a1j9o94

You are right that someone could go through that effort bat that only really makes sense if they give you some.idea how to use it.

Reading the exchange it seems like you got caught up on the illustrative examples the other person used. If a web server isn't a good example of a simple problem for F* the landing page should have an example of something that is. It shouldn't take going through that many steps to understand the so what.

Disclaimer I work at Zapier, but we're doing a ton of this. I have an agent that runs every morning and creates prep documents for my calls. Then a separate one that runs at the end of every week to give me feedback

This is effectively how I treat my AI agents. A lot of the reason this doesn't work well for people today is due to context/memory/harness management that makes it too complex for someone to set up if they don't want a full time second job or just like to tinker.

If you productize that it will be an experience a lot of people like.

And on the UI piece, I think most people will just interact through text and voice interfaces. Wherever they already spend time like sms, what's app, etc.

You would only use the base model during training. This is a distillation technique

I fall into this trap a lot. The platonic ideal argument is a fun mental exercise but doesn't get anything done

This is an interesting space. Right now we've gotten to a point where agents can do most tasks, but they will get lazy/skip steps if you're not precise in the requirements. We need ways to validate that expands beyond software tests. This is a good direction but a few thoughts: 1. From what I can tell the agent who does the task is running the validation. Keeping the validation agent as a separate context avoids the validator knowing what the software is supposed to do vs what it does 2. There's a lot of prior art around org structures to validate things that we've built out over the last ~100 years that we can apply in this space. E.g. look at the way that blind trials are run

Honestly just didn't think about it. Added it.

Hey HN,

  I spent the last few weeks exploring whether AI systems could benefit from generating video predictions before making decisions—like how humans mentally simulate "what happens if I pour this coffee?" before acting.

  The idea: Show an AI an image, ask "what happens if I push this?", have it generate a video prediction, then compare that prediction to reality. If the prediction looks wrong, maybe the AI could catch its own mistakes.

  The result: Current models can't do this. But I learned some interesting things along the way.

  What I tested:
  - 7 different architectures for predicting future video frames from VLM latent space
  - Whether perceptual similarity (LPIPS) between predicted and actual video correlates with correctness
  - Self-correction loops where the model gets feedback on its predictions

  Key findings:

  1. VLMs can't predict the future – Every architecture I tried performed worse than just copying the current frame as the "prediction." The model understands what's in an image but can't predict what will change.
  2. Visual similarity ≠ semantic correctness – This one surprised me. Wrong predictions often looked MORE similar to reality than correct ones (LPIPS correlation: 0.106). You can't use "does it look right?" to catch mistakes.
  3. Some things worked – Hybrid encoders (DINOv2 + VLM) preserve spatial information that VLMs lose. VLMs understand generated video well (93% semantic retention). Small adapters (10M params) work better than large ones (100M).

  I'm releasing this as a benchmark proposal. Video generation is improving fast—capabilities that don't exist today might emerge in future models. Seems worth tracking.

  Links:
  - Demo video: https://youtu.be/YJxDt_zCrUI
  - Code + paper: https://github.com/a1j9o94/foresight
  - Live demo: https://foresight-demo-kappa.vercel.app

  Built with Qwen2.5-VL, LTX-Video, Modal (GPUs), and the Something-Something v2 dataset.

  Happy to answer questions about the experiments or methodology.

Pretty much every major LLM client has web search built in. They aren't just using what's in their weights to generate the answers.

When it gives you a link, it literally takes you to the part of the page that it got its answer from. That's how we can quickly validate.

I would argue that's just your coworker giving you a bad answer. If you prompt a chatbot with the right business context, look at what it spits out, and layer in your judgement before you hit send, then it's fine if the AI typed it out.

If they answer your question with irrelevant context, then that's the problem, not that it was AI

Honestly if you have a working relationship/communication norms where that's expected, I agree just send the 5 bullets.

In most of my work contexts, people want more formal documents with clean headings titles, detailed risks even if it's the same risks we've put on every project.

