blourvim

You could have limited-instruction llms where the model does one thing, for example summaries. It could accept a limited amount of instructions for example, first token for verbosity, second for style etc...

These models rely on knowledge that are embedded in their weights, if a new library is released, a new linux version comes out, some new protocol succeeds the previous one, you want your llm to know about it. Sure you can just add that into the context window, but that has its own problems.

Unless new research, there are a few which look promising, gives a new method, training is going to be a constant cost sink.

On top of this, if you stop training, it is 6 months until someone releases an open weights model and now you are competing to give the lowest price for the same product.

Also we can't forget that this is a business that *has to* be in the global labor industry, not just a tech tool, they have to have much better models to justify the trillion dollar evaluation

The web server is able to detect and change the response of a curl | sh thus a malicious actor could avoid some detection.

Having a binary also allows for antivirus software to be ran on it before deciding to run the binary itself or not.

I am not sure if curl | sh bypass sandboxing by default, but definitely is the case when compared to flatpak or just regular packaging

Training is expensive, to keep selling tokens, they have to have the best model. They have to keep spending money on model training

A product or a platform could give them a moat, but they don't quite have that yet although I think they are looking to get there

quite satisfying to hear the clicks

What I don't understand is that if a lawsuit happens, then must the plaintiff produce their source code for verification ? Even so a git tree is trivial to change into some other arbitrary code even if a license violation has occurred. I also heard if proven the consequences are that they would lose all revenue starting from when the violation has occured

Thats a big sometimes for me. I am unfamiliar with the app development process. For backend,I would rather design the whole thing from scratch rather than try to fix a broken architecture. For frontend, maybe that is true, but I haven't had the chance to build anything complex, just forms and tables.

I am a backend dev, and building a mobile app for the first time, It is a toy sized project right now and it works, would have indeed taken me 5x the time to get to that point on my own.

I know from experience that to make it a non-toy project I am probably going to need to spend some, if not most of that saved time cleaning up in the future due to technical debt.

For backend, it is mildly useful at best, helpful with boilerplate and when I know exactly what needs to be built. I have not yet firsthand experienced these 10x, 100x productivity gains.

Frontend, forms,tables and generic dashboards is pretty good, I am sure one could get it done faster and better over the long term with proper technique and methods, but I just hate css

Ah, wise words. There is still time indeed

Mildly disappointed that it is not the dystopian adpocalypse we were warned against

License in question: https://github.com/papermark/papermark?tab=License-1-ov-file It is AGPL, basically means:

You have to share the source code even when the user interacts over the network with the software.

The project which uses that code, must also be AGPL,

There are ways to separate it and go around it, for example, using an AGPL auth server shouldn't affect the code where your business logic lives

I am sure they could have found a way to design their product to be compliant, especially following past drama.

This is assuming the code is indeed copied, since we don't know that for sure, it does look very similar but I am not sure how that is enforced

I am not a lawyer, from what I understand that the precedent is that you can use copyrighted material in ML process. Even though meta has, allegedly, pirated the material, the cost of violation would be pennies compared to the ai spend, since that is the violation, not that they used those materials,

Benchmarks in general are a little iffy, the whole industry is going off of vibes anyways. Can't decide before trying it out

I would guess that they would want to at the very least 10x their prices. Remember they need to make up for training, marketing, etc.. and make a big chunk of profit on top of that to justify their trillion dollar evaluation

I have been playing nonograms for a year now, there was never insta death, controls are a little off, title is hidden until its solved to avoid spoilers This is obviously programmed by someone who doesn't play nonograms. Safe to assume it was vibe-coded

i like it

I heard that the rule changes which would allow SpaceX to be auto bought by those funds has been blocked, previous stock seasoning rules will apply

I'd like you to set up here.now: the web hosting service for agents.

Install as a skill if I have npm: npx skills add heredotnow/skill --skill here-now -g

If not, do this instead: curl -fsSL https://here.now/install.sh | bash

We really need to stop with the curl pipe bash pattern

Article says,this is a software issue. Where GPU'S are unable to get to be fully utilized due to scaling issues. I dont know how hardware that scale works, but it could very well be that they still need all of their hardware to get their current compute

It was fun, I couldn't figure out what secrets going on at the treehouse, the book margins are too hard to read nor the locked room in the building

cool to see you in the wild, for me, it does work out of the box however, some sites will break or have too complex of a navigation, especially with iframes. and will have to swap to a mouse which is a bummer, which I understand is an inherent limitation of the tech, since web is not built today to do that.

solid extension, big fan

This should actually be fine, no different than any other web server

git will check the .git folder, find a hook, and run it where it is applicable. If you are cloning a remote repository may inherit you with malicious hooks. These hooks run before you git operations, for example it is useful if you want lint the code a certain way before pushing, it does it automatically.

You can disable this behavior globally. Yes, the agent should have no git access this way, however you could always do a local sub repository if you want to. You track your changes twice, but should work

You can keep the git folder in your machine, then use sshfs to mount the remote into your directory where: project-root -.git/ -code/ <= remote/code

This way you don't have to give git access to the potentially unsafe server. Git hook attacks are still possible so disable those by defaultç

This is an unusual folder structure but works fine, let me know if there is anything iffy

nice one, discovered a trick where holding space space while holding left or right allows side stepping.

Most people who spend time in front of a keyboard should without effort should attain 60 words per minute naturally. With decent techniques, you can go up to 80 wpm. Now there are speed typers who can do 250 wpm. Can you really think meaningfully at that speeds, in a way that it can be understood and is useful ?

At that point one could argue that running your own server to get instant responses from your llm matters, while we are at it, compile times, tests, search queries, could always use be shaved some milliseconds per.

With speed reader tools I can comfortably read and understand 500 wpm, should we now switch to speed readers since we read code more often than we write it ?

No, in reality, you create software faster when the code you write the first time doesn't result in errors, which requires knowledge of your stack.

Even if you are not looking to read your code, LLMs also benefit from good programming practices, some of these programming practices should produce code which is clear, changeable and extendable

Nvidia has done it before, I think they included a virus in their data which encrypted the stolen data

reminds me of those 2010's flash games, absolutely amazing

I am not really an ml dev so I don't understand most of it. It does sound ridiculous how it would even work work. Brilliant work and great article I enjoyed reading it

This sounds similar to the Kimi's mixture of experts architecture if I understood it correctly(likely I have not), can you comment on this ?

Time limit is brutal, couldn't do it without some googling, but managed to find an Irish flag in Paris at the Irish embassy. Fun premise for a game. I can see myself playing this with friends.

First play through I found the difficulty a bit much, but I think some geowizards will disagree