HN Simulatornew | past | comments | lists | submit | dkersten's commentslogin

For the vast majority of people, neither will AI.

It will just push you to do cheap menial work that is too costly to use AI for, or ever more complex work that AI can’t do. The time where AI can do your job slowing you to sit back and relax is ephemeral and temporary because companies will not hesitate to have large layoffs whenever they feel the machine can do the work of the employees.

We already see it a bit. Most people using AI to write code aren’t getting to sit back and relax, they’re asked for more and are spending more and more time on tedious, frustrating, boring code reviews, or at least on QA and being an agent nanny, instead of the fun and interesting problem solving programmers used to do.

We’re not quite there yet but of models keep improving and are able to do more and more with less and less supervision, then it’s only a matter of time before we’re relegated to the less fun tasks. Employers aren’t going to keep paying is unless we justify our value somehow. The utopian view of AI is a fantasy.


For the vast majority of people, neither will AI.

So let's hear your plan for getting us to a post-scarcity economy.


You could make it a requirement only for people who want to attribute “made by”. If you publish a no name extension, you could skip it.


Most of them appear to be small LLM’s fine tuned for the role.

That’s a different set of properties in terms of size, cost, and latency. Jev (apparently, not like I’ve seen its insides) is extremely cheap, extremely fast, doesn’t cost any output tokens as it speaks the output natively, can’t get the output wrong because it speaks the format natively, and (presumably based on the docs), the context is separate from the question, meaning it should be immune (or at least highly resistant) to prompt injection attacks.

It’s not just about the accuracy of the result, it’s a collection of all the properties that make Jev interesting.

Jev took years to develop, I strongly doubt that a copycat that was put together within days after Jev’s release will be able to match it on a sun of its properties. Even if fine tuned LLMs can outperform it on raw accuracy.


Jev is something your favorite LLM could zero-shot months ago, if you pointed it to the right arXiv paper (some of which are linked in this thread).


That probably explains why there were so many competitors around withing days of the Jev announcement. They are not starting with a moat, and there doesn't seem to be any moat in sight. Just buzzword recognition because everything is comparing to "jev".


There will always only be a very small percentage of people who want to discover and build things, even with llms, its a very small subset of people though, and most people want off the shelf solutions.

Also Jevs purpose isnt to become its own thing. It will get aquired in 18 months by one of Andressen Horowitz's incestuous circle of companies and everyone will make money, and the person who buys it wont necessarily care if Jev itself makes them a ton of money. They're just passing chips around the table.


Jev is the known idea, but with a slick website with grandiose claims and an impressive Doom demo on. It's also a catchy name - it caught on instantly, but mostly as a shorthand: it's easier to say "Jev" and "like Jev" than "classifier models" (or rather " classifier models").


Jev did not take years to develop. What it does was published in arxiv back in 2025. TypeSafe just marketed it.


What specific arxiv paper are you referencing here?



Yeah, I figured it was gonna be this one.

"SalesRLAgent: A Reinforcement Learning Approach for Real-Time Sales Conversion Prediction and Optimization"

Jev is a general-purpose thing. That is a specific-purpose thing. General-purpose thing is not the same as specific-purpose thing. What makes people think these are the same thing? I don't get it.


What do you mean? Jev is trivially different from what is described in this paper.


Bullshit. Below is copypaste from the paper in the section that outlines the "key contributions" of the paper. As you can see, it is focused on one specific problem: predicting sales conversions. So if you were to take this system and use it for some other task ("evaluate customer mood" for example), it would not work. Because, again, it is not describing a general purpose solution. It is describing a solution that is specific to one problem: sales conversions.

Copypasta:

• A reinforcement learning architecture specifically designed for sales conversation analysis and conversion prediction

• A synthetic data generation pipeline leveraging GPT-4O to create diverse and realistic sales conversations

• Novel state representation techniques using Azure OpenAI embeddings (3072 dimensions) with sales-specific features

• A meta-learning approach enabling the system to express confidence in its predictions based on conversation similarity to training data

• Integration mechanisms providing real-time guidance within existing sales platforms

• Extensive comparative evaluation demonstrating significant performance improvements over LLM-based approaches


> Novel state representation techniques using Azure OpenAI embeddings (3072 dimensions) with sales-specific features

Jev is basically the embeddings side of an LLM. Yes, it's a good idea, but the moat is non-existent.


No. It can't simultaneously be both general purpose and having task-specific embeddings.


Hard to take you seriously, fam. The distance between these two things is trivial. The author of the paper implemented the "generalized" version of same in less than 24 hours. It's not novel, it is a continuation of what already existed in a boring way


The comment I was responding to describes Jev's significance as being specifically about embeddings:

> Jev is basically the embeddings side of an LLM

This is how the paper you are referencing is describing the "key contribution" as it relates to embeddings:

> Novel state representation techniques using Azure OpenAI embeddings (3072 dimensions) with sales-specific features

And now you are claiming that the author of the paper created:

> "generalized" version of same

It's a bit hard to guess what you are trying to say, but if I were to steelman your argument, I would guess that you mean: training a model with a vocabulary that has some tokens representing things like "OPTION_A", "OPTION_B" is in your mind the same thing as "generalized version of sales-specific features in embeddings"? Is this what you were trying to say?


I said the author of the paper created a generalized version because they posted about doing so on reddit. They way they described the work, it sounded trivial.


No, what you said was:

> What it does was published in arxiv back in 2025

This is false. You made a statement which was false.

If you want to say that the author of the paper made a really bad knockoff of Jev in 2026 after Jev was published, then say that.


This is a fine-tuned model. The author even states that the model is competitive with Jev only if fine-tuned on the evaluation at hand.

