I’ve been diving into MCP the last few days and it feels quite messy.
It took an embarrassingly long time for me to figure out what the difference is between a client, server, and host are. Oh and servers can also be processes that you run locally and clients can spin up the server for you but you need to provide the NPX or UVX command.
The problem it’s supposed to solve is quite real, tools are annoying to build and share so I get that. Looking forward to it getting better but it’s not at the point where any dev can come in and connect things up with a nice dev experience yet.
That's awesome. The original discussion of bitnet made it seem like you needed to train a model from scratch but its neat they were able to adapt an existing model. This is quite exciting.
Yeah, its kind of a bummer. Dropbox doesn't really have a cohesive vision for their productivity suite? Dropbox Paper doesn't seem to have progressed and there haven't been any other notable new products released.
Funny enough, there's a DSL LLMs already tend to be trained on: JavaScript. Just teach the LLM the function signatures it has to work with, then execute the script in a VM
The original claim in the code dump is that no ML tools are used at all and the tool is just leaning on Playwright to automate specific actions on a website.
The CEO here is claiming that the ML code is being run outside this code base and that the original claim is being made by someone who doesn't know how the code works.
The CEO's mention of sanitized code isn't as clear to me, that can mean different things. Compiled code can be considered sanitized since it likely isn't human readable, obfuscated code makes that harder, and removing some code all together would be the most effective. The problem with removing code all together is that you would still find code paths that just can't be executed at all, leaving some trail of what code was removed. That wouldn't leak any secrets obviously, but would support the argument that code has been removed and the codebase is being misread.
The code is their "minions" to handle actions on websites. When you ask it to, say, book a trip, and it tries to search AirBnB.
"If someone spends enough time with the login minions they can extract these code"
AKA "Someone will figure out how this worked, but our code is secure, trust us".
The "rabbit hole" they mention is the whole "cloud" system that Rabbit talks about using to manage all of your services and integrations and 'rabbits' you create that run tasks.
It seems to just be a leak of their sandboxed headless browser setup and the API code for controlling it. Obviously such a thing will run arbitrary JS from the web so inevitably there will be something like a browser sandbox exploit, and subsequent dump of its filesystem.
The leak doesn't seem to contain what Rabbit calls the LAM, their purported AI model for interacting with UIs. And what the leakers are claiming is that Rabbit's automation is just handwritten scripts which seems to be completely unsubstantiated. The rabbit secret sauce could still turn out to be a scam but I didn't see anything to corroborate any of the leakers' claims. Grepping the files I found no reference to doordash, uber eats or midjourney, only a path reference to what appears to be a spotify integration library, but the source for that isn't there.
I recently forced myself to learn vim motions and it's been paying off pretty well. Once it clicked I found myself wanting vim motions everywhere.
I went deep and set up a neovim config, forcing myself to use neovim full time. However my new role requires me to use Intellij and I didn't want to spend hours upon hours figuring out how to replicate the java support in Neovim (and no, it isn't just installing the LSP, my work has some pretty specific intellij setup).
At this point I'm just using Intellij and VSCode with their respective vim plugins and I'm getting about 80% of the utility that I did with pure neovim. The one thing I really love about neovim is that it really forces you to keep your hands on the keyboard and you learn the keybinds very well because otherwise you just can't do the thing you wanted to! I find myself using the mouse a lot in VSCode and Intellij because they give you really handy buttons to click on.
I could probably go pure keyboard if I just sat down and learned all the various keybinds, but who has time for that?
Have you tried the neovim plugin instead of the regular vim one? I find they have different trade offs, but the neovim one is actually using neovim inside a vscode window and requires a neovim install.
Particularly, I’ve found the undo functionality to be far better with the neovim one, it never screws up. Holding alt and moving a line up and down can sometimes accidentally move your cursor one line higher or lower though.
Biggest personal issue with VS codes vim mode is that it only works in the text editor. The beauty of vim/neovim is that the keybindings work everywhere in the UI, not just the text edit window.
I basically had to disable entire VS Code. Didn't remove it yet and keep it just in case I need to remember how crappy it's. It has some advantages over neovim though, but they don't outhweigh neovim.
For IntelliJ, check out what you can do with the .ideavimrc file. You can turn pretty much anything into a vim motion and it has light plugin support as well. Also makes your setup portable.
Yeah, I've been meaning to spend more time on it, but I've been switching around to different editors a lot lately, so the constant jumping around means I'm trying to keep my vim keybinds as vanilla as possible.
The actual prompt is just a chain of thought prompt and the headline is just clickbait.
Also in the article, the author gets COT wrong.
> What "reasoning" they do (and "reasoning" is a contentious term among some, though it is readily used as a term of art in AI) is borrowed from a massive data set of language phrases scraped from books and the web. That includes things like Q&A forums, which include many examples of "let's take a deep breath" or "think step by step" before showing more carefully reasoned solutions. Those phrases may help the LLM tap into better answers or produce better examples of reasoning or solving problems from the data set it absorbed into its neural network weights.
Chain of thought has nothing to do with “tapping into better answers”. It’s simply asking the model to break up the output into smaller tasks and gives it more time and space to reason.
COT is not new or novel. Hell, it’s even listed in one of the guides in Open AI’s prompt guides as a strategy to improve prompts.
> Chain of thought has nothing to do with “tapping into better answers”. It’s simply asking the model to break up the output into smaller tasks and gives it more time and space to reason.
It doesn't give it more time and space to reason, though. Time isn't usually bounded on a single turn and space is context-window limited (and every turn in Chain of Thought is done within the context window, so it doesn't add any space.)
What it does is push the output toward a shape that resembles a particular idealization of (the explanation of) human reasoning, producing results that look more like an explanation of reasoning and sometimes producing more satisfying conclusions.
It does give more space in the sense that the model generates more tokens of steps/etc that it can then base its actual answer on, rather than being forced into generating the answer right away.
> It does give more space in the sense that the model generates more tokens of steps/etc that it can then base its actual answer on,
It doesn't give more space in the sense of increasing the upper bound on space used; it may bias the space used higher than a single naive prompt aiming to respond to the same question, but it doesn't alter the constraints.
It took an embarrassingly long time for me to figure out what the difference is between a client, server, and host are. Oh and servers can also be processes that you run locally and clients can spin up the server for you but you need to provide the NPX or UVX command.
The problem it’s supposed to solve is quite real, tools are annoying to build and share so I get that. Looking forward to it getting better but it’s not at the point where any dev can come in and connect things up with a nice dev experience yet.