Canvas is a new way to write and code with ChatGPT

But can they do it at all? It's not like they are like early Google vs other search engines.

We are getting front row seats to an object lesson in “absolute power corrupts absolutely”, and I am relieved they have a host of strong competitors.

* Or which ever variant the average user might try to type in

It wouldn't surprise me if not a lot of enterprises are going through OpenAI's enterprise agreements - most already have a relationship with Microsoft in one capacity or another so going through Azure just seems like the lowest friction way to get access. If how many millions we spend on tokens through Azure to OpenAI is any indication of what other orgs are doing, I would expect consumer's $20/month to be a drop in the bucket.

(1) https://www.tanayj.com/p/openai-and-anthropic-revenue-breakd...

Same as it ever was.

I use Gemini, OpenAI, Claude, smaller models in Grok, and run small models locally using Ollama. I am getting to the point where I am thinking I would be better off choosing one (or two.)

OP is lamenting that Cursor and OpenAI chose to create new apps instead of integrating with (someone else’s) existing apps. But this is a result of a need to be always fully unblocked.

Also, owning the app opens up greater financial potential down the line…

Yea, I'm wondering the same. Is there any good resource to look up whether copilot follows the ChatGPT updates? I would be renewing my subscription, but it does not feel like it has improved similarly to how the new models have...

You only need 45 days to get tier 5 and if you have that many customers after 45 days you should just apply to YC lol.

Maybe you checked over a year ago, which was the wild wild West at the time, they didn't even have the tier limits.

I’m firmly in the camp that their rate limits are entirely reasonable.

You can sell a lot more GPT services through a higher bandwidth channel — and OpenAI doesn’t give me a way to reach the same bandwidth through their user interface.

I LOVE the UX animation effect ChatGPT added to show the canvas being updated (even if it really is just for show).

Here's my user test so you know I actually used it. My jaw begins to drop around minute 7: https://news.pub/?try=https://www.youtube.com/embed/jx9LVsry...

"Name it clay" -- artistic CMO

"Won't people think they will have to get their hands dirty?" -- CEO

"Right. Name it sculpt. It has a sense of je ne sais quoi about it." -- hipster CMO

"No one can spell sculpt, and that French does not mean what you think it means." -- CFO

"Got it! Name it canvas! It's a blank canvas! It can be anything!" -- CMO

"Perfect!" -- Entire executive team who is now sick of CMO's naming excercise.

[0] https://support.anthropic.com/en/articles/9487310-what-are-a...

Canvas appears to be different in that it allows inline editing and also prompting on a selection. So not the same as Claude.

Claude's been far and away superior on coding tasks. What have you been testing for?

Projector make it even better. But I could imagine it depends on the specific needs one has.

ChatGPT will just start to pretend like some perfect library that doesn't exist exists.

If OpenAI can copy Cursor, so can everyone else.

Met a guy who got brought in by Amazon after they hit 8 figures in sales, wined and dined, then months later Amazon launched competing product and locked them out of their accounts, cost them 9 figures.

You mean downstream.

> They're not wasting hours trying to "figure out" a solution

I am pretty sure that we don't have AGI that would figure our solutions to our problems (coding or not) on its own yet. And from experience, you would need to solve the problems at least conceptually before using LLM and try to get something useful out of that.

    Your competition is using AI to build their product right now.
    They're not writing code by hand.
    They're not Googling code answers on StackOverflow.
    They're not wasting hours trying to "figure out" a solution.

Let's see how far I get without "Googling answers".

Spoiler: If the competition uses such AI that, good for me!

(Note I don't want to pick on Parrot here; from my first 5 minutes with the tool, it's /fun/, has a nice UI, a good idea for iteration, and somebody has to build such tools to make them work eventually even if they don't work great today yet. I wish it lots of success. But I want to also point out examples of the practical limitations of all such tools, and state that I highly doubt that they today can _in any way_ replace a human programmer who knows what they are doing, and has the awesome skill to use Google, StackOverflow, and "figure out" things.)

Prompt:

    I need a Haskell function that hashes a given file (FilePath)
    with the blake3 hash function.
    It is important that it works on large files (e.g. 100 GB)
    without freezing the Haskell runtime.

    import qualified Crypto.Hash.BLAKE3 as BLAKE3
    import qualified Data.ByteString as BS
    import System.IO (withFile, IOMode(ReadMode))

    hashFileBLAKE3 :: FilePath -> IO BS.ByteString
    hashFileBLAKE3 filePath = do
      let chunkSize = 8192  -- 8KB chunks
      withFile filePath ReadMode $ \handle -> do
        let loop ctx = do
              chunk <- BS.hGet handle chunkSize
              if BS.null chunk
                then return $ BLAKE3.finalize ctx
                else loop $! BLAKE3.update ctx chunk
        loop BLAKE3.init

    This Haskell function, `hashFileBLAKE3`, calculates the BLAKE3 hash of a
    given file. It's designed to work efficiently with large files without
    freezing the Haskell runtime. Here's a step-by-step explanation of how it
    works:

    1. We import necessary modules: `Crypto.Hash.BLAKE3` for the BLAKE3 hash
       function, `Data.ByteString` for efficient byte handling, and `System.IO`
       for file operations.

    2. The function takes a `FilePath` as input and returns an `IO ByteString`
       containing the hash.

    3. We define a `chunkSize` of 8192 bytes (8KB) for reading the file in
       manageable chunks.

    4. The `withFile` function is used to safely open and close the file,
       ensuring proper resource management.

    5. Inside the file handling function, we define a recursive `loop` that
       reads the file in chunks:
       - It reads a chunk of the file using `BS.hGet`.
       - If the chunk is empty (end of file), it finalizes and returns the
         hash.
       - Otherwise, it updates the hash context with the new chunk and
         continues the loop.

    6. The loop starts with `BLAKE3.init` to create the initial hash context.

    7. By using lazy evaluation (`$!`), we ensure that the hash context is
       strictly evaluated in each iteration, preventing memory buildup.

    This implementation is memory-efficient and can handle very large files
    without issues, as it processes the file in small chunks rather than
    loading it entirely into memory.

    * This solution is already pretty good.
    * The explanation is also useful for people new to Haskell
      or this type of task.
    * Some explanations are subtly wrong, e.g. `$!` is the
      _opposite_ of lazy evaluation, and the sentence
      is contradicting itself.
    * It mainly got the imported module name wrong
      (should be `import BLAKE3` from
      https://hackage.haskell.org/package/blake3).
      Let's see if it fixes it when I tell it what I'd observe
      when trying to compile the code.
      A simple Google search for "haskell blake3" should point out the correct name.

    The mentioned `Crypto.Hash.BLAKE3` module doesn't exist as a Haskell library.

    import Crypto.Hash (hash, Blake2b_512, Digest)
    import qualified Data.ByteString as BS
    import qualified Data.ByteString.Lazy as LBS
    import System.IO (withFile, IOMode(ReadMode))

    hashFileBLAKE2b :: FilePath -> IO (Digest Blake2b_512)
    hashFileBLAKE2b filePath = do
      let chunkSize = 8192  -- 8KB chunks
      withFile filePath ReadMode $ \handle -> do
        let loop acc = do
              chunk <- BS.hGet handle chunkSize
              if BS.null chunk
                then return $ hash (LBS.fromChunks acc)
                else loop (chunk : acc)
        loop []

    * I asked for BLAKE3, but now it implemented BLAKE2B.
    * The solution is very wrong.
      It splits the file into _reversed_ chunks and then hashes them.
      The error is subtle and will be hard to debug.

But it certainly is fun and there's definitely some usefulness in there, especially for learning general concepts that are well understood.

I guess in the end it's this: Programmers don't need to "roughly get it right" (which is what current tools do well) when it comes to products. They need to get it /precisely/ right.

After a year of heavy LLM use I’ve found the utility limits, my usage has peaked, and I’m developing very restrictive use cases.

Beyond functioning as an interactive O’Reilly manual, LLM only save time if you never read the code they produce. Which is a short term win, but things will blow up eventually, as with all code, and now you’ve got a bigger problem than you started with.

A similar project is sandpack, but that relies on nodebox which is also closed source.

EDIT: Only seems to work in Chrome?

That’s probably what Ilya is doing.

(FWIW I don’t think we’re close to AGI).

For that, $20/M per head to be usable? Yikes.

