[flagged]
A con like that wouldn't last very long.
This is for people who rely enough on ChatGPT Pro features that it becomes worth it. Whether they pay for it because they're freelance, or their employer does.
Just because an LLM doesn't boost your productivity, doesn't mean it doesn't for people in other lines of work. Whether LLM's help you at your work is extremely domain-dependent.
I really doubt that, actually. The only thing that LLMs are truly good for is to create plausible-sounding text. Everything else, like generating facts, is outside of its main use case and known to frequently fail.
Everything that is “word processing,” and that’s a lot.
This world is so messed up.
Or it already exists in some howto documentation, but nobody wanted to skim the documentation.
Versus the old way of asking them to write the contract, where they'll blatantly re-use some boilerplate (sometimes the name of a previous client's company will still be in there) and then take 2 weeks to get back to you with Draft #1, charging 10x as much.
I anlways ask our lawyer whether or not they have a boilerplate when I need a contract written up. They usually do.
I've even found that when lawyers send a document for one of my companies, and I give them a list of things to fix, including e.g. typos, the same typos will be in there if we need a similar document a year later for another company (because, well, nobody updated the boilerplate)
Do you ask about the boilerplate before or after you ask for a quote?
I could definitely see a large law firm (Orrick, Venable, Cooley, Fenwick) doing what you describe. I’ve worked with 2 firms just listed, and their billing practices were ridiculous.
I’ve had a lot more success (quality and price) working with boutique law firms, where your point of contact is always a partner instead of your account permanently being pawned off to an associate.
Email is in profile if you want an intro to the law firm I use. Great boutique firm based in Bay Area and extremely good price/quality/value.
EDIT: Added links.
https://www.cio.com/article/3540579/devs-gaining-little-if-a...
https://web.archive.org/web/20241205204237/https://llmreport...
(Archive link because the llmreporter site seems to have an expired TLS certificate at the moment.)
No improvement to PR throughput or merge time, 41% more bugs, worse work-life balance...
You missed this part. Being able to quickly fix things without deep thought while in flow saves you from the slowdowns of context switching.
Secondly, the "No improvement to PR throughput or merge time, 41% more bugs, worse work-life balance" result you quote came, per article, from a "study from Uplevel", which seems to[0] have been testing for change "among developers utilizing Copilot". That may or may not be surprising, but again it's hardly relevant to discussion about SOTA LLMs - it's like evaluating performance of an excavator by giving 1:10 toy excavators models to children and observing whether they dig holes in the sandbox faster than their shovel-equipped friends.
Best LLMs are too slow and/or expensive to use in Copilot fashion just yet. I'm not sure if it's even a good idea - Copilot-like use breaks flow. Instead, the biggest wins coming from LLMs are from discussing problems, generating blocks of code, refactoring, unstructured to structured data conversion, identifying issues from build or debugger output, etc. All of those uses require qualitatively more "intelligence" than Copilot-style, and LLMs like GPT-4o and Claude 3.5 Sonnet deliver (hell, anything past GPT 3.5 delivered).
Thirdly, I have some doubts about the very metrics used. I'll refrain from assuming the study is plain wrong here until I read it (see [0]), but anecdotally, I can tell you that at my last workplace, you likely wouldn't be able to tell whether or not using LLMs the right way (much less Copilot) helped by looking solely at those metrics - almost all PRs were approved by reviewers with minor or tangential commentary (thanks to culture of testing locally first, and not writing shit code in the first place), but then would spend days waiting to be merged due to shit CI system (overloaded to the point of breakage - apparently all the "developer time is more expensive than hardware" talk ends when it comes to adding compute to CI bots).
--
[0] - Per the article you linked; I'm yet to find and read the actual study itself.
That's not a problem. OpenAI need to get some cash from its product because the competition is intense from free models. Moreover, since they supposedly used most of the web content and pirated whatever else they could, improvements in training will likely be only incremental.
All the while, after the wow effect passed, more people start to realize the flaw in generative AI. So current hype, like all hype, as a limited shelf life and companies need to cash out now because it could be never.
They're bleeding money and are desperately looking for a business model to survive. It's not going very well. Zitron[1] (among others) has outlined this.
