department, and we'll reset to CIOs making decisions on some other metric, such as a guy they met at a conference.

mediaman··on You Need Customers to Succeed in Small Business
That doesn't mean anything. You can accomplish that by having a very large denominator, which is possible if creating a business is trivially easy.

In the US, it's generally very easy to create a "business," which doesn't even need to be a separate legal entity.

I'd guess that a substantial portion of the companies in the denominator have no employees and are just entities for owning a piece of property, or for someone's occasional graphic design work, or whatever. They are legally businesses, but not businesses in the way that we commonly think, such as a local auto shop or retail store.

mediaman··on Teaching LLMs how to solid model
The problem is that this isn't very useful except for the very earliest ideation stages of industrial design, which hardly need CAD anyway.

Most parts need to fit with something else, usually some set of components. Then there are considerations around draft, moldability, size of core pins, sliders, direction of ejection, wall thickness, coring out, radii, ribs for stiffness, tolerances...

LLMs seem far off from being the right answer here. There is, however, lots to make more efficient. Maybe you could tokenize breps in some useful way and see if transformers could become competent speaking in brep tokens? It's hand-wavy but maybe there's something there.

Mechanical engineers do not try to explain models to each other in English. They gather around Solidworks or send pictures to each other. It is incredibly hard to explain a model in English, and I don't see how a traditional LLM would be any better.

mediaman··on An update to our pricing
I find Cursor annoying too, with dumb suggestions getting in the way of my attempts to tab-indent. They should make shift-tab the default way to accept its suggestion, instead of tab, or at least let shift-tab indent without accepting anything if they really want to keep tab as default autocomplete.
mediaman··on Gemini 2.5 Flash
This was solved a hundred years ago.

It's the same problem factories have: they produce a lot of parts, and it's very expensive to put a full operator or more on a machine to do 100% part inspection. And the machines aren't perfect, so we can't just trust that they work.

So starting in the 1920s Walter Shewhart and Edward Deming came up with Statistical Process Control. We accept the quality of the product produced based on the variance we see of samples, and how they measure against upper and lower control limits.

Based on that, we can estimate a "good parts rate" (which later got used in ideas like Six Sigma to describe the probability of bad parts being passed).

The software industry was built on determinism, but now software engineers will need to learn the statistical methods created by engineers who have forever lived in the stochastic world of making physical products.

mediaman··on Gemini 2.5 Flash
That's really cool. I've been looking for a nicer solution to use with nvim.
mediaman··on Cursor IDE support hallucinates lockout policy, causes user cancellations
I'd add that the deductive abilities translate to well-defined spec. I've found it does well when I know what APIs I want it to use, and what general algorithmic approaches I want (which are still sometimes brainstormed separately with an AI, but not within the codebase). I provide it a numbered outline of the desired requirements and approach to take, and it usually does a good job.

It does poorly without heavy instruction, though, especially with anything more than toy projects.

Still a valuable tool, but far from the dreamy autonomous geniuses that they often get described as.

mediaman··on Pentagon to terminate $5.1B in IT contracts with Accenture, Deloitte
Agreed. Many people I know who work in state government are always on or planning their next vacation, or in the middle of another home renovation, and lightly joke about their “work from home.”

It’s just incredibly clear that there is no inventive to actually perform, and combined with environments such as WFH that need to be outcomes-based, it starts just looking like a fairytale life paid for by other people.

This happens in the private sector as well, but tends to correct over time - hence big tech’s layoffs, especially of managers and nontechnical staff, after the excesses of 2021-2022.

I don’t think this is true of all state employees, but it seems to happen more with the educated professional class with more abstract work.

mediaman··on Basic Income Pilot Project: Study results
This only seems to be an issue if the marginal cost of building is too high to expand supply, though, right? Otherwise if people have more money, they bid up housing, then new stuff gets built, and supply profits decline.

Of course, restricting supply is a problem. I also think this logic might break down in tightly restricted areas that are already vertically built out - Manhattan, for example, because costs per housing unit tend to follow a U curve with respect to height, where they decline with density but start increasing again at very high density, but that's not an issue in most places.

mediaman··on Apache ECharts
Agreed - geopolitics aside, it's been awesome to see how much Chinese companies have embraced open source in a number of different domains. I think the whole ecosystem is going to greatly benefit from the contributions of Chinese open source projects.
mediaman··on A startup doesn't need to be a unicorn
Your comment seems to suggest that you don't see any difference between "capital efficiency" and "having capital."

The terms mean very different things. Capital efficiency is measured by metrics such as the cash conversion cycle. It's possible to design a business model in such a way that you have negative cash conversion cycles, which cause you to actually generate cash as a function of growth (even when unprofitable by GAAP!), which is the opposite of most VC funded businesses whose burn rate is roughly a function of their growth rate.

mediaman··on A startup doesn't need to be a unicorn
There is a happy medium between "AI eating the world" and "AI is a scam," that is generally best described as "AI performing valuable but often boring services within enterprises."

We don't often see that because it gets totally drowned out between cynics and moralist scolds battling vapid hype-bros and pumpers, and also because it's boring.

It turns out there's actually a lot of tedious classification, OCR, entity extraction, and enterprise search to be done.

Amusingly much of the money made in SaaS was also in boring products!

mediaman··on A startup doesn't need to be a unicorn
Yes, this is what every venture capitalist says. You aren't doing a startup unless you want extreme growth, which requires our services and cuts us in. Building in a capital efficient way that generates substantial wealth for founders, but without giving VCs a cut of the pie, is, of course, "not a real startup," and often also slandered as a "lifestyle business" for low-ambition people.

PG is great in many ways but he's not the person I'd turn to for an unbiased opinion on what counts as a "startup."

The founders I'm particularly impressed with are the ones who have such a nuanced understanding of capital efficiency that they do not require VC, and only take money much later in the cycle when they can basically dictate terms and want hundreds of millions for liquidity or whatever (see, e.g., Joe Mansueto).

mediaman··on Reasoning models don't always say what they think
The context is whether this capability is required to qualify as AGI. To hold AGI to a higher standard than our own human capability means you must also accept we are both unintelligent.
mediaman··on We hacked Gemini's Python sandbox and leaked its source code (at least some)
This is an unusual opinion in industry, although common with consumers.

Currently, Google has the most cost effective model (Flash 2) for tons of corporate work (OCR, classifiers, etc).

They just announced likely the most capable model currently in the market with Gemini 2.5.

Their small open source models (Gemma 3) are very good.

It is true that they've struggled to execute on product, but the actual technology is very good and getting substantial adoption in industry. Personally I've moved quite a few workloads to Google from OpenAI and Anthropic.

My main complaint is that they often release impressive models, but gimp them in experimental mode for too long, without fully releasing them (2.5 is currently in this category).

mediaman··on The curious surge of productivity in U.S. restaurants
Most of the profession of economics in the last several decades has been focused on figuring out how to measure difficult-to-measure truths about how people behave. That's what behavioral economics is about. The other name on the UChicago institute you don't like, Gary Becker, won a Nobel for finding ways to creatively study behaviorism: family structure, why people commit crime, studying societal benefits from education, and quantifying damage of discrimination.

You might argue that we shouldn't measure things at all -- that it is better to live in darkness and ignorance, and make vague gestures about the way things are with no evidence or rigorous thought. Measuring is how we find truth. Sometimes it is hard. But throwing away our measuring tools is not the path to enlightenment.

Many people choose to ex ante dislike economists, and dismiss anything any of them have to say without further study, because there is a risk that the results of quantifying the world might disagree with their preconceptions.

mediaman··on The curious surge of productivity in U.S. restaurants
No, they're not glossing over inflation, because they were able to explain nearly the entire difference in real sales per employee between restaurants during the same time as a function of dwell time.

They are not just looking at an overall increase in sales, misunderstanding sector-specific inflation, and saying it must be dwell time.

It's a good paper! I would encourage reading it. They have specific dwell time data per location, along with sales data, and explain the difference in productivity by looking at differences in productivity between restaurants over the same time period, find that it is determined by changes in dwell time, and then find that the overall mix shift of reduced dwell time explains the overall change in real sales per employee across the sector. Inflation is not the explanation.

mediaman··on The curious surge of productivity in U.S. restaurants
If you read the paper, they explicitly define "productivity" as real sales per employee. Your point about service versus food would be valid except that, apparently, they weren't being paid well for the service, because real sales went up per employee as a function of reduced dwell time. That means that service was a low productivity offering.

That you can explain the increase in productivity by shift in mix from low-productivity to high-productivity products is, in fact, the entire point of the paper, and does not mean that a "surge in productivity" did not happen.

The authors point out that a change in consumer demand preferences is in fact a type of "technology" in that it can be a driver of sector productivity. Not all technology lives on circuit boards!

mediaman··on The curious surge of productivity in U.S. restaurants
The paper does not cover "mid-tier dining." It covers fast food because that's where they had good data.

Moreover, real restaurant spending in total is up, so while your personal dining decisions are interesting they don't appear to reflect the aggregate data.

I'm also not sure why you are calling Goolsbee a "Friedmanite." He served in the Obama administration and advocated strong government intervention during the financial crisis. Do you think every UChicago economist is a Friedman disciple?

Lastly, where did you infer the economists are "lauding" the productivity increases? It appears nowhere in the paper. They only defined productivity increases and explained why they occurred. Was there some other reason you wanted them to be lauding it, other than to justify your anger with people whose policy beliefs you don't seem to be familiar with?

mediaman··on Musk-led group makes $97B bid for control of OpenAI
Using BlackRock as a comparison is really disingenuous. They're basically a Vanguard. They hold assets as trustee for the actual owners of the wealth. BlackRock and Vanguard can't wake up one day and decide to put in a bid for OpenAI because of some personal vendetta.

You are confusing "assets under management" (AUM) with actual net worth. Blackrock's net worth is $43 billion, compared to $11 trillion in AUM, because 99.6% of the top-line number you chose is actually just client money. Musk, meanwhile, is around $400 billion, so you have the numbers backwards; Musk has about 10x the net worth of the nation's largest fund manager.

mediaman··on Ingesting PDFs and why Gemini 2.0 changes everything
Agree, BM25 honestly does an amazing job on its own sometimes, especially if content is technical.

We use it in combination with semantic but sometimes turn off the semantic part to see what happens and are surprised with the robustness of the results.

This would work less well for cross-language or less technical content, however. It's great for acronyms, company or industry specific terms, project names, people, technical phrases, and so on.

mediaman··on Ingesting PDFs and why Gemini 2.0 changes everything
Everything I tried previously had very disappointing results. I was trying to get rid of Azure's DocumentIntelligence, which is kind of expensive at scale. The models could often output a portion of a table, but it was nearly impossible to get them to produce a structured output of a large table on a single page; they'd often insert "...rest of table follows" and similar terminations, regardless of different kinds of prompting.

Maybe incremental processing of chunks of the table would have worked, with subsequent stitching, but if Gemini can just process it that would be pretty good.

mediaman··on OpenAI O3-Mini
You don't need to wait for Google. Their Jan 21 checkpoint for their fast reasoning model is available on AIStudio. It shows full reasoning traces. It's very good, much faster than R1, and although they haven't released pricing, based on flash it's going to be quite cheap.
mediaman··on OpenAI O3-Mini
It's not clear that writing poetry is a bad use case. Reasoning models seem to actually do pretty well with creative writing and poetry. Deepseek's R1, for example, has much better poem structure than the underlying V3, and writers are saying R1 was the first model where they actually felt like it was a useful writing companion. R1 seems to think at length about word choice, correcting structure, pentameter, and so on.
mediaman··on DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
You can just use it and see for yourself. It's quite good.

I do believe they were honest in the paper, but the $5.5m training cost (for v3) is defined in a limited way: only the GPU cost at $2/hr for the one training run they did that resulted in the final V3 model. Headcount, overhead, experimentation, and R&D trial costs are not included. The paper had something like 150 people on it, so obviously total costs are quite a bit higher than the limited scope cost they disclosed, and also they didn't disclose R1 costs.

Still, though, the model is quite good, there are quite a few independent benchmarks showing it's pretty competent, and it definitely passes the smell test in actual use (unlike many of Microsoft's models which seem to be gamed on benchmarks).

← PreviousPage 8 of 34Next →