I've done a ton of refactoring, from Python to Node / Deno, and it's surprisingly good — but not perfect.
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.
The single file aspect of many of these answers is what grinds me as well. I mean, it's fantastic for a short script, a function/class template, or a crack at the syntax error... but it becomes a huuuuge pain even when it's something as simple as a .h for the file you're working on.
This is cool, but I wish it were integrated into tools already used for coding and writing rather than having it be a separate app.
This also demonstrates the type of things Google could do with Gemini integrated into Google Docs if they step up their game a bit.
Honestly I’m scratching my head on OpenAI’s desire to double down on building out their consumer B2C use cases rather than truly focussing on being the infrastructure/API provider for other services to plug into. If I had to make a prediction, I think OpenAI will end up being either an infrastructure provider OR a SaaS, but not both, in the long-term (5-10 yrs from now).
The most amazing thing with notebooklm is that is can turn your docs into a very high quality podcast of two people discussing the content of your docs.
Finding signal in noise is not an easy job given clip things are moving along. Whatever content creators need to do to deliver quality distilled content - I'm here for it.
I've only used the "Deep Dive" generator a few times, and I'm already sensing the audio equivalent of "youtube face" in the style — not saying that's inherently bad, but this is definitely early days for this kind of tool, so consider Deep Dive as it is today to be a GPT-2 demo of things to come.
Do you have a reference for the "Juggling dog" thing? I've heard it with "singing dog", but I never managed to find any "official" reference or explanation of the thing.
He meant singing dog, likely conflated due to his linguistic interest.
"Juggling dog" has only been expressed a single time previously in our corpus of humanity:
During the Middle Ages, however, church and state sometimes frowned more sternly on the juggler. "The duties of the king," said the edicts of the Sixth Council of Paris during the Middle Ages, "are to prevent theft, to punish adultery, and to refuse to maintain jongleurs."(4) What did these jugglers do to provoke the ire of churchmen? It is difficult to say with certainty, since the jongleurs were often jacks-of-all-trades. At times they were auxiliary performers who worked with troubadour poets in Europe, especially the south of France and Spain. The troubadours would write poetry, and the jongleurs would perform their verses to music. But troubadours often performed their own poetry, and jongleurs chanted street ballads they had picked up in their wanderings. Consequently, the terms "troubadour" and "jongleur" are often used interchangeably by their contemporaries.
These jongleurs might sing amorous songs or pantomime licentious actions. But they might be also jugglers, bear trainers, acrobats, sleight-of-hand artists or outright mountebanks. Historian Joseph Anglade remarks that in the high Middle Ages:
"We see the singer and strolling musician, who comes to the cabaret to perform; the mountebank-juggler, with his tricks of sleight-of-hand, who well represents the class of jongleurs for whom his name had become synonymous; and finally the acrobat, often accompanied by female dancers of easy morals, exhibiting to the gaping public the gaggle of animals he has dressed up — birds, monkeys, bears, savant dogs and counting cats — in a word, all the types found in fairs and circuses who come under the general name of jongleur.”(5)
-- http://www.arthurchandler.com/symbolism-of-juggling
"Dogs were not aware of their shared interest in juggling until the invention of the internet, where like-minded canines would eventually congregate unto enclaves of specialty."
I suspect what I heard was a deliberate modification of this sexist quote from Samuel Johnson, which I only found by this thread piquing my curiosity: "Sir, a woman's preaching is like a dog's walking on his hind legs. It is not done well; but you are surprised to find it done at all." - https://www.goodreads.com/quotes/252983-sir-a-woman-s-preach...
Trying to find where I got my version from, takes me back to my own comments on Hacker News from 8 months ago, and I couldn't remember where I got it from then either:
> "your dog is juggling, filing taxes, and baking a cake, and rather than be impressed it can do any of those things, you're complaining it drops some balls, misses some figures, and the cake recipe leaves a lot to be desired". - https://news.ycombinator.com/item?id=39170057
Trying to find where I got my version from just brought me back to one of my own comments on Hacker News from 8 months ago:
> "your dog is juggling, filing taxes, and baking a cake, and rather than be impressed it can do any of those things, you're complaining it drops some balls, misses some figures, and the cake recipe leaves a lot to be desired". - https://news.ycombinator.com/item?id=39170057
I couldn't remember where I got it from then either.
He is adapting one of Samuel Johnson's most famous quotations, about the astonishing sight of seeing a woman preaching - like a dog walking, it may not be done well, but it's astonishing to see it done at all.
This feature is cool as fuck, but I noticed that podcasts it generates loose quite a lot of details from the original article. Even longreads turn into 13 mins chunks.
It is a cool concept, but anyone who listens to enough podcasts know that hosts have personalities and interests, and productions usually have their styles, focus and quality. These features make podcast channels unique and make you want to come back. That's why you may want to listen to podcast A instead of B even though they discuss the same topics. I doubt the Google thing will ever give us that -- likely just one hour of generic rambling that gets boring.
> Honestly I’m scratching my head on OpenAI’s desire to double down on building out their consumer B2C use cases rather than truly focussing on being the infrastructure/API provider for other services to plug into
I think it's because LLMs (and to some extent other modalities) tend to be "winner takes all." OpenAI doesn't have a long term moat, their data and architecture is not wildly better than xAI, Google, MS, Meta, etc.
If they don't secure their position as #1 Chatbot I think they will eventually become #2, then #3, etc.
At the moment this feels like a x10 speed run on the browser wars: lots of competitors very quickly churning who is "best" according to some metric, stuff getting baked into operating systems, freely licensed models.
How do you make money off a web browser, to justify the development costs? And what does that look like in an LLM?
LLMs are a more flexible platform than browsers. They can be prompted, finetuned or run locally. Even if a company wants to make their base model spit ads, it won't fly.
Depends how subtle they are about it, and what the rest of the ecosystem looks like.
Perhaps the ad/ad-blocker analogy would be: You can have the free genuinely open source LLM trained only on Wikipedia and out-of-copyright materials, or you can have one trained on current NYT articles and Elsevier publications that also subtly pushes you towards specific brand names or political parties that paid to sponsor the model.
Also consider SEO: every business wants to do that, nobody wants to use a search engine where the SEO teams won. We're already seeing people try to do SEO-type things to LLMs.
If (when) the advertisers "win" and some model is spitting out "Buy Acme TNT, for all your roadrunner-hunting needs! Special discount for coyotes!" on every other line, then I'd agree with you, it won't fly, people will switch. But it doesn't need to start quite so bold, the first steps on this path are already being attempted by marketers attempting to induce LLMs crawling their content to say more good things about their own stuff. I hope they fail, but I expect them to keep trying until they succeed.
Google and Facebook grew organically for a number of years before really opening the tap on ad intrusions in to the UX. Once they did, a tsunami of money crashed over both, quarterly.
The LLM companies will have this moment too.
(But your post makes me want to put a negative-prompt for Elsevier publications in to my Custom Instructions, just in case)
There is huge choice in open models. People won't adopt one with ads baked in, unlike Google and Facebook, because now there are more options. There are 100K LLM finetunes on HuggingFace.
I've got some of them on my experimentation laptop. They're only good enough to be interesting, not good in comparison to the private models, and the number of fine-tunes doesn't help with that. In particular I've had Microsoft's Phi 3.5 for less than a week and yet I've already had at least 4 cases of it spouting wild nonsense unrelated to the prompt — and I don't even mean that it was simply wrong, I mean the response started off with Chinese and then acted like it was the early GPT-3 "Ada" model doing autocomplete.
One of my machines also has a copy of Firefox on it. Not used that in ages, either. But Firefox is closer in quality to Chrome, than any of the locally-runnable LLMs I've tried are to the private/hosted LLMs like 4o.
