I don't think your example refutes the fairness heuristic at all.

The first case is completely fair because anybody else could have done the same thing without any special access required.

The second case is unfair because you had to work at the company to get access.

Okay, then imagine you overhear at a bar. Yes “anyone could have” theoretically, but not actually. In either case, you have material non-public information that your counterparty in the market does not.

you got that piece of non public information was not because you are an insider. As long as the bar is not exclusive to insider, i don't see any difference

Isn't it exclusive to people who live in the area of the bar?

What if the bar has a cover charge, so only those who pay get in?

What if the cover charge is $10,000 and the bar is advertised as "the place where public company execs love to come talk to each other about private deals"?

How should he have acted instead?

Yeah.

“Dario is saying the right thing and doing the right thing and not ever acting otherwise, but I think it’s just performative so I’m still disappointed in him.”

We don't know how the military intended to use Claude, and neither do we know nor does the military know whether Claude without RLHF-imposed safety would have been more useful to them.

Ergo, this is a very convenient PR opportunity. The public assumes the worst, and this is egged on by Anthropic with the implication that CLAUDE is being used in autonomous weapons, which I find almost amusing.

He can now say goodbye to $200 million, and make up for it in positive publicity. Also, people will leave thinking that Claude is the best model, AND Anthropic are the heroes that staved off superintelligent killer robots for a while.

Even setting this aside, Dario is the silly guy who's "not sure whether Claude is sentient or not", who keeps using the UBI narrative to promote his product with the silent implication that LLMs actually ARE a path to AGI... Look, if you believe that, then that is where we differ, and I suppose that then the notion that Amodei is a moral man is comprehensible.

Oh, also the stealing. All the stealing. But he is not alone there by any means.

edit: to actually answer your question, this act in itself is not what prompted me to say that he is an immoral man. Your comment did.

> to promote his product with the silent implication that LLMs actually ARE a path to AGI

That isn't implied. The thought process is a) if we invent AGI through some other method, we should still treat LLMs nicely because it's a credible commitment we'll treat the AGI well and b) having evidence in the pretraining data and on the internet that we treat LLMs well makes it easier to align new ones when training them.

Anyway, your argument seems to be that it's unfair that he has the opportunity to do something moral in public because it makes him look moral?

His actions seem pretty consistent with a belief that AI will be significant and societally-changing in the future. You can disagree with that belief but it's different to him being a liar.

The $200m is not the risk here. They threatened labelling Anthropic as a supply chain risk, which would be genuinely damaging.

> The DoW is the largest employer in America, and a staggering number of companies have random subsidiaries that do work for it.

> All of those companies would now have faced this compliance nightmare. [to not use Anthropic in any of their business or suppliers]

... which would impact Anthropic's primary customer base (businesses). Even for those not directly affected, it adds uncertainty in the brand.

Because the people who are consistently right will consistently win money and will make bigger bets which move the price more, in the limit case making the price converge on the true probability of the outcome.

This is the theoretical underpinning of prediction markets.

Equating being "consistently right" with having a sufficiently large stash of capital is ludicrous.

"right" people will wisely take most their winnings out of a high-variance market. "wrong" people with deep pockets (or lots of wrong people with shallow pockets) will continue to distort the market.

> will continue to distort the market.

they can only do so as long as they have enough capital to lose. Because every time they try to move the betting markets against the truth, they will simply lose that money when the event happens (and turns out they were wrong).

So any distortion will merely be temporary. Unless they have access to unlimited capital of course - which is not true yet for anyone (but the US gov't).

That only makes sense in a hermetically sealed system, which this is very much not.

Yes, but is this a problem? Haven't most betting markets turned out to offer accurate predictions?

Not particularly so. But even if it were, would that justify the social cost of this kind of gambling?

Well, the more often you're right, the more capital you will be able to accrue to bet with next time.

This is an unnecessarily cynical view.

People are excited about AI because it's new powerful technology. They aren't "pandering" to anyone.

I have been in dozens of meetings over the past year where directors have told me to use AI to enable us to fire 100% of our contract staff.

I have been in meetings where my director has said that AI will enable us to shrink the team by 50%.

Every single one of my friends who do knowledge work has been told that AI is likely to make their job obsolete in the next few years, often by their bosses.

We have mortgages to pay and children to feed.

People are afraid because they need to work to eat. People who don't need to work to eat are less likely to be afraid.

I have yet to meet anyone except managers be excited about LLM's or generative AI.

And the only people actually excited about the useful kinds of "AI", traditional machine learning, are researchers.

You don' have to look past this very forum, most people here seem to be very positive about gen AI, when it comes to software development specifically.

Lots of folk here will happily tell you about how LLMs made them 10x more productive, and then their custom agent orchestrator made them 20x more productive on top of that (stacking multiplicatively of course, for a total of 200x productivity gain).

I assume those people are managers, have a vested interest in AI, or have only just started programming.

How would you find out if you were wrong?

You're presented with hundreds of people that prove you wrong, and your response is "no, I assume I'm right"?

This is obviously a rhetorical statement. I'm not claiming a categorical fact, but a fuzzy one.

Most of these peoples are managers, investors, or junior.

I don't know what is your bubble, but I'm a regular programmer and I'm absolutely excited even if a little uncomfortable. I know a lot of people who are the same.

Interesting, every developer I've spoken to is extremely skeptical and has not found any actual productivity boosts.

Ok that's not true. I know one junior who is very excited, but considering his regular code quality I would not put much weight on his opinion.

I am using AI a lot to do tasks that just would not get done because they would take too long. Also, getting it to iterate on a React web application meant I can think about what I want it to do rather than worry about all the typing I would have to do. Especially powerful when moving things around, hand-written code has a "mental load" to move that telling an AI to do it does not. Obviously not everything is 100% but this is the most productive I have felt for a very long time. And I've been in the game for 25 years.

Why do you need to move things around? And how is that difficult?

Surely you have an LSP in your editor and are able to use sed? I've never had moving files take more than fifteen minutes (for really big changes), and even then most of the time is spent thinking about where to move things.

LLM's have been reported to specifically make you "feel" productive without actually increasing your productivity.

I mean there are two different things. One is whether there are actual productivity boosts right now. And the second is the excitement about the technology.

I am definitely more productive. A lot of this productivity is wasted on stuff I probably shouldn't be writing anyways. But since using coding agent, I'm both more productive at my day job and I'm building so many small hobby projects that I would have never found time for otherwise.

But the main topic of discussion in this thread is the excitement about technology. And I have a bit mixed feelings, because on one hand side I feel like a turkey being excited for the Thanksgiving. On the other hand, I think the programming future is bright. there will be so much more software build and for a lot of that you will still need programmers.

My excitement comes from the fact that I can do so much more things that I wouldn't even think about being able to do a few months ago.

Just as an example, in last month I have used the agents to add features to the applications I'm using daily. Text editor, podcast application, Android keyboard. The agents were capable to fork, build, and implement a feature I asked for in a project where I have no idea about the technology. Iif I were hired to do those features, I would be happy if I implemented them after two weeks on the job. With an agent, I get tailor made features in half of a morning. Spending less than ten minutes prompting.

I am building educational games for my kids. They learn a new topic at school? Let me quickly vibe the game to make learning it fun. A project that wouldn't be worth my weekend, but is worth 15 minutes. https://kuboble.com/math/games/snake/index.html?mode=multipl...

So I'm excited because I think coding agents will be for coding what pencil and paper were for writing.

I don't understand the idea that you "could not think about implementing a feature".

I can think of roughly 0 fratures of run-of-the-mill software that would be impossible to implement for a semi-competent software developer. Especially for the kinds of applications you mention.

Also it sounds less like you're productive and more like the vibeslop projects are distracting you.

I'm claiming it's both.

I produce more good (imo) production features despite being distracted.

The features I mention is something that I would be able to do, but only with a lot of learning and great effort - so in practical terms I would not.

It is probably a skill issue but in the past many times I downloaded the open source project and just couldn't build and run it. Cryptic build errors, figuring out dependencies. And I see claude gets the same errors but he just knows how to work around those errors. Setting up local development environment (db, dummy auth, dummy data) for a project outside of my competence area is already more work than I'm willing to do for a simple feature. Now it's free.

>I can think of roughly 0 fratures of run-of-the-mill software that would be impossible to implement for a semi-competent software developer.

Yes. I'm my area of competence it can do the coding tasks I know exactly how to do just a bit faster. Right now for those tasks I'd say it can one shot code that would take me a day.

But it enables me to do things in the area where I don't have expertise. And getting this expertise is very expensive.

Out of interest, could you give me an example of a feature that it one-shotted that would have taken you a whole day?

The example from yesterday:

I have a large C# application. In this application I have a functionality to convert some group of settings into a tree model (a list of commands to generate this tree). There are a lot of weird settings and special cases.

I asked claude to extract this logic into a separate python module.

It succesfully one-shot that, and I would estimate it as 2 days work for me (and I wrote the original C# code).

This is probabaly the best possible kind of task for the coding agents, given that it's very well defined task with already existing testcases.

Seems reasonable, but if it's just copy pasting, doesn't seem like that would take you a whole day. Maybe on the order of an hour at most.

Were you exaggerating earlier or do you have more examples?

This is a two-day task for me. If you could do it in one hour, then you're a 10x programmer compared to me.

You can browse the code at <my_username>.com/slop/hn_tb/

I have also sloped the simple code viewer. So you can make your judgement if it looks like 1 hour task.

This is nothing to do with politicians.

I wanted to ssh hello-hacker-news@knock-knock.net, but sadly it doesn't work because the site is hosted behind Cloudflare. You'd have to work out the true server IP address which is not easy to do.