I know I'm an outlier on HN, but I really don't care if AI was used to write something I'm reading. I just care whether or not the ideas are good and clear. And if we're talking about work output 99% of what people were putting out before AI wasn't particularly good. And in my genuine experience AI's output is better than things people I worked with would spend hours and days on.

I feel like more time is wasted trying to catch your coworkers using AI vs just engaging with the plan. If it's a bad plan say that and make sure your coworker is held accountable for presenting a bad plan. But it shouldn't matter if he gave 5 bullets to Chat gpt that expanded it to a full page with a detailed plan.

Not the person you're responding to, but I think there's a non trivial argument to make that our thoughts are just auto complete. What is the next most likely word based on what you're seeing. Ever watched a movie and guessed the plot? Or read a comment and know where it was going to go by the end?

And I know not everyone thinks in a literal stream of words all the time (I do) but I would argue that those people's brains are just using a different "token"

Having one tool that you can use to do all of these things makes a big difference. If I'm a financial analyst at a company I don't need to know how to implement and use 5 different specialized ML models, I can just ask one tool (that can still use tools on the backend to complete the task efficiently)

yy

The above is saying more precise not completely precise. The overall point they're making is you still are responsible for the code you commit.

If they are saying the code in this project was in line with what they would have written, I lean towards trusting their assessment.

Why do you say that? I would argue that as long as your tests and interfaces are clearly defined no reason it couldn't scale indefinitely.

I agree with this completely. I get the impression that a lot of people here think of software development as a craft, which is great for your own learning and development but not relevant from the company's perspective. It just has to work good enough.

Your point about management being vibe coding is spot on. I have hired people to build something and just had to hope that they built it the way I wanted. I honestly feel like AI is better than most of the outsourced code work I do.

One last piece, if anyone does have trouble getting value out of AI tools, I would encourage you to talk to/guide them like you would a junior team member. Actually "discuss" what you're trying to accomplish, lay out a plan, build your tests, and only then start working on the output. Most examples I see of people trying to get AI to do things fail because of poor communication.

The point is devs aren't sales/client facing. So from the customers perspective, it's just a delivery detail.

How do you do things like compare prices in plain text?

Wouldn't phone books and catalogues count as advertising?

This is really interesting. Has anyone seen a good registry of MCP servers somewhere? What I would love to be able to do is ask for a task and have the agent figure out what servers and tools it needs to accomplish that task directly.

Probably not the whole model, but the first step was "fine tuning" the base model on ~800 chain of thought examples.

Those were probably from OpenAI models. Then they used reinforcement learning to expand the reasoning capabilities.

Ideally yes, but most people in corporate settings don't communicate clearly. So if they have a culture of dropping hints, and you don't take them you're not going to do well at the company.

It's not good, but it's an accurate reflection of the work environment.

I was part of a group that did something like this. We used a crypto token system where if you joined the group you got X number of tokens and you could use that to get other people in the network to help with various tasks.

Then the network would take a small percentage of equity in the companies of all the members.

Cursor isn't designed to do long running tasks. As someone mentioned in another comment it's closer to a function call than a process like Devin.

It will only do one task at a time that it's asked to do.

It's also an American company building the project. The cultural values of the US are relevant.

It's HR's entire job to set policies for hiring. They can say a candidate has to have a college degree. Why wouldn't they have the right to set this policy as well?

If you're using ChatGPT directly as opposed to the API, the system prompts could be driving it.

Also, in section 3.6 of the paper, they talk about just switching fishing email, to email helps.

Or said differently, tell it that it's for a marketing email, and it will gladly write personalized outreach

This is super interesting.

I'm still waiting for a model that generates images as part of its thought process. That's what I was hoping this was!

I wonder what it would take to train a proof of concept. Maybe start with videos of people whiteboarding and explaining math or solving engineering problems visually. Then, create captions that frame these as word problems. Train the model so that, during inference, it generates images—not necessarily for human viewing, but as part of its internal reasoning.