A terminal emulator in a GUI environment such as Linux is expected to play nice with the GUI and support mouse-based select, copy and paste, as well as being a terminal emulator, and this means that the terminal itself is consuming mouse events to support text selection.
If you wanted to write a shell that has mouse support you could certainly do so, and this would be based on sending escape codes to the terminal to tell it to return mouse events as input rather than let the terminal consume them itself. The shell could then itself support the clipboard if it wanted to, as well as support mouse editing.
I just googled it, and apparently "fish shell" does exactly this, but your hypothetical user is more likely to stumble upon a bash shell which is letting the terminal emulator consume the mouse events and support copy and paste.
Two years ago, I considered investing in Anthropic when they had a valuation of around $18B and messed up by chickening out (it was available on some of the private investor platforms). Up 20x since then ...
It was always obvious that Anthropic's focus on business/API usage had potential to scale faster than OpenAI's focus on ChatGPT, but the real kicker has been Claude Code (released a year ago).
It'd be interesting to know how Anthropic's revenue splits between Claude Code, or coding in general, other API usage, and chat (which I assume is small).
Eh, I think you made the best decision you could given the info you had.
I’ve poked around on EquityZen and was shocked at how little information is available to investors. In some cases I did not even see pitch decks, let alone one of the first companies I looked at had its top Google result: CEO recently arrested for fraud and business is almost worthless now.
Unless you are willing to take a blind punt or have insider information, those platforms are opaque minefields and I don’t fault you for not investing.
Matt Levine has a fun investment test: when presented with an opportunity, you should always ask, “and why are you offering it to me?”
Meaning, by the time it gets offered to retail investors (even accredited ones are retail) we’re getting the scraps that no one else wants.
Hiive and Forge Global are the ones I know of. You must be an "accredited investor" which means nothing at all except that you have a million dollars or make $200k/yr.
Like you can buy shares of Anthropic as long as you prove you make over 200K? That easy? Shouldn't they approve of the purchase? Sorry, noob in this space!
They have to approve and it's not as simple - it's just that if you make $200k a year or have $1m in the bank, the government assumes you're a knowledgeable investor and allows you to bypass certain protections.
If you are NOT knowledgeable and simply have money ... well it'll soon be parted.
The secondary platform verifies you and then you indicate interest. If there’s a seller you may get to buy. Company may ROFR. Priority goes to bigger buyers.
Well, it could be a batch editor, such as linux's sed, invoked from the command line, or with "computer use" the model could indeed potentially drive a real interactive editor.
Part of the problem though is that tools like Claude Code don't want to assume too much of the environment - that a specific editor is available, or even that it is running on a particular OS. The way it remains platform agnostic and not reliant on specific tools is by only having a dependency on Node.js, which provides file read/write support, so to implement an edit request the agent uses Node.js to read the file, itself implements the edit, then again uses Node.js to create the new updated file.
That's not quite how it works, and anyways if the model can't generate an accurate find/replace string, why would you expect it to do any better generating accurate commands to drive your editor (assuming it knew how do do that in the first place) ?!
The way edits happen is that the agent (local) first tells the model (typically remote) that it has an edit tool (e.g. taking parameters file name, find string and replace string). If the model decides it wants to edit a file then it'll invoke this edit tool, which just results in a blob of JSON being put in the model's response specifying the edit (filename, etc). The agent then receives the response, intercepts this JSON blob, sees that it is an edit request and does what is asked.
The problem the article is describing is that the edit request (tool invocation) generated by the model isn't always 100% accurate. Even if the agent told the model it had a tool to invoke an actual editor, say sed, assuming the model knew how to use sed, this is still going to fail if the edit request cannot be interpreted literally by the editor (due to being inaccurate).
Seems like it's veering towards a per-model protocol similar to the expectation that these models will develop their own languages to speak among themselves as agents.
The trouble is though, because it's all indeterminant slop, every model will break in small ways that you're back to indeterminancy and building a harness ontop of the harness.
Still, <nerd snipe>, there's probably a way to get the local model and arbitrary remote model to agree on how to make a method call. But the only way that will be fruitful if you find a highly reproducible set of tuples within the model's shared space.
> Everyone thinks their use of AI is perfectly justified while the others are generating slops
No doubt, but I think there a bit of a difference between AI generating something utilitarian vs something expected to at least have some taste/flavor.
AI generated code may not be the best compared to what you could hand craft, along almost any axis you could suggest, but sometimes you just want to get the job done. If it works, it works, and maybe (at least sometimes) that's all the measure of success/progress you need.
Writing articles and posts is a bit different - it's not just about the content, it's about how it's expressed and did someone bother to make it interesting to read, and put some of their own personality into it. Writing is part communication, part art, and even the utilitarian communication part of it works better if it keeps the reader engaged and displays good theory of mind as to where the average reader may be coming from.
So, yeah, getting AI to do your grunt work programming is progress, and a post that reads like a washing machine manual can fairly be judged as slop in a context where you might have hoped for/expected better.
Yes, it seems the open benchmark results that are normally reported, such as SWE-bench, SWE-bench Verified, and Terminal-bench, aren't really that indicative of success in more general use cases.
According to Gemini, SWE-bench is actually a very narrow test, consisting of fixing GitHub issues drawn from 12 large Python projects (with Verified being a curated subset of that), and Terminal-bench (basically agentic computer tool use) is more focused on general case rather than use of the tools used by a typical coding agent such as Claude Code, Codex CLI or Gemini CLI.
Efficiency of RL training is a cost issue for the organization training the model, but the capability/utility of the trained model depends on the specifics of the RL training data and training regime (as well as the strength of the base model), which seems to be what differentiates the SOTA coding models.