> OpenAI's monthly revenue hit $300 million in August, and the company expects to make $3.7 billion in revenue this year (the company will, as mentioned, lose $5 billion anyway), yet the company says that it expects to make $11.6 billion in 2025 and $100 billion by 2029, a statement so egregious that I am surprised it's not some kind of financial crime to say it out loud. […] At present, OpenAI makes $225 million a month — $2.7 billion a year — by selling premium subscriptions to ChatGPT. To hit a revenue target of $11.6 billion in 2025, OpenAI would need to increase revenue from ChatGPT customers by 310%.[1]
Surprise surprise, they just raised the price.
They have also added a new, even higher performance model which can leverage test time compute to scale performance if you want to pay for that GPU time. This is no different than AWS offering some larger ec2 instance tier with more resources and a higher price tag than existing tiers.
Models are becoming a commodity. It's game theory. Every second place company (eg. Meta) or nation (eg. China) is open sourcing its models to destroy value that might accrete to the competition. China alone has contributed a ton of SOTA and novel foundation models (eg. Hunyuan).
https://www.nytimes.com/2024/09/27/technology/openai-chatgpt...
Roughly 10 million ChatGPT users pay the company a $20 monthly fee, according to the documents. OpenAI expects to raise that price by $2 by the end of the year, and will aggressively raise it to $44 over the next five years, the documents said.
We'll have to see if the first bump to $22 this year ends up happening.
I can't read the article. Any mention of the API pricing?
I'm hard pressed to identify any users to whom LLMs are providing enough value to justify $20/month, but not $44.
On the other hand, I can see a lot of people to whom it's not providing any value being unable to afford a higher price.
Guess we'll see which category most OpenAI users are in.
So the question is it worth $200/month and to how many people, not is it over hyped, or if it has flaws. And does that support the level of investment being placed into these tools.
Models are about to become a commodity across the spectrum: LLMs [1], image generators [2], video generators [3], world model generators [4].
The thing that matters is product.
[1] Llama, QwQ, Mistral, ...
[2] Nobody talks about Dall-E anymore. It's Flux, Stable Diffusion, etc.
[3] HunYuan beats Sora, RunwayML, Kling, and Hailuo, and it's open source and compatible with ComfyUI workflows. Other companies are trying to open source their models with no sign of a business model: LTX, Genmo, Rhymes, et al.
[4] The research on world models is expansive and there are lots of open source models and weights in the space.
Does Apple charge a premium? Of course. Do Apple products also tend to have better construction, greater reliability, consistent repair support, and hold their resale value better? Yes.
The idea that people are buying Apple because of the Apple premium simply doesn't hold up to any scrutiny. It's demonstrably not a Verblen good.
The lack of repairability is easily Apple's worst quality. They do everything in their power to prevent you from repairing devices by yourself or via 3rd party shops. When you take it to them to repair, they often will charge you more than the cost of a new device.
People buy apple devices for a variety of reasons; some people believe in a false heuristic that Apple devices are good for software engineering. Others are simply teenagers who don't want to be the poor kid in school with an Android. Conspicuous consumption is a large part of Apple's appeal.
Maybe not as true in the US, but reading about the green bubble debacle, it's also a lot about status.
Now that is a trope when you're talking about Apple. They may use more premium materials that and have a degree of improved construction leveraging those materials - but at the end of the day there are countless numbers of failure prone designs that Apple continued to ship for years even after knowing they existed.
I guess I don't follow the fact that the "Apple Premium" (whether real or otherwise) isn't a factor in a buyer decision. Are you saying Apple is a great lock-in system and that's why people continue to buy from them?
There’s something to be said for buying something and knowing it will interoperate with all your other stuff perfectly.
It's very hard to explain to people who haven't dug into macOS that it's a great system for power users, for example, especially because it's not very customizable in terms of aesthetics, and there are always things you can point to about its out-of-the-box experience that seem "worse" than competitors (e.g., window management). And there's no one thing I can really point to and say "that, that's why I stay here"; it's more a collection of little things. The service menu. The customizable global keyboard shortcuts. Automator, AppleScript (in spite of itself), now the Shortcuts app.