When they are focusing on just being an API provider then they will be in a market with (long term) razor thin margins and high competition - most likely unable to build a deep moat. But if you can shape customers habits to always input "chatgpt.com" into the browser whenever they want to use AI then that's a very powerful moat. Those customers will also most likely be on a subscription basis, meaning much more flexibility in pricing and more rent for openAI (people using it less then what OpenAI calculates for subscription costs).
From Wikipedia, for that don’t know the term: “a concept in economics that describes a process in which new innovations replace and make obsolete older innovations.”
Ironically, I had to google it, and agree with the comment.
You should read The Innovator's Dilemma as well, as it goes into detail on this concept, basically explaining why and how technological disruption occurs from the point of view of the disruptor and disruptee.
Google isn't a startup, they aren't desperate to impress anyone. I don't even think they consider "AI" to be a product, which is probably correct. These AI enabled features are background processes that ideally integrate into products over time in ways that don't require you to explicitly know they're even there.
Given how widely used Google Docs is, for serious work, disrupting people's workflows is not a good thing. Google has no problem being second, they aren't going to die in the next three months just because people on Twitter say so.
Same here. I feel like Google's products have become such a labyrinth of features, settings, integrations, separate (but not really) products, that navigating them requires an expert. Sadly, I don't see a way back - each new additional feature or product is just bolted on top and adds more complexity. Given the corporate structure of Google, there's zero chance of an org-wide restructuring of the labyrinth.
75% of OpenAI's revenue is coming from their consumer business - the better question is the long term viability of their public API.
But if they believe they're going to reach AGI, it makes no sense to pigeonhole themselves to the interface of ChatGPT. Seems like a pretty sensible decision to maintain both.
75%? Thats astonishing to me. Where are you able to see those details?
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.
It may be pretty minimal but i can personally vouch for 20ish techies in my own social orbit who's businesses wont authorise or wont pay for OpenAI yet and are doing so out of their own pockets; i share an office with four of them.
Maybe the consumer side will slide as businesses pick up the tab?
This very good analysis estimates 73%, which includes team and enterprise. Given that enterprise access is limited and expensive, it seems Plus and Teams are mostly carrying this.
The whole financial breakdown is fascinating and I’m surprised to not see it circulating more.
Your source is a blog post by a polemic author whose own source is second-hand by NYT, an organization that is in lawsuit with OpenAI. I would have rather have heard it from the horse's mouth. What financial information about OpenAI does NYT have that I don't? Do they have privileged access to private org financials?
In my estimation, you're not qualified for this conversation.
ChatGPT itself is them copying their own API users, this is just them building out more features already built by users. My guess is they know they don't have a long term edge in models alone, so they are going to rely on expanding ChatGPT for better margins and to keep getting training data from users. They obviously want to control the platform, not integrate with other platforms
I only use Gemini in Colab perhaps 5% of the times I use Colab, yet it is nice to have.
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.)
LLM as a service is much easier to replicate than physical data centers and there's a much lower potential user base than consumers, so I'd imagine they're swimming upstream into B2C land in order to justify the valuation
I wouldn't be surprised to see Apple add something like this to Pages and some of their other apps. Their approach to AI, from what we've seen so far, has been about integrating it into existing apps and experiences, rather than making a separate AI app. I have to imagine this is the way forward, and these stand alone apps are basically tech demos for what is possible, rather than end-state for how it should be consumed by the masses.
I agree with you on where OpenAI will/should sit in 5-10 years. However, I don't think them building the occasional tool like this is unwarranted, as it helps them show the direction companies could/should head with integration into other tools. Before Microsoft made hardware full time, they would occasionally produce something (or partner with brands) to show a new feature Windows supports as a way to tell the OEMs out there, "this is what we want you to do and the direction we'd like the PC to head." The UMPC[0] was one attempt at this which didn't take off. Intel also did something like this with the NUC[1]. I view what OpenAI is doing as a similar concept, but applied to software.
Every app with a significant installed user base is adding AI features.
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…
How many people use Pages these days? I don't think Apple even mentions the product in their WWDC these days. My guess is that most people either use Microsoft suite as required by their employer or use cloud based knowledge base/notes tools like Notion/Quip/Obsidian/Confluence etc. I doubt Apple thinks it worthwhile to invest in these products.
People who need to make the occasional document outside of work, who don’t need to invest in paying for Office, use iWork. I count myself in that list. I use Office at work (99% of that usage is Excel), but at home I use the iWork apps. Mostly Numbers, but Pages as well. I hear many of my friends and family doing the same, because it’s what they have, it’s good enough, and it’s free.
Few people outside of tech circles know what those other apps you mentioned are. I use Confluence at work, because it’s what my company uses. I also tried using it at home, but not for the same stuff I’d use Pages for. I use Obsidian at work to stay organized, but again, it doesn’t replace what I’d use Pages for, it’s more of a Notes competitor in my book. A lot of people don’t want their documents locked away in a Notion DB, and it’s not something I’d think to use if I’m looking to print something.
I went back and looked at the last WWDC video. Apple did mention the apps briefly, to say they have integrated Image Playgrounds, their AI image generation, into Pages, Keynote, and Numbers. With each major upgrade, the iWork apps usually get something. Office productivity isn’t exactly the center of innovation these days. The apps already do the things that 80% of users need.
> but I wish it were integrated into tools already used for coding
Unless I'm missing something about Canvas, gh CoPilot Chat (which is basically ChatGPT?) integrates inline into IntelliJ. Start a chat from line numbers and it provides a diff before applying or refining.
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...
According to this (1), they are using the 4o model. And looks like you'll be able to pick your model(2) in the starting with version 1.94 released this September.
I check the GitHub blog[0] from time to time. They also have a RSS feed if you'd prefer that. The is also a waitlist for o1 access you may sign up for[1]
> the type of things Google could do with Gemini integrated into Google Docs
Google already does have this in Google Docs (and all their products)? You can ask it questions about the current doc, select a paragraph and ask click on "rewrite", things like that. Has helped me get over writer's block at least a couple of times. Similarly for making slides etc. (It requires the paid subscription if you want to use it from a personal account.)
That's there too; see https://support.google.com/docs/answer/14206696 — you can click on the "Ask Gemini ⟡" and carry on a conversation, e.g. "summarize emails about <topic>" and use those to paste into the doc. (I haven't found all that much use for referencing other files though. But the "proper chat" is useful for saying things like "no actually I meant something more like: …" and carrying on.)
> demonstrates the type of things Google could do with Gemini integrated into Google Docs
Or Microsoft!
> think OpenAI will end up being either an infrastructure provider OR a SaaS, but not both
Microsoft cut off OpenAI's ability to execute on the former by making Azure their exclusive cloud partner. Being an infrastructure provider with zero metal is doable, but it leaves obvious room for a competitor to optimise.
Their API is unusable due to rate limits. Myself and my wife have both had ideas, started using it, and found other approaches after hitting rate limits. I tried funding more money in the account to increase the rate limits and it did not work. I imagine they see poor growth there because of this.
You need to use it for some time to get into their higher tiers of usage. I used to also have this problem and it annoyed me greatly, but once I got to usage tier 4 it never happened again (except for o1-preview but that just wastes tokens IMO).
It's pretty trivial to get increased limits, I've used the API for a few consulting projects and got to tier 4 in a month. At that point you can burn near $200 a day and 2 million tokens per minute.
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.
But my subscription at $20/mo is a fraction of my API usage at $5/day (about $100/mo).