It's neat that Z.ai are opensourcing slime, and are themsleves using DeepSeek's Sparse Attention - a different approach to that of the big US companies.
ctags just gives you locations of symbol definitions.
TreeSitter will also give you locations of symbol usages, which is obviously very useful to an AI agent. You can basically think of Treesitter as having full syntactic knowledge of the code it is looking at - like a compiler's AST.
There is also a more powerful cousin of ctags, cscope (C/C++) and Pycscope (python) that additonally gives usage locations, and more, as well as gtags that does similar, but supports more languages.
LSP is designed to help editors, not AI agents, and provides query by cursor position. Treesitter supports query by symbol (find definition, usages, etc), and so is much better suited to what an AI agent may want to do.
Thanks - I wasn't aware of that, although it still doesn't seem to be what would be most useful to an AI agent.
For example, if the agent wants to modify a function, it may want to know all the places the function is called, which AFAIK Treesitter can provide directly, but it seems with LSP you'd have to use that DocumentSymbol API to process every source file to find the usages, since you're really searching by source file, not by symbol.
> Apple doesn't need to solve AI. It's not core to their business
Perhaps, but it depends on what business they are really in...
One classic business failure ("Marketing Myopia") is to define the business you are in as the product or service you sell, rather than the customer need that you fill.
It's certainly been a long time since Apple was in the "phone business", and Nokia is an example of what happened to a company that thought that was the business they were in.
For now, AI is largely being packaged in a way that is somewhat orthogonal to what a smartphone does - as a service (e.g. AI chat) that it can consume - but as AI becomes more pervasive that will change, and it seems that increasingly the mobile device in your pocket will become more like your do-it-all personal assistant rather than a pocket computer that you use to run different applications do to different things.
So, do Apple think they are in the smartphone business, with AI as someone else's business, a service that their phones can consume, or are they correctly anticipating where things are heading?
My take is that the future will look something like this:
- You pay for a personal AI assistant from a cloud vendor (most people) or you run it yourself on your own hardware (not yet common, likely somewhat common in the future as hardware becomes cheaper and open weight models keep getting better). This assistant won't be a chatbot but an autonomous agent (like OpenClaw today). Some of these will be free but heavily subsidized through aggressive ads.
- This AI assistant has hooks into whatever personal services you want (email, cloud storage, photos, messages, etc.).
- You own a variety of devices in different form factors, each of which increasingly acts as a way of interacting with the same AI assistant, which exists independent of device. Some of these form factors will be new ones that don't meaningfully exist yet today, like true high-end AR glasses.
- Many apps and websites will eventually just become on-demand generative interfaces spawned by your AI assistant. Some "fixed" or "pre-programmed" interfaces will still exist, though.
For Apple, there are really two questions: (1) do they need to create their own frontier AI assistant to play a significant role in this future long-term? (2) if the answer to the former is "yes", when do they need that by, and how does it strategically weigh against creating the next generation of compute form factors that show up in the third item above?
Given that Apple has openly stated that they intend to create personalized intelligence across their ecosystem and that they don't know that the smartphone will be the dominant form factor in a decade, I think their answers are: (1) yes; (2) they need to have one eventually, but it's even more important that they prepare for next-gen form factors, and so they're okay being late to the game on AI assistants as long as they get there soon enough.
I'm personally not convinced about "next-gen form factors", although I know that's what many companies are focusing on - some type of smart glasses, or whatever kind of (screen-less?) device Jony Ive and OpenAI are working on.
Most people are too appearance and fashion-conscious to want to wear tech on their face, and I don't see many people wanting to carry TWO expensive tech gadgets (and worry about charging/losing/forgetting them), so, seeing as photos and video is core to what people want from their mobile device, it seems that the smartphone will continue to be the form-factor of the future, and I expect these other next-gen form factors to fail.
I think Apple's brand loyalty buys them some lead time in being a fast-follower, but the danger to them would be if things change so fast and profoundly that they get left behind a la Nokia. What if Google or someone else comes out with an AI-centric "personal assistant" device so compelling that it massively ups the bar as to what customers expect from a mobile device (in same way that iPhone did at launch)? I wouldn't expect it to kill Apple overnight, but it seems that they are in effect gambling that "we can always pay for AI if we have to", and "someone will always license it to run on-device if we need to".
> Most people are too appearance and fashion-conscious to want to wear tech on their face
Have you seen the Ray Ban meta glasses? They already look pretty close to existing fashionable sunglasses, albeit with a visible camera.
> and I don't see many people wanting to carry TWO expensive tech gadgets (and worry about charging/losing/forgetting them)
They already do; plenty of people carry a smart phone, a smart watch, and airpods.
> seeing as photos and video is core to what people want from their mobile device, it seems that the smartphone will continue to be the form-factor of the future, and I expect these other next-gen form factors to fail.
People use smartphones to avoid being bored, but there are situations when it's unacceptable to use them (i.e. in a meeting); I could see smart glasses being used for that niche.
> What if Google or someone else comes out with an AI-centric "personal assistant" device so compelling that it massively ups the bar as to what customers expect from a mobile device (in same way that iPhone did at launch)?
Knowing Google, that personal assistant would probably be shut down within a year.
If you wanted to write a shell that has mouse support you could certainly do so, and this would be based on sending escape codes to the terminal to tell it to return mouse events as input rather than let the terminal consume them itself. The shell could then itself support the clipboard if it wanted to, as well as support mouse editing.
I just googled it, and apparently "fish shell" does exactly this, but your hypothetical user is more likely to stumble upon a bash shell which is letting the terminal emulator consume the mouse events and support copy and paste.