And, sure, they tend to push their hardware in some ways, not always wisely. Nobody asked for the world's thinnest, most fragile keyboards, nor did we want them to spend five or six years fiddling with it and going "We think we have it now!" (Narrator: they did not.) But I really do like how solid my M1 MacBook Air feels. I really appreciate having a 2880x1800 resolution display with the P3 color gamut. It's a good machine. Even if I could run macOS well on other hardware, I'd still probably prefer running it on this hardware.
Anyway, this is very off topic. That ChatGPT Pro is pretty damn expensive, isn't it? This little conversation branch started as a comparison between it and the "Apple tax", but even as someone who mildly grudgingly pays the Apple tax every few years, the ChatGPT Pro tax is right off the table.
But name one other company whose hardware truly matches Apple’s standards for precision and attention to detail.
Example: the 4o or Claude are great for coding, summarizing and rewriting emails. So which domains require a slightly better model?
I suppose if the error rate in code or summary goes down even 10%, it might be worth $180/month.
I made a very exhaustive question to ChatGPT o1-preview, including all information I though is relevant. Something like good forums question. Well, 10 seconds later it spew me a working solution. I was ashamed, because I have 20 years of experience under my belt and this model solved non-trivial task much better than me.
I was ashamed but at the same time that's a superpower. And I'm ready to pay $200 to get solid answers that I just can't get in a reasonable timeframe.
(Now I think if of the idiom, when did we switch to 9-6? I've never had a 9-5).
The NFT market lasted for many years and was enormous.
Never underestimate the power of hype.
Bernie Madoff ran his investment fund as a Ponzi scheme for over a decade (perhaps several decades)
> we use a stricter evaluation setting: a model is only considered to solve a question if it gets the answer right in four out of four attempts ("4/4 reliability"), not just one
This surely makes the other models post smaller numbers. I'd be curious how it stacks up if doing eg 1/1 attempt or 1/4 attempts.
I suspect this is a key driver behind having a higher priced, individual user offering. It gives pricing latitude for enterprise volume licenses.
As someone who has both repeatedly written that I value the better LLMs as if they were a paid intern (so €$£1000/month at least), and yet who gets so much from the free tier* that I won't bother paying for a subscription:
I've seen quite a few cases where expensive non-functional things that experts demonstrate don't work, keep making money.
My mum was very fond of homeopathic pills and Bach flower tinctures, for example.
* 3.5 was competent enough to write a WebUI for the API so I've got the fancy stuff anyway as PAYG when I want it.
Just zero curiosity, only skepticism.
Let's say I run a company call AndSoft.
AndSoft has about 2000 people on staff, maybe 1000 programers.
This solution will cost 200k per year. Or 2.4 million per year.
Llama3 is effectively free with some liberation. Is ChatGPT pro 2.4 million a year better than Llama3. Of course Open AI will offer volume discounts.
I imagine if I was making north of 500k a year I'd subscribe as a curiosity... At least for a few months.
If your time is worth 250$ a hour, and this saves you an hour per month it's well worth it.
I'm talking of the generally available one, haven't had the chance to try this new version.
For tasks where bullshitting or regurgitating common idioms is key, it works rather well and indeed takes you 80% or even close to 100% of the way there. For tasks that require technical precision and genuine originality, it’s hopeless.
So far, given my range of projects, I have seen it struggle with lower level mobile stuff and hardware (ESP32 + BLE + HID).
For things like web (front/back), DB, video games (web and Unity), it does work pretty well (at least 80% there on average).
And I'm talking of the free version, not this $200/mo one.
I’m not saying that LLMs can’t be useful, but I do think it’s a darn shame that we’ve given up on creating tools that deterministically perform a task. We know we make mistakes and take a long time to do things. And so we developed tools to decrease our fallibility to zero, or to allow us to achieve the same output faster. But that technology needs to be reliable; and pushing the envelope of that reliability has been a cornerstone of human innovation since time immemorial. Except here, with the “AI” craze, where we have abandoned that pursuit. As the saying goes, “to err is human”; the 21st-century update will seemingly be, “and it’s okay if technology errs too”. If any other foundational technology had this issue, it would be sitting unused on a shelf.
What if your compiler only generated the right code 99% of the time? Or, if your car only started 9 times out of 10? All of these tools can be useful, but when we are so accepting of a lack of reliability, more things go wrong, and potentially at larger and larger scales and magnitudes. When (if some folks are to believed) AI is writing safety-critical code for an early-warning system, or deciding when to use bombs, or designing and validating drugs, what failure rate is tolerable?
This does not follow. By your own assumptions, getting you 80% of the way there in 10% of the time would save you 18% of the overall time, if the first 80% typically takes 20% of the time. 18% time reduction in a given task is still an incredibly massive optimization that's easily worth $200/month for a professional.
160 hours a month * $100/hr programmer * 9% = $1400 savings, easily enough to justify $200/month.
Even if 1/10th of the time it fails, that is still ~8% or $1200 savings.
People around here feel seriously threatened by ML models. It makes no sense, but then, neither does defending the Luddites, and people around here do that, too.
Sometimes it does what you want it to do, but still creates a bug.
Asked the AI to write some code to get a list of all objects in an S3 bucket. It wrote some code that worked, but it did not address the fact that S3 delivers objects in pages of max 1000 items, so if the bucket contained less than 1000 objects (typical when first starting a project), things worked, but if the bucket contained more than 1000 objects (easy to do on S3 in a short amount of time), then that would be a subtle but important bug.
Someone not already intimately familiar with the inner workings of S3 APIs would not have caught this. It's anyone's guess if it would be caught in a code review, if a code review is even done.
I don't ask the AI to do anything complicated at all, the most I trust it with is writing console.log statements, which it is pretty good at predicting, but still not perfect.
I use LLMs maybe a few times a month but I don’t really follow this argument against them.
It would be pretty easy for most code reviewers to miss this type of bug in a code review, because they aren't always looking for that kind of bug, they aren't always looking at the AWS documentation while reviewing the code.
Yes, people could also make the same error, but at least they have a chance at understanding the documentation and limits where the LLM has no such ability to reason and understand consequences.
If it's easier mentally, just put that second sentence in from of every chatgpt answer.
Yeah the Junior dev gets better, but then you hire another one that makes the same mistakes, so in reality, on an absolute basis, the junior dev never gets any better.
New echnology allows those signs to be counterfeited quickly and cheaply, and it tricks our subconscious despite our best efforts to be hyper-vigilant. (Our brains don't want to do that, it's expensive.)
Perhaps a stopgap might be to make the LLM say everything in a hostile villainous way...
If an intern set their Slack status to "There's no guarantee that what I say will be accurate, engage with me at your own risk." That wouldn't excuse their attempts to answer every question as if they wrote the book on the subject.
There seems to be two camps: People who want nothing to do with such flawed interns - and people who are trying to figure out how to amplify and utilize the positive aspects of such flawed, yet powerful interns. I'm choosing to be in the latter camp.
The only point I wanted to make was that an LLM's ability and propensity to generate plausible falsehoods should, in my opinion, elicit a much deeper sense of distrust than one feels for an intern, enough so that comparing the two feels a little dangerous. I don't trust an intern to be right about everything, but I trust them to be self aware, and I don't feel like I have to take a magnifying glass to every tidbit of information they provide.
It just seems to me that you really need to know the answer before you ask it to be over 90% confident in the answer. And the more convincing sounding these things get the more difficult it is to know whether you have a plausible but wrong answer (aka "hallucination") vs a correct one.
If you have a need for a lot of difficult to come up with but easy to verify answers it could be worth it. But the difficult to come up with answers (eg novel research) are also where LLMs do the worst.
Which is a very valuable lesson, worth more than $200
Now I'll forever be using a second rate model because I'm not rich enough.
If I'm stuck using a second rate model I may go find someone else's model to use.
The problem is whether it can saves you $180/mo more than Claude does.
I love this back-to-back pair of statements. It is like “You can never win three card monte. I pay a monthly subscription fee to play it.”
I think at least LLMs are more receptive to the idea that they may be wrong, and based on that, we can have N diverse LLMs and they may argue more peacefully and build a reliable consensus than N "intelligent" people.