Literally misses the point of Jev, which you don't need to fine-tune to get accuracy nor - and no other model has this - some sort of out of sample calibration


> I strongly doubt that a copycat that was put together within days after Jev’s release will be able to match it on a sun of its properties

But why? If the simplest way to achieve Jev's capabilities (accuracy, cost, latency) is by fine tuning a small model, what makes you think that this isn't exactly what Typesafe did?

And even if they did something different - what makes you think it was a good idea in the first place, given how easy their results were replicated without any "secret sauce"?


My point is that it’s not replicated. You replicate the accuracy, but not the other properties. The Jev-competitors only proved that you can get or beat the accuracy, nothing about the other properties. Especially the “zero hallucination” output and the (if it works how the documentation make it sound) prompt injection resistant architecture. You can’t get that with a fine tuned LLM.


> zero hallucination

Plenty has been said about this claim. If you're still falling for this, I feel sorry for you.

If you remove wheels from your car, your car will get a "no speeding ticket" property, and yet there's nothing exciting about it.

> You can’t get that with a fine tuned LLM

Of course you can. All these claims are nothing but marketing.


I understand most major model providers support passing a JSON schema that is strictly followed in the output, accomplishing the same 'zero hallucination' and prompt injection resistance.

The only difference I am aware of is that probabilities are better calibrated with these decision models compared to regular LLMs which can output hallucinated numbers where your schema allows a number.


I’m not claiming you can’t get some of the properties in some ways with LLMs, just that Jev is the sum of its parts, not just one or two properties.

LLMs can produce structured output by limiting the next token based on a grammar, that ensures correctly formed output.

But (presumably, again I haven’t seen Jev’s insides) Jev doesn’t need to do that, it just has to output probabilities for each answer, and can do it natively without artificially limiting output tokens.

So jev is cheap, fast, never produces malformed output, its output always carries correct semantic meaning (but this doesn’t mean it always answers correctly), and it separates correct from prompt (which if it truly does that is the most exciting part).


Took me about 20 seconds to prompt inject the demo instance.... y'all would be great scammer targets


Can you share a sample prompt demonstrating it? I’d be interested to see, especially since you’re saying it’s incredibly easy (20 seconds would mean it’s easier than on most normal LLMs even).


JEV's moat is calibration nothing else. All other llms excel at system one like judgement versus JEV, but that's not a dunk on JEV. Calibration is very important for some niche use cases.


Jev is almost certainly not immune to prompt injection


It separates context/state from prompt/questions. That’s a key ingredient for prompt injection hardening. If it is designed and trained to never treat state as instruction, and you only ever put untrusted user input into state, and never the questions, then why wouldn’t it be?

The jev documentation says this is what you should do. It doesn’t mean that it’s actually designed for prompt injection resistance, but it could be. We won’t know for sure until they either release more details or someone proves otherwise.

But the split is one that doesn’t exist in normal LLMs and it’s the exact split that is needed for immunity or resistance to prompt injection.


So "Deno" isn't joining Cloudflare, Deno is effectively defunkt and its former team is joining Cloudflare.


jsr is moving to cloudflare.


isn't that fairly irrelevant if Deno dies out? Does anything else use JSR?


I tried using JSR for about a week, but the restrictions were so tight I gave up.


It’s immutable npm, you can’t unpublish I believe but yeah it’s trick of allowing one to publish typescript packages rather than compiled js only applies to deno.


JSR packages can run on Node.js, Bun, Cloudflare Workers, etc. The maintainer uploading the package decides which runtimes are supported.


Ah, fair enough, makes the whole situation a bit more lame


> if minor data or chat sessions are mishandled without explicit parental consent

I know this isn't what you meant, but I can't help but laugh at the thought of a parent consenting: "I consent to the mishandling of my child's data"

As to your actual point, yes, I assume they are trying to reduce liability, especially if they're gearing up to an IPO.


Since I see so many small kids with FB, Insta and TikTok accounts made for them by their parents - I find "I consent to the mishandling of my child's data" only funny in a very dark way.


Well, yes. Amusing might have been a better word than funny.


Interestingly, for me on iOS, all of the examples worked except the textarea one, which didn’t allow me to interact at all.


It’s also glitchy and flickery for me on iOS.


I enjoyed this talk. It’s long, but it’s interesting and goes into a lot of “lost” history.


I stopped using Claude because of this BS. If I pay for a service, I want to know what I’m paying for, I want it to be predictable.

Anthropic have been anything but. Flip flopping on model availability, model access behind an opaque filter, their past behaviour of model degradation as they prepared their next model… these are not signs of a reliable service.

I’ve mostly settled on using a mixture of open weights models through Together.ai and Fireworks.ai, a MiniMax subscription for high-token-use tasks that don’t need the best model (for $20 I get what feels like infinite tokens), and codex for the occasional high complexity task, although with Kimi K3 and hopefully soon GLM 5.3, it’s becoming increasingly less important. Deepseek 4 flash is my cheap main with delegation to other models as needed.

I’ve also found LFM2.5 8B surprisingly useful for single-focus tasks like “does this diff touch anything that isn’t related to the task”, and it’s incredibly cheap ($0.03/0.12 per M in/out).


I run LFM2.5 8B locally. It's not very smart, to say the least.

But then, there are plenty of mindless, menial tasks out there, and it would go through those like a champ. And quickly, too.


Exactly, not all tasks require smarts, they just need to be good enough. I’ve found that if prompts are really focused, the tiny models do alright.


> LLMs can now write perfect code in low-level languages, that is frankly _safer_ and _exponentially_ faster than implementations in high-level languages.

Your experience does not match mine.

It writes ok code, it does not write good code. I see it leaving a lot of performance on the table, doing silly things like wrapping everything in a global mutex. Or certainly doesn’t write idiomatic code in Rust or C++


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: