> There's no lint or compiler that can check for correctly constructed contracts.

There are definitely linters and this exists https://catala-lang.org/

> What I don't see is vast areas of industry finding $10s to $100s of billions of value in LLMs.

Translation between languages.

That value dwarfs all programming value that can be had. Economically, culturally, scientifically, spiritually.

Between human languages? The problem will be Google is giving that away for free.

The quality of Google Translate is so low that it isn't even in the picture.

You can do that with single words or phrases.

Now do a full movie's subtitles with google translate, in an automated fashion. So that it understands the context and still translates it correctly.

Social issues aside, that's a tiny slice of the economics of AI.

Just by nature of how often it's used etc.

You need workloads for AI be cost effective: software, automation etc.

International trade is a giant sector already, and potentially much bigger than that, now that LLM assisted translation makes it easy to offer your products and services in any language, or make complicated and sensitive deals without a common language.

Hackers always rage and down vote every time I mention this, because they are unable to see beyond their small world. Why didn't they learn that their part of the internet is 0,000000001% of what the world uses the internet for today. It's going to be the same with LLMs. Programming and hacker stuff is going to be 0,0000000000000000000000000000001% of what the world uses AI for. But translation is going to be in the top 5 of use cases.

I think quite cheap models will probably work with translation (I think for many use cases Google's now fairly old tech works well enough and is free?).

The main problem is I think you're assuming because the current translation market is large (I'm just going to assume it's ~100B in size just from a cursory search), then it will remain large with LLMs. If LLMs are much cheaper than humans, even with a lot of growth in translation volume the total spend may not compensate for it (again most of the volume will probably be using almost free models?). Another is assuming that because something is valuable you can charge a lot for it. Like, oxygen from air is extremely valuable to us. If oxygen somehow depleted we would die almost instantly. It does not mean everyone goes around purchasing oxygen or even less that you can charge absurd amounts for it.

LLMs are creating a new translation market.

Before AI translation became available, I would hire professional translators. Their rate was about $50-100 for a detailed product page into one language, and it would take them a few days to deliver.

Now with AI translation, I pay about $120 per year for unlimited translation. Meaning dozens of product pages into 5 languages, plus e-mail back and forth with hundreds of customers. All instantly at my convenience.

What this means is that a whole lot of people, sectors and businesses who would never hire professional translator can now have high quality translation at their disposal for a cheap price.

> Another is assuming that because something is valuable you can charge a lot for it.

Even if LLM translation won't deliver trillions of dollars in income to the AI companies, it will without a doubt deliver trillions of dollars in value to customers and users within the coming few years.

Good human translators are still higher in quality than any AI and will always be. But AI translation is currently far beyond good enough for all use cases, except fine literature and maybe complicated juridical stuff. But I'm not familiar with those sectors.

I think OP's comment is about volume/scale more so than practical application.

Product marketing and contracts will be a 'common use case' we think of, but it will be 0.000001% of tokens consumed.

A developer using sub-agents will consume more tokens 1 Day than a marketing manager will consume 1 Month, easily.

Unless there is something inherently automated about the nature of the AI, it will be a tiny % use case.

Even a lawyer, using AI daily for contracts - that will be relatively light use. They'll make more use doing legal research etc.

Developers and Automation are the 'primary' uses cases for AI, and in the future, we'll start to see AI integrated into Apps - that will be 85% of tokens consumed.

Yes - once translation becomes realtime, and we have our Star Trek Universal Translators, then translation will become more visible, but even by then, a relatively small part of overall consumption, even if it's more highly visible.

If a salesperson uses 100 000 tokes for translation, which gives him $10 000 in sales.

And a programmer uses one million tokes, which gives him $100 in sales (or value).

What is then the value of a token?

The comment I answered asked where there is an industry finding 10s or 100s of billion of dollars in value from LLMs. The answer is translation. It's the value they as customers get out of the LLMs, not what cost they are paying for the LLMs.

Value has to be counted in production, not consumption.

> Developers and Automation are the 'primary' uses cases for AI

Just like programmers and scientists were the primary users of the Internet when it began. But things change rapidly.

1) A sales person won't make $10K from a little translation - we can already automate basic translation using tech and humans.

2) 'Because there was a bit of translation involved in the job' - does not mean that AI is really driving that economic activity.

Again: translation is extremely common. The AI might make that part a bit more efficient, it will be barely noticeable.

3) That level of AI activity will just be some 'compute' on your local machine, nobody will be counting tokens.

Think of it like this:

"Translation is going to get easier because you can download some software that does it". It happens to use the GPU instead of CPU.

The vast majority of this kind of compute will be automation, software and a few other things.

'Translation' will just be a little app on your iPhone.

Of course a sales person can make $10k from a transaction. It depends entirely on what he is selling. But you can split it into smaller chunks if you so wish. Since unlimited high quality LLM translations are about $120 per year right now.

But why are you talking about "basic translation"? There is no commercial need for that. It has to be high quality, and LLMs deliver that. Human translators deliver that as well, but at a much higher price.

Your number 2 point, what do you mean? In many cases if there was no translation, then there wouldn't have been a sale. It's a direct driver.

Your number 3 point: No matter how cheap it is for the user, the economic value from LLM translation will still be counted in the trillions of dollars. Just like e-mail (which is cheap or free for the user). But you will need an enormous amount of storage space on your local machine to have perfect quality translation between dozens of languages.

I don't know why hackers want to diminish the value of LLM translations. The commercial need for translation is not for finding out what a word means. It needs to be of the highest quality, comparable to professional human translators.

But translations don't require anything close to SOTA level models. Translations will be high volume, low margin transactions. That will not save Anthropic.

Maybe not Anthropic, but LLM translations has a value counted in the trillions of dollars easily.

I see it's important but trillions? How did you quantify that number?

Tourism and travel is 10% of global GDP, or about 10 trillion dollars. And it's not the only sector where translation between languages is a benefit. However, I am not sure that the trillions of dollars of value from LLM translation in the coming few years will translate into trillions of dollars of income for the AI companies.

Yep translation is going to be a free feature on all devices using on-device LLMs. I see no revenue flowing to AI companies.

Commercial grade translation between dozens of languages would take enormous amounts of disk space.

But we'll see. Text editing tools are included on all digital devices, yet companies pay for commercial solutions like MS Office. Cameras and basic image editing tools are included on all smart phones, yet there are millions of advertising agencies around the world. And so on.

LLM frontier companies are more akin to infrastructure like chips and airplanes and the internet than applications built on top of it. There will be products built on top of them that generate revenue.

eg. processors are needed for everything on the planet. No Intel isn't going to generate annual revenue in the trillions because of that. In the case of frontier companies the moat is even smaller with dozens of players competing. LLMs will become a low-cost commodity. You meanwhile can make billions building applications using them though.

> eg. processors are needed for everything on the planet. No Intel isn't going to generate annual revenue in the trillions because of that.

No, but Intel processors are going to generate revenue in the trillions for the economy, which was the original question as far as I understood it.

A model bundled in your phone can do the translation, you're thinking too much of it.

A model bundled in your phone cannot do commercial grade quality of translations.

While there's no linter for writing contracts, my experience (as a commercial lawyer) is that frontier LLMs are far better and error checking and far quicker at writing than the average senior lawyer. The main thing holding back further deployment (in my jurisdiction) are concerns around data residency, privilege and how fundamentally it will break an industry that is so heavily reliant on time based billing.

Trouble is they're unreliable.

I compared insurance quotes last week. Needed cover for 2 brands, 1 company. Opus jumps up and down saying both brands need listing on the policy schedule. Human broker said not.

I told opus and it's the usual "thanks you're right" bollocks because it bothered to read in more detail and found that all business activities are covered.

I agree with you in respect to the greatest risks in that industry. Just thought it's worth mentioning that LLM's unsurprisingly are not immune to hallucinating cases. https://www.scientificamerican.com/article/why-lawyers-keep-...

Using them for coding makes it easy to self check its work (assuming those pieces of work are "verifiable").

As proofreading will still need to happen, what do you think the appetite for lawyers is to do this kind of work? Do you think it will drive fees down significantly? Empower younger lawyers at firms who probably are the ones doing this checking for the partners? (Or will that just create a further divide).

I'm genuinely asking as I am not in law but all my family is and it's nice to see someone here that's thought about the impacts in that space.