Even if it was a consensus opinion among all HN users, which hardly seems to be the case, it would have little impact on the other billion plus potential customers…
A LLM will just pretend to care about the error and happily repeats the error over and over.
bit of a stretch.
Try asking an LLM complex questions in your area of expertise. Interview it as if you needed to be confident that it could do your job. You'll quickly find out that it can't do your job, and isn't actually capable of reasoning.
People also pull this figure out of their ass, over or undertrust themselves, and lie. I'm not sure self-reported confidence is that interesting compared to "showing your work".
I think there's a huge difference because individuals can be reasoned with, convinced they're wrong, and have the ability to verify they're wrong and change their position. If I can convince one person they're wrong about something, they convince others. It has an exponential effect and it's a good way of eliminating common errors.
I don't understand how LLMs will do that. If everyone stops learning and starts relying on LLMs to tell them how to do everything, who will discover the mistakes?
Here's a specific example. I'll pick on LinuxServer since they're big [1], but almost every 'docker-compose.yml' stack you see online will have a database service defined like this:
services:
app:
# ...
environment:
- 'DB_HOST=mysql:3306'
# ...
mariadb:
image: linuxserver/mariadb
container_name: mariadb
environment:
- PUID=1000
- PGID=1000
- MYSQL_ROOT_PASSWORD=ROOT_ACCESS_PASSWORD
- TZ=Europe/London
volumes:
- /home/user/appdata/mariadb:/config
ports:
- 3306:3306
restart: unless-stopped
Another Docker related example would be a Dockerfile using 'apt[-get]' without the '--error-on=any' switch. Pay attention to Docker build files and you'll realize almost no one uses that switch. Failing to do so allows silent failures of the 'update' command and it's possible to build containers with stale package versions if you have a transient error that affects the 'update' command, but succeeds on a subsequent 'install' command.
There are tons of misunderstandings like that which end up being so common that no one realizes they're doing things wrong. For people, I can do something as simple as posting on HN and others can see my suggestion, verify it's correct, and repeat the solution. Eventually, the misconception is corrected and those paying attention know to ignore the mistakes in all of the old internet posts that will never be updated.
How do you convince ChatGPT the above is correct and that it's a million posts on the internet that are wrong?
1. https://docs.linuxserver.io/general/docker-compose/#multiple...
## Restrict Host Ports for Security
If app and mariadb are only communicating internally, you can remove 3306:3306 to avoid exposing the port to the host machine:
```yaml ports: - 3306:3306 # Remove this unless external access is required. ```
So, apparently, ChatGPT doesn't need any more convincing.
I don't understand how it gets there though. How does it "know" that's the right thing to suggest when the majority of the online documentation all gets it wrong?
I know how I do it. I read the Docker docs, I see that I don't think publishing that port is needed, I spin up a test, and I verify my theory. AFAIK, ChatGPT isn't testing to verify assumptions like that, so I wonder how it determines correct from incorrect.
That said, my "issue" might be that I usually work alone and I don't have anyone to consult with. I can bother people on forums, but these days forums are pretty much dead and full of trolls, so it's not very useful. ChatGPT was that thing that allows me to progress in this environment. If you work in Google and can ask Rob Pike about something, probably you don't need ChatGPT as much.
Is it possible that they have subsidized the infrastructure for free and paid users and they realized that OpenAI requires a higher revenue to maintain the current demand?
Losing $5-10b per year also is insane. People are still looking for the added value, it's been 2 whole years now
I think the ship has sailed on whether GPT is useful or a con; I've lost track of people telling me it's their first search now rather than Google.
I'd encourage skeptics who haven't read this yet to check out Nicholas' post here:
It's the cost of a new, shiny, Apple laptop every year.
"Barely good enough to replace interns" is worth a lot to businesses already.
(On that note, a founder of a SAP competitor and a major IT corporation in Poland is fond of saying that "any specialist can be replaced by a finite number of interns". We'll soon get to see how true that is.)
Since when does SAP have competitors? ;-P
A friend of mine claims most research is nowadays done by undergraduates because all senior folks are too busy.
$200 seems pretty cheap for a 24/7 [remote] intern with these abilities. That kind of money doesn't even buy a month's worth of Big Macs to feed that intern with.
