dennisy

I too found this extremely interesting, jotted some notes here: https://dennisy.me/notes/programs-as-weights

Will keep noodling on this, I feel there must be ways to scale this up!

That would be interesting.

Do you feel most of the speed upgrade will come from the software or hardware side?

Thanks for the replies!

Would love to swap notes at some point if you are up for it?

Hey!

This looks great and congratulations on the launch.

I am also building in this space and wanted to get your views on a few things.

1. Are you building your own connectors to 3p systems? 2. How are you finding the sales motion? I found people to get the problem fast, but actually converting them seems rather slow.

Good luck!

Great read, I feel many of the topics discussed firsthand.

It gives me some relief to know there are others out there who struggle with some similar issues, but I was hoping the piece may offer some guidance, but sadly I do not feel it has.

Very cool! Not GitHub repo?

I love this and wanted to build this - but https://www.alphaxiv.org/ already exists, and it gets no social action (hardly any papers have comments), so this makes me doubtful about this.

I am interested to hear if anyone knows why the format may not resonate with researchers or those reading papers in general?

My own reason is that to get value from a "social" site the number of interactions has to be high and of a fast speed for people to continue to engage, which is maybe not possible to hit on research papers.

This is very cool!

But why do we need this? An agent can just have a local DB using SQLite for example.

This is a fair question, but not one I feel we can let people self answer.

I doubt many people will honestly admit they did no design, testing and that they believe the code is sub par.

It does give me an idea that maybe we need a third party system which can try and answer some of the questions you are asking… of course it too would be LLM driven and quite subjective.

True! But this is a very naive implementation, a proper implementation could surpass these challenges.

Congratulations on the launch!

There is lots of competition in this space, how is your tool different?

I think this used to be the trick, but now with AI it is such a general purpose technology I am not sure that makes sense. Users can “niche down” in a generic app, using prompts and configs.

Fair play for launching this, it looks like a neat project.

However I feel it will be an uphill battle competing with OpenAI and Anthropic, I doubt your harness can be better since they see so much traffic through theirs.

So this is for those who care about the harness running on their own infra? Not sure why anyone would since the LLM call means you are sending your code to the lab anyway.

Sorry I don’t want to sound negative, I am just trying to understand the market for this.

Good luck!

I feel for the startups sweating each one of these frontier lab releases.

How many more are thinking “am I next?”

So what are you suggesting do not allow companies to sell such tools?

Ohh this is very cool!

I think most people would agree.

However it is less clear on how to do this, people mostly take the easiest path.

That is good advice!

I am too late for that as I am full time but also lucky to have had a previous exit.

I am planning to go the VC route this time, because the problem I am going after feels VC size.

However I feel bootstrap or small F&F round gives you more flexibility when it comes to exits.

Congrats again!

Congratulations!

Can you talk a bit more about your journey without raising funds?

Also what does HN think of that path today when trying to launch a new AI startup?

I was thinking something that runs in the same process, and does not require docker or a third party API.

For example Monty by the pydantic team, or the Anthropic sandbox which I believe uses OS level primitives.

Congrats on the launch, the agentbox-sdk looks interesting, but seeing as the first commit was 3 days ago - I feel a little wary to use it just yet!

One question, do you have plans for any other forms of sandboxing that are a little more "lightweight"?

Also how do you add more agent types, do you support just ACP?

Let’s just shorten it to AaS?

The “next gen of models” argument is a valid one and one I think of often, but if you truly buy it, it would stop do creating anything - since the next gen of models could make it obsolete.

Spicy! Upvoted just because I would look forward to the HN debate that would happen on this topic!

My personal view here has shifts quite often, yes software is getting easier to build at an alarming rate, but designing good software is not (yet), and new novel software is not in the training set so should remain hard to build, forever (?)

To address the main comment, yes lots of apps are just wrappers, and I feel this is not sustainable, we will not need all these specialised apps since AI is such a general interface to computing.

Just came across this and found it to be quite interesting - they propose to evolve the memory architecture per task, which seems like a nice idea since it is so hard to know up front what your agent needs to learn / remember.

It reminds me of GEPA & Textgrad.

Would love to hear if anyone has worked on anything of this sort in practice.

How did you guess it was AI slop? I read it and answered before seeing your comment :)

Whilst this is a very valid view, I am not sure there is evidence for it, if you have it please do provide it.

It is not obvious to me that one credible citation is better than having appeared many times in the training data.

Also knowledge releases of models are too slow to be depended on for marketing in my view. It is the search / retrieval of LLMs for which one can optimise.

A final point would be that none of this changes the point made, which is that there too much noise due to too many products being shipped.

Could you please expand on “none of them find the tools very useful”?

I would love to know how big your sample is, in what way the tools fail, what features are missing etc.

I have thought about this idea many times and think it would be amazing!

Also could think of it a little like a “Wikipedia of flashcards”.

Would you be interested in working on something like this?

Are we only able to think of these systems as some form of human and probe them from the outside like a therapist?

Surely these sorts of problems must be worked upon from a mathematical standpoint.