Sadly? Intentionally! The IP is hiding behind Cloudflare mainly to make it much harder for the bots to figure it out. Blocking you from messing with the stats is just icing on the cake. :-)

I don't think hosting the site behind Cloudflare will affect the number of SSH brute-force attempts, these bots are just brute-forcing the entire IPv4 space aren't they?

Exactly right. I just didn’t want the bad actors to know my own IPv4 address so that they might try to cause havoc or treat my site differently.

Roger.

And cool project btw.

I always wonder what the pen plotter is adding?

You can look at SVG lineart on the screen without plotting it, and if you really want it on paper you can print it on any printer.

And particularly:

> This was an experiment I would like to push further. I would like to reduce the feedback loop by connecting Claude directly to the plotter and by giving it access to the output of a webcam.

You can do this in pure software, the hardware side of it just adds noise.

"You can do this in pure software, the hardware side of it just adds noise."

That "noise" changes the context, connects it to different parts of the training corpus.

Removing the "physical art" part would likely change the responses to be much more technical (because there is way more technical talk surrounding SVGs) and less art-critic (there is more art-critic talk around physical art).

This is art though. Whether you like the results or not, I'd say that the OP is using tools to make visual art but also that the process is part of the art as well. The process of art making doesn't have to be optimized - especially for the latest technology. We still paint when we have photography, we still make darkroom prints when we have color screens, etc.

Sure, you could just do it in software. Maybe it would produce something interesting though, to have that extra layer through the physical world?

It does. It makes for a more catchy title and feeds into illusions of it understanding something about the world.

Didn't it notice in the photo that the opacity of the lines was 100% IRL?

> While this entity-attribute-value model enables the developer to break out from the structure imposed by an SQL database, it loses out on all the benefits,[1] since all of the work that could be done efficiently by the RDBMS is forced onto the application instead. Queries become much more convoluted,[2] the indexes and query optimizer can no longer work effectively, and data validity constraints are not enforced. Performance and maintainability can be extremely poor.

This is only true if you try to do this for all of your data.

I've used key-value tables loads of times, it's convenient for storing things like global configuration.

What else can you do? Make a table that has every configuration value as a separate column and populate it with only a single row? That seems absurd and worse.

> What else can you do? Make a table that has every configuration value as a separate column and populate it with only a single row?

If the values are singletons, what you're describing is the most efficient. What the author is describing is a surprisingly common anti-pattern where someone has a table with three columns: entity id, property name, property value. Almost like a graph database. Fetching data for one entity (normally one row in a properly built db) is fine, but fetching the data for multiple entities is instantly a mess.

It's worth noting, though, that unless the configuration values are all the same type, you lose type safety with just one column for values. "I'll parse the data as JSON" means your service will fail hard at runtime if someone changes the configuration and uses invalid data.

Personally I’d pass all my JSON-y stuff through Pydantic or similar which would take care of these concerns.

It wouldn't. If you load configuration and it's the wrong type, now Pydantic is throwing an error. You can't run code if the config is invalid no matter what tools you use: failing validation is still an error.

This paragraph rubbed me the wrong way too.

> work that could be done efficiently by the RDBMS the RDBMS? You're only using one? Why not spread the work out a little? Even if you think you write all your queries efficiently, nothing stops your teammates from DOS'ing your efficient queries by writing inefficient queries themselves. Last week our team started started piling up write timeouts because another team was modifying one of their tables. Not in their db, in the db.

> Queries become much more convoluted Please, every ounce of effort invested in ORMs like EF/LINQ is to make code look less like querying and more like plain old object access. For the most part, devs want to work with objects and store objects. If you didn't go the RDBMS route, you wouldn't need EF/LINQs help in decomposing your objects and scattering their parts into separate tables. The least convoluted query possible is to just grab the object you wanted directly.

Make a table with a key and a json column. Use this table as key-value store.

Omit the key and store it in a single json column. The worst of all worlds!

If the json has a natural key, you can add a json index pointing to it, at least in postgres. But this will probably have worse performance characteristics than adding a separate key column.

Also natural keys suck - if a bugged json comes with repeated key (or no key at all) you will be unable to save it to process it further (well unless you have a separate table to save them). Also some jsons won't be keyed

But storing something with no index at all is probably okay if you don't plan to query much (or if the dataset is tiny)

If you think it's so easy, you can quit and start your own company and find out.

Unlike buying a house, writing software can be done without up front investment in one's spare time

Turns out the value of it actually isn't just the code though...

> without up front investment

If you don’t count a computer and the time it takes to learn.

I'm assuming people comparing only earning a Bay Area salary to chattel slavery have reached that point already

But that doesn't mean the average person will produce all that stuff from a standing start within 12 months just because they would like to drill a hole.

Nah, they just bought their drilling tool from whatever was their local version of Walmart in that ancient era.