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 suspect they are building their B2C products because it gives them better data to train on. It's a lot harder to control the quality of data when you have no idea how API inputs were produced, what the UI is like, or who the users are. You don't know the provenance of the data, or the context. Or even if multiple unrelated client products are being commingled through the same key.
If you control the UI, you have none of those problems.
To be honest I think they’re having less success than it appears with their B2B offerings. A lot of cloud providers services like AWS have their own things they sell through those channels and I think a lot of businesses are finding those solutions to be cheaper and “good enough”
I think this is already built into Microsoft's Office365 "CoPilot" (which I assume is a ChatGPT frontend. You can ask the AI to make changes to your Office documents.
If I'm reading this right; it's been in VSCode as Copilot Chat for a fair bit now. I use it often, when they added context (provide extra files to reference or even the entire @workspace if it's small enough), absolute gamechanger.
Aren't we talking about, say, GitHub Copilot? That's integrated into Visual Studio/VSCode. I just started using it again as they've done some small upgrades, and the results can often be phenomenal. Like, I will visualize an entire block of code in my mind, and I'll type the first couple of characters and the entire block will just appear. I'm literally that predictable.
Copilot is only using GPT3.5 for most of the results though, seemingly. I'd be more excited if they would update the API they're using.
Do they not understand that the example text they are using in the first image is so laughably banal that it makes an entire segment of its potential audience not want to engage at all?
Shoot me in the face if my own writing is ever that bad.
ETA: just to be clear... I am not a great writer. Or a bad one. But this is a particular kind of bad. The kind we should all try to avoid.
I am trying to convince myself that I am not insane and everyone else is. The platform was literally down for me for a good 12 hours or so because they had an auth problem or bug. Their interface is subpar yet they are trying to convince people that this is replacing knowledge worker any minute now. I recommended to a friend that he uses chatGPT to write some English content and it did a bad job. I checked bolt yesterday and the code it produced for a very simple app was complete garbage hallucination.
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.
Garbage in, garbage out. It is not going to imagine your perfect scenario and then create it for you. I take anyone saying it is garbage with a grain of salt because it is incredibly useful for me. And others think so too, so how can your bad experience negate that. It can't. If you can craft the right prompts it can make you much more efficient. Anyone saying it is going to replace whole people en masse is just part of the hype machine. But if all it does is make every human on earth 1% more efficient then that is an obscene amount of value it is creating.
I'm with you. I feel like I'm losing my mind. Everyone around me is talking about the looming AGI, death of the knowledge worker and how "everything" has changed. But every time I try to use these text generators I get nothing useful from them. It's like the whole world has bought into a mass hallucination.
It makes more sense when you realize that while sure, there might be slight variation in output, generally speaking the people tripping over themselves in how [current version] is so amazing aren't being totally honest about why they think it's amazing.
For them, the ability to generate so much trash is the good part. They might not even be fully aware that it's trash, but their general goal is to output more trash because trash is profitable.
It's like all those "productivity systems". Not a single one will produce a noticeable increase in productivity magically that you can't get from just a $1 notebook, they just make you feel like you are being more productive. Same with RP bots or AI text editors. It makes you feel so much faster, and for a lot of people that's enough so they want in on a slice of the AI moneypit!
Its a tool, like any other tool a software developer would use. In areas where I have a lot of repetition or need to pour through verbose (but simple) documentation, its such a game changer. I can spend 5 minutes thinking about what I want the machine to do, give it some samples of what I expect the output to be and wala, it generates it, often times 100% correct if I've got the prompt put in properly, sometimes its good enough with a bit of refinement. This is something I would normally have delegated to a junior team member or sub-contractor, but now I'm saving in time and money.
Occasionally I sink 1-2 hours into a tweaking something I thought was 90% correct but was in reality garbage. I had that happen a lot more with earlier models, but its becoming increasingly rare. Perhaps I'm recognizing the limitations of the tool, or the systems indeed are getting better.
This is all anecdotal, but I'm shipping and building faster than I was previously and its definitely not all trash.
Person A: Me try make this code work but it always crash! maybe the server hate or i miss thing. any help?
Person A with AI: I've been trying to get this code to work, but it keeps crashing. I'm not sure if I missed something or if there's an issue with the server. Any tips would be appreciated!
For a non-native English speaker, it's much better professionally to use AI before sending a message than to appear authentic (which you won't in another language that you aren't fluent so better to sound robotic than write like a 10 years old kid).
Person A with AI:
In the bustling world of software development, where lines of code intertwine to create the intricate tapestry of our digital lives, I find myself facing a challenge that has proven to be both perplexing and frustrating. I’ve spent over a decade honing my skills as a developer. Known for my analytical mind and commitment to excellence, I’ve navigated various programming languages, frameworks, and projects that I’m proud to have contributed to.
Recently, I stumbled upon a bug that initially seemed minor but quickly revealed itself to be a formidable adversary. It disrupted the seamless user experience I had meticulously crafted, and despite my best efforts, this issue has remained elusive. Each attempt to isolate and resolve it has only led me deeper into a labyrinth of complexity, leaving me frustrated yet undeterred.
Understanding that even the most seasoned developers can hit a wall, I’m reaching out for help. I’ve documented the symptoms, error messages, and my various attempts at resolution, and I’m eager to collaborate with anyone who might have insights or fresh perspectives. It’s in the spirit of community and shared knowledge that I hope to unravel this mystery and turn this challenge into an opportunity for growth.
It's pretty good for native English speakers at work who need/want a reverse anger translator.
Me: This is the most garbage code I've ever seen. It's bad and you should feel. It's not even wrong. I can't even fathom the conceptual misunderstandings that led to this. I'm going to have to rewrite the entire thing at this rate, honestly you should just try again from scratch.
With AI: I've had some time to review the code you submitted and I appreciate the effort and work that went into it. I think we might have to refine some parts so that it aligns more closely with our coding standards. There are certain areas that are in need of restructuring to make sure the logic is more consistent and the flow wouldn't lead to potential issues down the road.
I sympathize with the sibling comment about AI responses being overly-verbose but it's not that hard to get your model of choice to have a somewhat consistent voice. And I don't even see it as a crutch, this is just automated secretary / personal assistant for people not important enough to be worth a human. I think a lot of us on HN have had the experience of the stark contrast between comms from the CEO vs CEO as paraphrased by their assistant.
Aw thanks! I at least have the benefit of being a fluent writer.
For lots of East Asian researchers it's really embarrassing for them to send an email riddled with typos, so they spend a LOT of time making their emails nice.
> For lots of East Asian researchers it's really embarrassing for them to send an email riddled with typos, so they spend a LOT of time making their emails nice.
OK -- I can see this. But I think Grammarly would be better than this.
It does now, perhaps, for complete rewrites. I've not looked recently.
But its suggestion system, where it spots wordy patterns and suggests clearer alternatives, was available long before LLMs were the new hotness, and is considerably more nuanced (and educational).
Grammarly would take apart the nonsense in that screenshot and suggest something much less "dark and stormy night".
Thanks for saying this. Whenever Grammarly puts a red line under a slightly superflouos part of the sentence I get more and more agitated at this small nudging to robotic writing.
There's more to writing than clarity, though. Not all written communication needs to abide of the efficient/clear writing style of technical documentation FFS
But there's not much more important, stylistically, to writing an business email or document than clarity. It's absolutely the most important thing. Especially in customer communications.
In the UK there is/used to be a yearly awards scheme for businesses that reject complexity in communucations for clarity:
But anyway, you don't have to act on all the suggestions, do you? It's completely different from the idea of getting an AI to write generic, college-application-letter-from-a-CS-geek prose from your notes.
With enough repetitive suggestions asking for the same thing, it will just continuously push your writing style towards this ultra-dry writing. Plus, even in business emails it's important to show a human side in writing.
It's not like Grammarly's push for clear writing actually helps in any way. Most times it just outright suggests removing relevant info from the sentence. They just push for this service as a way to incentivise subscriptions, writing quality be damned.
They plateaued on model performance and they are hype based. They need to keep the momentum going by "releasing" stuff, so they are garbage out at the moment. Given that open weight models are so close to gpt-4, their value is exactly 0 unless they can produce a new model with a significant jump in coherence.
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.
I thought the same thing: the “blog post” in the example image is an example of the absolute trash that’s being spewed onto the internet by these tools. 10+ sentences and yet somehow nothing actually said.
I see this all the time from AI boosters. Flashy presentation, and it seems like it worked! But if you actually stare at the result for a moment, it’s mediocre at best.
Part of the issue is that people who are experts at creating ML models aren’t experts at all the downstream tasks those models are asked to do. So if you ask it to “write a poem about pizza” as long as it generally fits the description it goes into the demo.
We saw this with Gemini’s hallucination bug in one of their demos, telling you to remove film from a camera (this would ruin the photos on the film). They obviously didn’t know anything about the subject beforehand.
> Part of the issue is that people who are experts at creating ML models aren’t experts at all the downstream tasks those models are asked to do.
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.
Well, comedians still exist, despite the fact that ChatGPT can write an endless stream of “jokes” for next to zero cost. So do musicians. I know less about poetry and CAD but I assume people who seek out those modalities aren’t going to be impressed with generic garbage. A person who seeks out poetry isn’t going to be easily impressed.
No. But then all of these products are marketed to people who are, at some domain-specific level, still towards the "but I wore the juice!" end of the scale, right?
Unskilled and unaware of it. Or rather, unskilled and unaware of what a skilled output actually involves. So, unaware of the damage they do to their reputations by passing off the output of a GPT.
This is what I mean about the writing, ultimately. If you don't know why ChatGPT writing is sort of essentially banal and detracts from honesty and authenticity, you're the sort of person who shouldn't be using it.
(And if you do know why, you don't need to use it)
Looks like you're missing an AI to auto-close the PRs for you.
Seriously though, I'm tired of the "helpful" GitHub bots closing issues after X days of inactivity. Can't wait for one powered by AI to decide it's not interested in your issue.
Those bots are great - the maintainers dont bother fixing the issue, so it goes stale, gets auto-closed, and boom! Metrics go up. Another issue successfully closed! Another bug fixed! Until someone else opens a new issue for it, which can then be closed as duplicate. I love it!
It seems sort of weird to keep pushing the chat interface so hard into programming. For 'real' usage, it seems like Cursor or Aider approaches work better, since you end up having AI write code, you manually edit, AI updates further, and back and forth. In a chat interface, copy/pasting updated code gets old fast.
On the other hand, I did have good luck w/ Anthropic's version of this to make a single page react app with super basic requirements. I couldn't imagine using it for anything more though.
I'm really happy to see ChatGPT doing this. The idea of a canvas made me really enjoy using Claude as I felt it to be the (so far) most "appropriate interface" to AI Chatbots as you are often doing two unique things in an AI chat:
- holding in your mind a "thing" (i.e. some code)
- talking about a "thing" (i.e. walking through the code)
The same applies to non-code tasks as well. The ability to segregate the actual "meat" from the discussion is an excellent interface improvement for chatbots.
Why do you ask? I did use "4o with canvas" shortly after writing the above. To be clear, my original comment was not about the actual OpenAI implementation of a "canvas style chatbot", but rather that I have found the canvas-chat-UX to be the most effective way to interact with a chatbot. I am basing this on many hours with both ChatGPT (non-canvas until today) and Claude/Anthropic (who has had this feature). As of this writing, I prefer Claude both because of (a) the content of its output and (b) the canvas style, which allows my brain to easily parse what is the topic vs. discussion about the topic.
Not the op, but I just tried it. I agree with his point that this is a huge step up in having the discussion and then a separate canvas where the work takes place, and then iterations are done to the canvas.
I LOVE the UX animation effect ChatGPT added to show the canvas being updated (even if it really is just for show).
Slightly OT, but one thing I noticed further into the demo is how you were prompting.
Rather than saying “embed my projects in my portfolio site” you told it to “add an iframe with the src being the project url next to each project”. Similarly, instead of “make the projects look nice”, you told it to “use css transforms to …”
If I were a new developer starting today, it feels like I would hit a ceiling very quickly with tools like this. Basically it looks like a tool that can code for you if you are capable of writing the code yourself (given enough time). But questionably capable of writing code for you if you don’t know how to properly feed it leading information suggesting how to solve various problems/goals.
> Basically it looks like a tool that can code for you if you are capable of writing the code yourself (given enough time).
Yes, exactly. I use it the way I used to outsource tasks to junior developers. I describe what I need done and then I do code review.
I know roughly where I want to go and how to get there, like having a sink full of dirty dishes and visualizing an empty sink with all the dishes cleaned and put away, and I just instruct it to do the tedious bits.
But I try and watch how other people use it, and have a few other different styles that I employ sometimes as well.
> I use it the way I used to outsource tasks to junior developers.
Is this not concerning to you, in a broader sense? These interactions were incredibly formative for junior devs (they were for me years ago) - its how to grew new senior devs. If we automate away the opportunity to train new senior devs, what happens to the future?
I wasn't. I saw this post on HN, opened a new tab for ChatGPT, and saw that I had access to the model. I assume it's rolling out incrementally over a few hours to all paid users.
As another data point: I wasn't notified either. I didn't have it when I first read about Canvas here earlier but coming back to the topic now I do have the feature. This is for a Teams account, the post says Plus and Teams are the first to get it rolled out with Enterprise and other in a week.
The easiest way to check if you have access is it will appear as an explicit choice in the "Model" selector.
The most surprising part of this announcement was the team who worked on this — more people worked on and contributed to this than many startups. There are 16 people working on this project!!
If each was paid $300k (that's a minimum...) and they spent a year on this, it'd make it a $5M project...
Here's an idea: If AI like this is so brilliant and can think for itself, why don't we just tell it to come up with its own next iteration? Surely if it can write code for medical devices, cars, planes, etc. (where no doubt junior engineers are extensively using it), then why not AI?
Cant we just tell ChatGPT to make e.g. TensorFlow faster, better, cleaner? Why do people put in so much work anymore, if the AI is so damn good?
I believe it's not solving a real problem. I believe that the human skills of reasoning, logical thinking, etc. make it possible for any able minded human to do the things I just mentioned, given time and money is provided. AI can't do that. Let's call that process "programming". It cant do programming. It pretends to program.
You are entitled to interpret the word "programming" in multiple different ways depending on the context of the conversation. What you propose here as "programming" is valid but not comprehensive of what everyone may think of when considering the "programming" process.
By many definitions of "programming", these AI tools are indeed programming. In the same way, many definitions of "reasoning, ..." may include the AI tools in them. However, there are alternate definitions (reasonable in their own way) in which it is clear these AI tools fall short.
So, I think you are proposing an argument of semantics but presenting it as if it is an argument on actual capabilities of these AI tools. In general, we all are in agreement on what these systems are capable of.
These AI tools are able to do a small subset of what a professional human is able to do, with greatly reduced flexibility, e.g. tasks like leetcode where there is an abundance of concentrated training data.
I would bet that the current paradigm in ChatGPT will never be able to replace a minimally competent human at real world programming - the kind of programs people actually pay for. Let’s see.
> I would bet that the current paradigm in ChatGPT will never be able to replace a minimally competent human at real world programming - the kind of programs people actually pay for. Let’s see.