It just seems like a lot (or even absurd) for a subscription to a service on teh Interweb, akin to "$200 for access to a web site? lolwut?"
Or each user doing an o1 model prompt is probably like, really expensive and they need to charge for it until they can get cost down? Anybody have estimates on what a single request into o1 costs on their end? Like GPU, memory, all the "thought" tokens?
The truth is that there are people who value the marginal performance -- if you think it's insane, clearly it's not for you.
Those people want to purchase status. Unless they ship you a fancy bow tie and a wine tasting at a wood cabin with your chatgpt subscription this isn't gonna last long.
This isn't about marginal performance, it's an increasingly desperate attempt to justify their spending in a market that's increasingly commodified and open sourced. Gotta convince Microsoft somehow to keep the lights on if you blew tens of billions to be the first guy to make a service that 20 different companies are soon gonna sell for pennies.
Interesting to compare this $200 pricing with the recent launch of Amazon Nova, which has not-equivalent-but-impressive performance for 1/10th the cost per million tokens. (Or perhaps OpenAI "shipmas" will include a competing product in the next few days, hence Amazon released early?)
A fun question I tried a couple of times is asking it to give me a list with famous talks about a topic. Or a list of famous software engineers and the topics they work on.
A couple of names typically exist but many names and basically all talks are shamelessly made up.
This hallucination is usually a very good choice to name the option / API.
I wish it could just say "There is not a good approximation of this API existing - I would suggest reviewing the following docs/sources:....".
I certainly don’t see why mere prediction can’t validate reasoning. Sure, it can’t do it perfectly all the time, but neither can people.
Have you been introduced to their CEO yet? 5 minutes of Worldcoin research should assuage your curiosity.
Forgive me for not finding your argument persuasive.
As a customer, I don’t care about the people. I’m not interested in either argument by authority (if Altman says it’s good it must be good) or ad hominem (that Altman guy is a jerk, nothing he does can have value).
The actual product. Have you tried it? With an open mind?
To be honest, it doesn't matter what the price of producing AI is, though. $200/month is, and will be a stupid price to pay because OpenAI already invented a price point with a half billion users - free. When they charged $10/month, at least they weren't taking advantage of the mentally ill. This... this is a grift, and a textbook one at that.
You don’t sound like you’re very familiar with the chatgpt product. They have about 10m customers paying $20/month. I’m one of them, and I honestly get way more than $200/month value from it.
Perhaps I’m “mentally ill”, but I’d ask you to do some introspection and see if leaping to that characterization is really the best way to explain people who get value where you see none.
Such a silly conclusion to draw based on a gut feeling, and to see all comments piggyback on it like it's a given feels like I'm going crazy. How can you all be so certain?
What I doubt though is that it can reach a mass market even in business. A good large high resolution screen is something that I absolutely consider to deliver the value it costs. Most businesses don’t think their employees deserve a 2k screen which will last for 6-10 years and thus costs just a fraction of this offering.
Apparently the majority of businesses don’t believe in marginal gains
I'll pay $200 a month, no problem; right now o1-preview does the work for me of a ... somewhat distracted graduate student who needs checking, all for under $1 / day. It's slow for an LLM, but SUPER FAST for a grad student. If I can get a more rarely distracted graduate student that's better at coding for $7/day, well, that's worth a try. I can always cancel.
[1] https://www.investopedia.com/terms/p/price_discrimination.as...
It makes me wonder why they don't want to offer a usage based pricing model.
Is it because people really believe it makes a much worse product offering?
Why not offer some of the same capability as pay-per-use?
Yeah, not really fixed: https://imgur.com/a/counting-letters-with-chatgpt-7cQAbu0
Just to put that into perspective.
I also really don't find comparisons like this to be that useful. Any subscription can be converted into an exchange rate of coffee, or meals. So what?
So, AI market is capped by Starbucks revenue/valuation.
In Italy, an espresso is ca. 1€.
In other words, do you have proof that this medium of information output is doomed to forever be useless in producing information that adds value to the world?
These are of course rhetorical questions that you nor anyone else can answer today, but you seem to have a weird sort of absolute position on this matter, as if a lot depended on your sentiment being correct.