With access to good databases and tool use, hallucinations are basically a thing of the past (and very easy to verify). The hallucination issue typically comes from people using free ChatGPT or similar - tech usage and literacy is often poor and there's a lack of understanding of the risks of using something without proper legal database tools. Most lawyers would not have the faintest idea of what tool use in the context of an LLM even means. It's still seen as a magic box rather than a tool for real work.

Will it drive down fees significantly? Doubt it, there's not enough pressure on them, and the industry is resistant to change. Firms don't want to make a big deal about using it because clients will then ask why they're not getting a discount.

Will it empower younger lawyers? Not as much as I'd like. I'm very fortunate to have an employer that lets me use Claude Code for my work (to a limited extent). I think for 99.9% of lawyers it's not an option available to them, through a mix of concerns around AI usage and concerned IT departments. There could be great benefits, but it would rely on having to break out of traditional private practice which would make it difficult to get enough work.

I think in the near term LLMs will have much the same impact on the legal industry as it has on the software industry.

But isn’t “writing a linter for contracts” the thing we need and probably would solve?

Coding agents are great because they have compilers, linters, test cases etc to ground themselves in.

With tools like OKF I’m sure most knowledge work would be distilled to its core data - it’s AST if you will, and then allow models to guard against hallucinations.

Checking if a case law exists is a tool call, you can demand provenance, it’s all _buildable_.

Hallucinations are “solvable” this way, so the rest is just time and adoption…

> concerns around data residency

Funny you should say that. I read somewhere (not on HN, but I think it was a post linked from here) that a number of law firms who deal with extremely sensitive documents have started buying amped-up Macbooks with 512GB of memory to be able to run local models.

These are businesses who literally - and for once this word fits - cannot afford to let some of those documents get anywhere outside their corporate walls.

Macbook Pros max out at 128GB

Turns out the books do, studio setups can go up to 512GB.

There was a thread earlier this year about the option getting pulled from selection (https://news.ycombinator.com/item?id=47296302) but I'd guess you can still top the thing up yourself.

>my experience (as a commercial lawyer)

How do you think the industry will respond to contract clause slop ??

I can foresee each side inserting 100s of innocent clauses with minute dependencies that provide hidden advanatges

Good question - but the state of commercial contracts (pre-AI slop) is horrifically dysfunctional, inefficient, poorly thought through and inaccessible to the average person, so the bar is quite low. I think there will be some new malicious behaviours like that to keep an eye out for but I haven't seen them in practice (yet).

I have a family member that is an attorney in housing law. She claimed that LLMs are not particularly useful for her work. If she asked a simple question like, "Find all the <insert specific housing laws> for all 50 states," then she still has to go and check every single one of the laws. Since the legislature is modified so often, she cannot look at, say, Maryland's law and know if it the LLM output was the 1990, 2014, 2018, or 2026 version of the codified law. In order to fact check the law, she has to look it up, and by that point in time, she has the answer she did the work of the LLM.

This doesn’t sound as much that “LLMs can’t help me” as much as “trying to one-shot a complicated answer in a chatbot can’t help me”. Output straight from a generative model definitely should not be trusted.

But that doesn’t mean that an appropriate harness (even Claude code and a dynamic workflow) can’t complete, and validate, an answer.

Seems like it would be a massive market to make such a harness, and yet you're making it sound very easy to do. It would be surprising no one has done it if it were.

Harvey and Clio both make entire products around this. They have raised billions of dollars.

That's not really the point though; it doesn't have to be "legal specific".

The key is in understanding that "Basic chatbots (claude, chatgpt, etc) hallucinate. Independently verify the output (even if it's another model doing the verification) before assuming it's correct".

You can reproduce these results even if you're not a lawyer, if you've dabbled in any MMO that has been around for a few years and had multiple patches.

Ask any LLM about a random game mechanic and then be prepared to verify if the answer is for the patch in 2025, 2023 or 2020...

This is my experience if you need the data to be accurate often it isn't and you only know that because you had to check because it was important. Things not important you never check.

LLMs are not magic boxes, they hallucinate and deviate quite a lot when asked hard to verify questions.

The only way we got a head start of using it for coding and maths was to have some formal method of validating the output as part of the training and inference time.