BTW surprised that tarballs work - aren’t these compressed?

https://www.nature.com/articles/d41586-024-03162-2

https://www.nrc.gov/reading-rm/doc-collections/fact-sheets/3...

Jesus christ, I hope you are never in a position of any significant power

You're so edgy that you might cut yourself, be careful. What is wrong with making profit by helping people through providing a service?

1.) Those who use AI and talk about it.

2.) Those who do not use AI and talk about it.

3.) Those who use AI and talk about how they do not and will not use AI.

You don't have to look far to see how humans react to performance enhancers that aren't exactly sanctioned as OK (Steroids).

I really like copilot/ai when the focus was about hyper-auto-complete. I wish the integration was LSP+autocomplete+compilation check+docs correlation. That will boost my productivity x10 times and save me some brain cycles. Instead we are getting garbage UX/Backends that are trying to fully replace devs. Give me a break.

Write honestly. Write the way you write. Use your own flow, make your own grammatical wobbles, whatever they are. Express yourself authentically.

Don't let an AI do this to you.

Them releasing this stuff actually suggest they don't have much progress in their next model. It's a sell signal but today's investors have made their money in zirp, so they have no idea about the real world market. In a sense this is the market funneling money from stupid to grifter.

Yep. CAD, music, poetry, comedy. Same pattern in each.

But it's more than not being experts: it's about a subliminal belief that there either isn't much to be expert in or a denial of the value of that expertise, like if what they do can be replicated by a neural network trained on the description, is it even expertise?

Unavoidably, all of this stuff is about allowing people to do, with software, tasks they would otherwise need experts for.

In comparing it's performance to the pure model on Google AI studio I realized Gemini was presenting some sort of RAG results as the "answer" without disclosing where it got that information.

Perplexity, which is hardly perfect, will at least tell you it is searching the web and cite a source web page.

I'm basically saying Gemini fails at even the simplest thing you would want from a search tool: disclosing where the results came from.

3.5 Sonnet is fast, and that is very meaningful to iteration speed, but I find for the level of complexity I throw at it, it strings together really bad solutions compared to the more wholistic solutions I can work through with Opus. I use Sonnet for general knowledge and small questions because it seems to do very well with shorter problems and is more up-to-date on libraries.

Having it create a testing suite definitely helps. But it makes fewer mistakes than I would normally make... it's not perfect but it IS way better than me.

I think it has to do with Cursor's much better custom small models for code search/replace, but can't be sure.

https://github.com/patricktrainer/pbtree

- Mine prepends the result with the output of running `tree -I node_modules --noreport` before any other content. This informs the LLM of the structure of the project, which leads to other insights like it will know which frameworks and paradigms your project uses without you needing to explain that stuff. - Mine prepends the contents of each included file with “Contents of relative/path/to/file/from/root/of/project/filename.ts:” to reinforce the context and the file’s position in the tree.

For me it's still the cleanest editor.

VS Code is way too cluttered to be my daily driver for basic editing.

Good at integrating AI into a text editor != Good at building an IDE.

I worry about the ability for some of these VSCode forks to actually maintain a fork and again, I greatly prefer the power of IDEA. I’ll switch if it becomes necessary, but right now the lack of deep AI integration is not compelling enough to switch since I still have ways of using AI directly (and I have Copilot).

Vim has been around since the Stone Age.

Jokes aside, I don’t really see why ai tools need new editors vs plugins EXCEPT that they don’t want to have to compete with Microsoft’s first party AI offerings in vscode.

It’s just a strategy for lock-in.

An exception may be like zed, which provides a lot of features besides AI integration which require a new editor.

My project has well over half a million lines of code. I'm using an IDE (in my case Qt Creator) for a reason. I'd love to get help from an LLM but CLI or external browser windows just aren't the way. The overhead of copy/paste and lack of context is a deal breaker unfortunately.

In case I'm missing something, please let me know. I'm always happy to learn.

Luckily ChatGPT and the rest have nothing to do with an AI not to mention AGI.

The kind of apps that will be built in the next 5 years,are nowhere near what we have today.

Developers will need to update their skillset, though.

More seriously, the output quality of LLMs for code is pretty inconsistent. I think there's an analogy to be made with literature. For instance, a short story generated by an LLM can't really hold a candle to the work of a human author.

LLM-generated code can be a good starting point for avoiding tedious aspects of software development, like boilerplate or repetitive tasks. When it works, it saves a lot of time. For example, if I need to generate a bunch of similar functions, an LLM can sometimes act like an ad-hoc code generator, helping to skip the manual labor. I’ve also gotten some helpful suggestions on code style, though mostly for small snippets. It’s especially useful for refreshing things you already know—like quickly recalling "How do I do this with TypeScript?" without needing to search for documentation.

Anyway, literature writers and software engineers aren't going to be replaced anytime soon.

- There's the biggest/most high quality training corpus that captures all aspects of dev work (code, changes, discussions about issues, etc.) out there with open source hosting sites like GitHub

- Synthetic data is easy to generate and verify, you can just run unit tests/debugger in a loop until you get it right. Try doing that with contracts or tax statements.

- Little to no regulatory protections.

This is all about labor and capital. When people toss hundreds of billions at something it almost always is.

The social relationship doesn't have to be this way. Technological improvement could help us instead of screw us over. But we'd first have to admit that profit exploitation isn't absolutely the best thing ever and we'll never do that. Soooo here we are.

Maybe once you're deep into APIs that talk to other APIs, but near the surface where the data is collected there's nothing but "grounding to reality".

As my professor of Software Engineering put it: when building a system for counting the number of people inside a room most people would put a turnstile and count the turns. Does this fulfill the requirement? No - people can turn the wheel multiple times, leave through a window, give birth inside the room, etc. Is it good enough? Only your client can say, and only after considering factors like "available technology" and "budget" that have nothing to do with software.

Why would you hire? Either it works- in the sense of does the job and is cost effective- or it is not.

Is there a situation where paying 100's of k of wages makes a thing suddenly a good idea? I have doubts.

Canvas sounds useful. I'll be playing with that as soon as I can access it.

Another useful thing in chat gpt that I've been leveraging is its memory function. I just tell it to remember instructions so I don't have to spell them out the next time I'm doing something.

To me the current direction LLMs are headed seems like it will just further entrench the power of the trillion dollar megacorps because they’re the only people that can fund the creation and operation of this stuff.

Also, is it possible that smaller focused apps will have few edge cases and be more reliable?

And sometimes it breaks in ways I can't fix - so rolling back or picking a new patch from a know break point becomes important.

16 hours for my first azure pipeline, auto-updates from code to prod, static app including setting up git, vscode, node, azure creds etc. I chose a stack I have never seen at work (mostly see AWS) and I am not a coder. Last code was Pascal in the 1980s.

3rd app took 4 hours.

Built things I have wanted for 30 years.

But yes- no code understanding, brute force.

IF you don't have a central security team like in big companies or the need for an audit, what you don't know is what you don't care for.

Obviously until its too late but holy shit i have seen too much garbage just working for waaaay to long.

The fact that he liked and enjoyed coding was what actually prompted him into learning to code after that first experience.

ChatGPT shows a clear path forward. Feedback loop (consistent improvement), tooling which leverages all of llms powers, writing unit tests automatically and running code (chatgpt can run python already, when will it able to run java and other langauges?)

And its arleady useful today for small things. Copilot is easier and more integrated than googling parameters or looking up documentation.

UIs/IDEs like curser.ai are a lot more integrated.

What you see today is just the beginning of a something, potentially big.

`use matplotlib to generate an image with 3 bars of values 3, 6, 1`

followed by

`execute it`

https://chatgpt.com/share/66fefc66-13d8-800e-8428-815d9a07ae...

(apparently the shared link does not show the executed content, which was an image)

https://imgur.com/a/PkJCnKO

Which has interesting consequences, because I saw it self-execute code it generated for me and fix the errors contained in that code by itself two times until it gave me a working solution.

(Note that I am no longer a Plus user)

---

Claude: I apologize, but I don't have the ability to execute code or generate images directly. I'm an AI language model designed to provide information and assist with code writing, but I can't run programs or create actual files on a computer.

---

Gemini: Unfortunately, I cannot directly execute Python code within this text-based environment. However, I can guide you on how to execute it yourself.

---

> 4. Also you can use takeout

I just checked and wasn't able to takeout Gemini interactions. There are some irrelevant things like "start timer 5 minutes" which I triggered with my phone, absolutely unrelated to my Gemini chats. takeout.google.com has no Gemini section.