Depends on definition of "minimally competent". There WILL be over-engineered enterprise solutions that employ 10x more AI-enabled code monkeys than is strictly required. Think about it: we can create a "fully automated AI coder" and then hire another FTE (or team) to handhold and clean up after it. It will be done. It must be done.
Why must LLMs or “AI” beat or match the smartest and most capable humans to be considered to solve a real problem? There’s been a lot of technology invented and in widespread use that solves real problems without having human-like intelligence.
It seems like this only supports "JavaScript, TypeScript, Python, Java, C++, or PHP". I would be so happy if it worked with FeatureScript (which is similar to JavaScript, but is for 3D CAD in Onshape).
I wonder what it would take to expand the languages it supports?
As with anything else that is helpful, there is a balancing act to be aware of. This is too much for my taste. Just like github copilot is too much.
It's too dumb like this. But chatgpt is insanely helpful in a context where I really need to learn something I am deep diving into or where I need an extra layer of direction.
I do not use the tool for coding up front. I use them for iterations on narrow subjects.
I haven’t used it yet, but couldn’t you just copy paste a chunk of existing code into canvas and have it help there? If so, that does seem more useful than the original of just pasting a lot of code into chat and hoping it pulls the correct context from your description. If I’m understanding it correctly, I’m canvas you can paste a bunch of code and then ask for help on specific contexts within the code by highlighting it. If done properly that seems super useful to me.
Don't fight this. Try to profit from it. People love these tools and they will become utterly, utterly dependent.
Using a spell-checker, I have gradually lost my ability to spell. Using these LLM tools, large parts of the population will lose the ability to think. Try to own them like farm animals.
The large number of tokens being processed by iterative models requires enormous energy. Look at the power draw of a Hopper or Blackwell GPU. The Cerebras wafer burns 23 KW.
One avenue to profit is to invest in nuclear power by owning uranium. This is risky and I do not recommend it to others. See discussion here: https://news.ycombinator.com/item?id=41661768
Speaking of energy use. Microsoft is literally in talks to restart a nuclear plant where they will buy all the power[1]. It happens to be "Three Mile Island" plant [2]
The Sprott Physical Uranium Trust owns 65,711,826 pounds of triuranium octoxide (U3O8) stored at uranium hexafluoride (UF6) conversion facilities in Canada, France, and the United States.
Cameco Corporation, ConverDyn, and Orano Chimie-Enrichissement
individually act as custodians on behalf of the Trust for the
physical uranium owned by the Trust.
If as GP says, people are actively choosing to become passive and dependent like farm animals, is it immoral to become the farmer? The alternative is to join them. It’s not in anyone’s power to force the rest of society to “be free.”
I think you should try to give tools like this another chance. If Andrej Karpathy can say AI-assisted programming is a productivity boost for him (https://x.com/karpathy/status/1827143768459637073), it can be a productivity boost for probably any programmer.
Research contributions are almost entirely separate from engineering quality. This is true for the vast vast vast majority of work people do in grad school.
This is a ridiculous statement. It's possible he isn't an absolutely amazing programmer*, but given he's probably in the top 0.01% in terms of implementing novel (and often groundbreaking) ideas via programming and that he finds AI-assisted programming tools useful, it suggests AI-assisted programming is now more than just a CRUD code monkey assister.
openai gpt is a service not a product but canvas is the first product that
openai build. i guess we will see new products in future. canvas is a new product, thats why they didn't introduce it at devday.
i think chatgpt is not a product but a service like dropbox. i don't say it is useless or can't be sold. it is just a feature for a product. GPT, in itself, is a powerful technology or tool that enhances user experiences in various applications. It provides natural language processing capabilities like answering questions, generating text, assisting with tasks, and so on. However, without a specific context or integration into a larger platform, GPT is more akin to a feature that can improve or enable specific functionalities within products.
I don’t understand what distinction you’re trying to make. Dropbox is a product too. People buy subscriptions for ChatGPT, that’s the product they want. It could be based on whatever model, that’s just the platform used to build the product. Users want the brand and the UX of ChatGPT, the whole bundle. That’s the product OpenAI sells
Slick interface but the example they gave is depressing.
We taught the model to open a canvas for prompts like “Write a blog post about the history of coffee beans”.
If you're not heavily editing this post to say something genuinely new, then congratulations you've added even more drivel to the Great Pacific Garbage Patch of the internet.
Some people really think that they are now "content writers" with ChatGPT, just as those spitting out sort-of-working scripts think they are "software engineers".
I am sorry, dears, but this is not how it works. To be good at any of this, you should be able to do it yourself without any "prompt engineering", and the only path is through work, time, trial/error, and tons of frustration.
I get the argument against AI tools when it's about them not working as well as the hype says they do, but not when they are the "how dare you use a tool that makes it easier for you! That's cheating!" argument. When I was in school cheap pocket calculators were just becoming available. All of my teachers banned them as they saw them as an existential threat to teaching math and science. These days students are generally allowed calculators -- the teachers finally accepted that it's a good thing that tools can automate the rote parts so that teaching can move on to the more interesting and thought-provoking parts.
I'm playing around with this right now and it's pretty sweet. It real-time shows which lines it's "thinking" about working and feels very dynamic, like I'm working with a machine in real-time.
It can't display markdown and formatted code side-by-side which is kind of a surprise.
I haven't tried doing anything super complex with it yet. Just having it generate some poems, but it's smart enough to be able to use natural language to edit the middle of a paragraph of text without rewriting the whole thing, didn't notice any issues with me saying "undo" and having data change in surprising ways, etc. So far so good!
I'm not very skilled at creating good "test" scenarios for this, but I found this to be fun/interesting: https://i.imgur.com/TMhNEcf.png
I had it write some Python code to output a random poem. I then had it write some code to find/replace a word in the poem (sky -> goodbye). I then manually edited each of the input poems to include the word "sky".
I then told it to execute the python code (which causes it to run "Analyzing...") and to show the output on the screen. In doing so, I see output which includes the word replacement of sky->goodbye.
My naive interpretation of this is that I could use this as a makeshift Python IDE at this point?
Is it possible to have the local context be a directory and all files within it or something? Ie to just ask it questions on the side of your WIP repo? Use your normal editor/etc.
Why don't companies learn from the really best times of Apple; announce; immediately available. Sure I know why but that used to be why I liked them. This marketing grift is terrible.
Funny timing. StackBlitz announced Bolt.new (https://bolt.new/) today with multi-file edit, emulated filesystem, arbitrary npm installs, and is open source. I feel ChatGPT is still chasing after Claude 3.5 artifact.
Has anyone had much experience with it, that can share their findings? I'm happy with Claude Sonnet and can't try every new AI code tool at the rate they are coming out. I'd love to hear informed opinions.
Just tried it, it wrote promising code but in the end only the last file was created and the other files had no content (even though I saw them being written). Seems a bit bugged at the moment.
I tried Bolt this morning for about 5 minutes, and it did output valid React code, but things like inputs and buttons it created didn't work without additional prompting, and then it started throwing "There was an error processing your request"
You get tighter feedback loops and better products if you own the vertical.
OpenAI already A/Bs test the responses it generates. Imagine if they own the text editor or spreadsheet you work on too. It’ll incorporate all of your edits to be self-correcting.
Excuse my possible ignorance but I had a question: Does ChatGPT have 'up to date knowledge' in terms of latest versions of software?
When I have used it I asked it to help me with VueJS but whilst it knew about version 3 (which is substantially different to V2) it didn't know the very latest version.
Isn't this going to be problematic if it's trying to help code but has outdated information? Won't this result in more code being written/tied to older languages/framework versions?
In my experience, because GPT outputs code tied to particular versions of software, I use those as it's far quicker to copy/paste/edit than reading the latest docs, where GPT is absolutely useless as it simply doesn't know about the latest code and isn't able to produce working snippets.
So the real deal would be if GPT knows about the latest software instantly the moment it's published (ingesting docs data, any stackoverflow Q/A's etc).
Here is what GPT 4o-mini told me
P: can you list me the latest verisons of ALL major languages such as python, node etc. and the latest versions of typescript, angular, vuejs etc
R:
> Python: 3.12.0 (released in October 2023)
> Node.js: 20.5.0 (released in September 2023)
> Java: 21 (released in September 2023)
> C#: 12 (part of .NET 8, released in November 2023)
> Go: 1.21.0 (released in August 2023)
> Ruby: 3.2.2 (released in August 2023)
> PHP: 8.3.0 (released in August 2023)
> Rust: 1.72.0 (released in September 2023)
> TypeScript: 5.3.0 (released in September 2023)
> Angular: 16.2.0 (released in September 2023)
> React: 18.2.0 (released in June 2023)
> Vue.js: 3.3.0 (released in September 2023)
You can count on ChatGPT to know the exact versions and release dates of software just the same as you can rely on your friendly neighborhood dev to know it off the top of their head - not at all. A dev would likely have to look it up too. A language model would also need to look it up through function calling (or just including in your prompt).
This kind of scenario, where there are concrete answers in some datastore somewhere and the relevance of the content that the model was trained on varies chronologically (a blog post on the latest version of React circa 2015 could deliver a wrong answer), are the ones you want to engineer around. This is where you start using SDKs and binding tools to your LLM so you can ensure grounding context is available to generate correct answers.
This happens with other LLMs too. They don't know what they don't know. They ought to check what version you are using and at least issue a warning if the major number is different.
I don't care about not having IDE integration with this - this is a solved problem that I don't care about. It would be nice to have two-way synchronization though so I could run code locally, edit it with my local editor (not IDE) and sync it back up to make more changes in ChatGPT.
700 comments
[ 13.0 ms ] story [ 299 ms ] threadThis is closer to Cursor for writing than Claude Projects.
For that, $20/M per head to be usable? Yikes.
https://code.visualstudio.com/updates/v1_94
Or is it meant to be used on just a single file?
Do you want AI to do this for you? Do you trust that it will do a good job?
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.
This also demonstrates the type of things Google could do with Gemini integrated into Google Docs if they step up their game a bit.
Honestly I’m scratching my head on OpenAI’s desire to double down on building out their consumer B2C use cases rather than truly focussing on being the infrastructure/API provider for other services to plug into. If I had to make a prediction, I think OpenAI will end up being either an infrastructure provider OR a SaaS, but not both, in the long-term (5-10 yrs from now).
This is exactly what Google’s NotebookLM does. It’s (currently) free and it reads your Google Docs and does RAG on them.
https://notebooklm.google/
https://github.com/BenWheatley/Timeline-of-the-near-future
I've only used the "Deep Dive" generator a few times, and I'm already sensing the audio equivalent of "youtube face" in the style — not saying that's inherently bad, but this is definitely early days for this kind of tool, so consider Deep Dive as it is today to be a GPT-2 demo of things to come.
"Juggling dog" has only been expressed a single time previously in our corpus of humanity:
These jongleurs might sing amorous songs or pantomime licentious actions. But they might be also jugglers, bear trainers, acrobats, sleight-of-hand artists or outright mountebanks. Historian Joseph Anglade remarks that in the high Middle Ages:"We see the singer and strolling musician, who comes to the cabaret to perform; the mountebank-juggler, with his tricks of sleight-of-hand, who well represents the class of jongleurs for whom his name had become synonymous; and finally the acrobat, often accompanied by female dancers of easy morals, exhibiting to the gaping public the gaggle of animals he has dressed up — birds, monkeys, bears, savant dogs and counting cats — in a word, all the types found in fairs and circuses who come under the general name of jongleur.”(5) -- http://www.arthurchandler.com/symbolism-of-juggling
I suspect what I heard was a deliberate modification of this sexist quote from Samuel Johnson, which I only found by this thread piquing my curiosity: "Sir, a woman's preaching is like a dog's walking on his hind legs. It is not done well; but you are surprised to find it done at all." - https://www.goodreads.com/quotes/252983-sir-a-woman-s-preach...
Trying to find where I got my version from, takes me back to my own comments on Hacker News from 8 months ago, and I couldn't remember where I got it from then either:
> "your dog is juggling, filing taxes, and baking a cake, and rather than be impressed it can do any of those things, you're complaining it drops some balls, misses some figures, and the cake recipe leaves a lot to be desired". - https://news.ycombinator.com/item?id=39170057
My comment there predates this Mastodon thread, but the story in Mastodon may predate whoever told me the version I encountered: https://social.coop/@GuerillaOntologist/112598462146879765
Trying to find where I got my version from just brought me back to one of my own comments on Hacker News from 8 months ago:
> "your dog is juggling, filing taxes, and baking a cake, and rather than be impressed it can do any of those things, you're complaining it drops some balls, misses some figures, and the cake recipe leaves a lot to be desired". - https://news.ycombinator.com/item?id=39170057
I couldn't remember where I got it from then either.
I think it's because LLMs (and to some extent other modalities) tend to be "winner takes all." OpenAI doesn't have a long term moat, their data and architecture is not wildly better than xAI, Google, MS, Meta, etc.
If they don't secure their position as #1 Chatbot I think they will eventually become #2, then #3, etc.
But can they do it at all? It's not like they are like early Google vs other search engines.
How do you make money off a web browser, to justify the development costs? And what does that look like in an LLM?
Perhaps the ad/ad-blocker analogy would be: You can have the free genuinely open source LLM trained only on Wikipedia and out-of-copyright materials, or you can have one trained on current NYT articles and Elsevier publications that also subtly pushes you towards specific brand names or political parties that paid to sponsor the model.
Also consider SEO: every business wants to do that, nobody wants to use a search engine where the SEO teams won. We're already seeing people try to do SEO-type things to LLMs.
If (when) the advertisers "win" and some model is spitting out "Buy Acme TNT, for all your roadrunner-hunting needs! Special discount for coyotes!" on every other line, then I'd agree with you, it won't fly, people will switch. But it doesn't need to start quite so bold, the first steps on this path are already being attempted by marketers attempting to induce LLMs crawling their content to say more good things about their own stuff. I hope they fail, but I expect them to keep trying until they succeed.
Google and Facebook grew organically for a number of years before really opening the tap on ad intrusions in to the UX. Once they did, a tsunami of money crashed over both, quarterly.
The LLM companies will have this moment too.
(But your post makes me want to put a negative-prompt for Elsevier publications in to my Custom Instructions, just in case)
One of my machines also has a copy of Firefox on it. Not used that in ages, either. But Firefox is closer in quality to Chrome, than any of the locally-runnable LLMs I've tried are to the private/hosted LLMs like 4o.
Ironically, I had to google it, and agree with the comment.
https://chatgpt.com/share/66ff28e2-ea74-800b-a230-86d562f60f...
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
I've been using it for about a year.
Given how widely used Google Docs is, for serious work, disrupting people's workflows is not a good thing. Google has no problem being second, they aren't going to die in the next three months just because people on Twitter say so.
But if they believe they're going to reach AGI, it makes no sense to pigeonhole themselves to the interface of ChatGPT. Seems like a pretty sensible decision to maintain both.
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.
Maybe the consumer side will slide as businesses pick up the tab?
The whole financial breakdown is fascinating and I’m surprised to not see it circulating more.
https://www.wheresyoured.at/oai-business/
This analysis is just doing basic math based on reporting from the NYT and Post on OpenAI’s financials.
In my estimation, you're not qualified for this conversation.
(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.)
I agree with you on where OpenAI will/should sit in 5-10 years. However, I don't think them building the occasional tool like this is unwarranted, as it helps them show the direction companies could/should head with integration into other tools. Before Microsoft made hardware full time, they would occasionally produce something (or partner with brands) to show a new feature Windows supports as a way to tell the OEMs out there, "this is what we want you to do and the direction we'd like the PC to head." The UMPC[0] was one attempt at this which didn't take off. Intel also did something like this with the NUC[1]. I view what OpenAI is doing as a similar concept, but applied to software.
[0] https://en.wikipedia.org/wiki/Ultra-mobile_PC
[1] https://en.wikipedia.org/wiki/Next_Unit_of_Computing
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…
Few people outside of tech circles know what those other apps you mentioned are. I use Confluence at work, because it’s what my company uses. I also tried using it at home, but not for the same stuff I’d use Pages for. I use Obsidian at work to stay organized, but again, it doesn’t replace what I’d use Pages for, it’s more of a Notes competitor in my book. A lot of people don’t want their documents locked away in a Notion DB, and it’s not something I’d think to use if I’m looking to print something.
I went back and looked at the last WWDC video. Apple did mention the apps briefly, to say they have integrated Image Playgrounds, their AI image generation, into Pages, Keynote, and Numbers. With each major upgrade, the iWork apps usually get something. Office productivity isn’t exactly the center of innovation these days. The apps already do the things that 80% of users need.
Unless I'm missing something about Canvas, gh CoPilot Chat (which is basically ChatGPT?) integrates inline into IntelliJ. Start a chat from line numbers and it provides a diff before applying or refining.
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...
1.https://code.visualstudio.com/updates/v1_92#_github-copilot
2.https://code.visualstudio.com/updates/v1_94#_switch-language...
[0] https://github.blog/changelog/label/copilot/
[1] https://github.blog/changelog/2024-09-19-sign-up-for-openai-...
Google already does have this in Google Docs (and all their products)? You can ask it questions about the current doc, select a paragraph and ask click on "rewrite", things like that. Has helped me get over writer's block at least a couple of times. Similarly for making slides etc. (It requires the paid subscription if you want to use it from a personal account.)
https://support.google.com/docs/answer/13951448 shows some of it for Docs, and https://support.google.com/mail/answer/13447104 is the one for various Workspace products.
Or Microsoft!
> think OpenAI will end up being either an infrastructure provider OR a SaaS, but not both
Microsoft cut off OpenAI's ability to execute on the former by making Azure their exclusive cloud partner. Being an infrastructure provider with zero metal is doable, but it leaves obvious room for a competitor to optimise.
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.
What for? If someone has already a business and customers he's already far off the average YC startup.
It depends really on the business and what not.
I’m firmly in the camp that their rate limits are entirely reasonable.
Take a look at cursor.com
Professionals instead don't love to change the tools once they got used to it for small incremental gains.
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.
If you control the UI, you have none of those problems.
Copilot is only using GPT3.5 for most of the results though, seemingly. I'd be more excited if they would update the API they're using.
Shoot me in the face if my own writing is ever that bad.
ETA: just to be clear... I am not a great writer. Or a bad one. But this is a particular kind of bad. The kind we should all try to avoid.
They don't care. Their goal is to accelerate the production of garbage.
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.
For them, the ability to generate so much trash is the good part. They might not even be fully aware that it's trash, but their general goal is to output more trash because trash is profitable.
It's like all those "productivity systems". Not a single one will produce a noticeable increase in productivity magically that you can't get from just a $1 notebook, they just make you feel like you are being more productive. Same with RP bots or AI text editors. It makes you feel so much faster, and for a lot of people that's enough so they want in on a slice of the AI moneypit!
Occasionally I sink 1-2 hours into a tweaking something I thought was 90% correct but was in reality garbage. I had that happen a lot more with earlier models, but its becoming increasingly rare. Perhaps I'm recognizing the limitations of the tool, or the systems indeed are getting better.
This is all anecdotal, but I'm shipping and building faster than I was previously and its definitely not all trash.
If you accept that we live in a world where blind lead the blind, it's less surprising.
Other people in our lab (from China, Korea, etc.) also find this kind of thing useful for working / communicating quickly
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.
Recently, I stumbled upon a bug that initially seemed minor but quickly revealed itself to be a formidable adversary. It disrupted the seamless user experience I had meticulously crafted, and despite my best efforts, this issue has remained elusive. Each attempt to isolate and resolve it has only led me deeper into a labyrinth of complexity, leaving me frustrated yet undeterred.
Understanding that even the most seasoned developers can hit a wall, I’m reaching out for help. I’ve documented the symptoms, error messages, and my various attempts at resolution, and I’m eager to collaborate with anyone who might have insights or fresh perspectives. It’s in the spirit of community and shared knowledge that I hope to unravel this mystery and turn this challenge into an opportunity for growth.
Me: This is the most garbage code I've ever seen. It's bad and you should feel. It's not even wrong. I can't even fathom the conceptual misunderstandings that led to this. I'm going to have to rewrite the entire thing at this rate, honestly you should just try again from scratch.
With AI: I've had some time to review the code you submitted and I appreciate the effort and work that went into it. I think we might have to refine some parts so that it aligns more closely with our coding standards. There are certain areas that are in need of restructuring to make sure the logic is more consistent and the flow wouldn't lead to potential issues down the road.
I sympathize with the sibling comment about AI responses being overly-verbose but it's not that hard to get your model of choice to have a somewhat consistent voice. And I don't even see it as a crutch, this is just automated secretary / personal assistant for people not important enough to be worth a human. I think a lot of us on HN have had the experience of the stark contrast between comms from the CEO vs CEO as paraphrased by their assistant.
For lots of East Asian researchers it's really embarrassing for them to send an email riddled with typos, so they spend a LOT of time making their emails nice.
I like that tools like this can lift their burden
OK -- I can see this. But I think Grammarly would be better than this.
But its suggestion system, where it spots wordy patterns and suggests clearer alternatives, was available long before LLMs were the new hotness, and is considerably more nuanced (and educational).
Grammarly would take apart the nonsense in that screenshot and suggest something much less "dark and stormy night".
But there's not much more important, stylistically, to writing an business email or document than clarity. It's absolutely the most important thing. Especially in customer communications.
In the UK there is/used to be a yearly awards scheme for businesses that reject complexity in communucations for clarity:
https://www.plainenglish.co.uk/services/crystal-mark.html
But anyway, you don't have to act on all the suggestions, do you? It's completely different from the idea of getting an AI to write generic, college-application-letter-from-a-CS-geek prose from your notes.
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.
I see this all the time from AI boosters. Flashy presentation, and it seems like it worked! But if you actually stare at the result for a moment, it’s mediocre at best.
Part of the issue is that people who are experts at creating ML models aren’t experts at all the downstream tasks those models are asked to do. So if you ask it to “write a poem about pizza” as long as it generally fits the description it goes into the demo.
We saw this with Gemini’s hallucination bug in one of their demos, telling you to remove film from a camera (this would ruin the photos on the film). They obviously didn’t know anything about the subject beforehand.
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.
Unskilled and unaware of it. Or rather, unskilled and unaware of what a skilled output actually involves. So, unaware of the damage they do to their reputations by passing off the output of a GPT.
This is what I mean about the writing, ultimately. If you don't know why ChatGPT writing is sort of essentially banal and detracts from honesty and authenticity, you're the sort of person who shouldn't be using it.
(And if you do know why, you don't need to use it)
Seriously though, I'm tired of the "helpful" GitHub bots closing issues after X days of inactivity. Can't wait for one powered by AI to decide it's not interested in your issue.
On the other hand, I did have good luck w/ Anthropic's version of this to make a single page react app with super basic requirements. I couldn't imagine using it for anything more though.
- holding in your mind a "thing" (i.e. some code)
- talking about a "thing" (i.e. walking through the code)
The same applies to non-code tasks as well. The ability to segregate the actual "meat" from the discussion is an excellent interface improvement for chatbots.
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...
Slightly OT, but one thing I noticed further into the demo is how you were prompting.
Rather than saying “embed my projects in my portfolio site” you told it to “add an iframe with the src being the project url next to each project”. Similarly, instead of “make the projects look nice”, you told it to “use css transforms to …”
If I were a new developer starting today, it feels like I would hit a ceiling very quickly with tools like this. Basically it looks like a tool that can code for you if you are capable of writing the code yourself (given enough time). But questionably capable of writing code for you if you don’t know how to properly feed it leading information suggesting how to solve various problems/goals.
Yes, exactly. I use it the way I used to outsource tasks to junior developers. I describe what I need done and then I do code review.
I know roughly where I want to go and how to get there, like having a sink full of dirty dishes and visualizing an empty sink with all the dishes cleaned and put away, and I just instruct it to do the tedious bits.
But I try and watch how other people use it, and have a few other different styles that I employ sometimes as well.
> I use it the way I used to outsource tasks to junior developers.
Is this not concerning to you, in a broader sense? These interactions were incredibly formative for junior devs (they were for me years ago) - its how to grew new senior devs. If we automate away the opportunity to train new senior devs, what happens to the future?
https://i.imgur.com/R5PQQoi.png
You don't get the new experience until you give it a prompt though, which is kinda weird.
The easiest way to check if you have access is it will appear as an explicit choice in the "Model" selector.
If each was paid $300k (that's a minimum...) and they spent a year on this, it'd make it a $5M project...
Cant we just tell ChatGPT to make e.g. TensorFlow faster, better, cleaner? Why do people put in so much work anymore, if the AI is so damn good?
By many definitions of "programming", these AI tools are indeed programming. In the same way, many definitions of "reasoning, ..." may include the AI tools in them. However, there are alternate definitions (reasonable in their own way) in which it is clear these AI tools fall short.
So, I think you are proposing an argument of semantics but presenting it as if it is an argument on actual capabilities of these AI tools. In general, we all are in agreement on what these systems are capable of.
I would bet that the current paradigm in ChatGPT will never be able to replace a minimally competent human at real world programming - the kind of programs people actually pay for. Let’s see.
Depends on definition of "minimally competent". There WILL be over-engineered enterprise solutions that employ 10x more AI-enabled code monkeys than is strictly required. Think about it: we can create a "fully automated AI coder" and then hire another FTE (or team) to handhold and clean up after it. It will be done. It must be done.
I mean, there's strong incentive for you (it would be worth tens of millions, possibly billions if your iteration is superior enough).
How much time do you need?
I wonder what it would take to expand the languages it supports?
As with anything else that is helpful, there is a balancing act to be aware of. This is too much for my taste. Just like github copilot is too much.
It's too dumb like this. But chatgpt is insanely helpful in a context where I really need to learn something I am deep diving into or where I need an extra layer of direction.
I do not use the tool for coding up front. I use them for iterations on narrow subjects.
Using a spell-checker, I have gradually lost my ability to spell. Using these LLM tools, large parts of the population will lose the ability to think. Try to own them like farm animals.
The large number of tokens being processed by iterative models requires enormous energy. Look at the power draw of a Hopper or Blackwell GPU. The Cerebras wafer burns 23 KW.
One avenue to profit is to invest in nuclear power by owning uranium. This is risky and I do not recommend it to others. See discussion here: https://news.ycombinator.com/item?id=41661768
https://www.nature.com/articles/d41586-024-03162-2
https://www.nrc.gov/reading-rm/doc-collections/fact-sheets/3...
Fortunes will be made selling electricity to people who develop serious cognitive dependence on LLMs.
There is no need for you to participate in the profits. I respect your life choices and I wish you well.
Please see the discussion here:
https://news.ycombinator.com/item?id=41661768
for serious warnings. This is not suitable for you.
Jesus christ, I hope you are never in a position of any significant power
(did you just take the lord's name in vain?)
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).
Remove all the educator stuff and karpathy would still be one of the most accomplished of his generation in his field.
Idk just seems like a weird comment.
*(it's also possible he is)
We taught the model to open a canvas for prompts like “Write a blog post about the history of coffee beans”.
If you're not heavily editing this post to say something genuinely new, then congratulations you've added even more drivel to the Great Pacific Garbage Patch of the internet.
I am sorry, dears, but this is not how it works. To be good at any of this, you should be able to do it yourself without any "prompt engineering", and the only path is through work, time, trial/error, and tons of frustration.
But to be less negative, this (or NotebookLM) could be useful to re-arrange and enrich one's own notes.
Sadly the amount of LLM slop on the internet is already out of control, and I'm afraid there's no going back.
It can't display markdown and formatted code side-by-side which is kind of a surprise.
I haven't tried doing anything super complex with it yet. Just having it generate some poems, but it's smart enough to be able to use natural language to edit the middle of a paragraph of text without rewriting the whole thing, didn't notice any issues with me saying "undo" and having data change in surprising ways, etc. So far so good!
I'm not very skilled at creating good "test" scenarios for this, but I found this to be fun/interesting: https://i.imgur.com/TMhNEcf.png
I had it write some Python code to output a random poem. I then had it write some code to find/replace a word in the poem (sky -> goodbye). I then manually edited each of the input poems to include the word "sky".
I then told it to execute the python code (which causes it to run "Analyzing...") and to show the output on the screen. In doing so, I see output which includes the word replacement of sky->goodbye.
My naive interpretation of this is that I could use this as a makeshift Python IDE at this point?
A similar project is sandpack, but that relies on nodebox which is also closed source.
https://news.ycombinator.com/item?id=41733485
Has anyone had much experience with it, that can share their findings? I'm happy with Claude Sonnet and can't try every new AI code tool at the rate they are coming out. I'd love to hear informed opinions.
EDIT: Only seems to work in Chrome?
OpenAI already A/Bs test the responses it generates. Imagine if they own the text editor or spreadsheet you work on too. It’ll incorporate all of your edits to be self-correcting.
When I have used it I asked it to help me with VueJS but whilst it knew about version 3 (which is substantially different to V2) it didn't know the very latest version.
Isn't this going to be problematic if it's trying to help code but has outdated information? Won't this result in more code being written/tied to older languages/framework versions?
In my experience, because GPT outputs code tied to particular versions of software, I use those as it's far quicker to copy/paste/edit than reading the latest docs, where GPT is absolutely useless as it simply doesn't know about the latest code and isn't able to produce working snippets.
So the real deal would be if GPT knows about the latest software instantly the moment it's published (ingesting docs data, any stackoverflow Q/A's etc).
Here is what GPT 4o-mini told me
P: can you list me the latest verisons of ALL major languages such as python, node etc. and the latest versions of typescript, angular, vuejs etc
R:
This kind of scenario, where there are concrete answers in some datastore somewhere and the relevance of the content that the model was trained on varies chronologically (a blog post on the latest version of React circa 2015 could deliver a wrong answer), are the ones you want to engineer around. This is where you start using SDKs and binding tools to your LLM so you can ensure grounding context is available to generate correct answers.