Once every 36 seconds? Even for simple transcription, that seems insane. This reads like the hustle culture memes...
2:00am wake up
2:05am cold shower
2:15am breakfast: almonds, breast milk bought off facebook, 50mg adderal
2:30am begin transcribing thoughts into the machine
I read the whole article, and It could be bragging
or a cry for help.
In any case its clear that the world of performance metrics has achived some kind of unimaginable gravity and angular momentum, and has started to orbit itself
> Within the last 24 months it’s clear that AI has become an essential coworker
I'm curious why the personification of these tools? Like, with the same logic we could call our dishwashers and washing machines employees and co-workers.
My guess is it is some form of hype-speak to raise the level of perceived importance of the technology for financial gain.
Because you don't start your dishwasher by saying to it "alright, I need you to wash my dishes. they are going to probably need some extra soaking because I burned dinner yesterday." Instead, you push the pots and pans button and off to work the robot goes.
> In fact, I find myself growing reliant on the AI to the extent that I no longer remember some of the R syntax, A sign of working at a higher level of abstraction : one of the promises of AI.
> A sign of working at a higher level of abstraction.
Working at a higher level of abstraction while loosing knowledge at the lower level also means by some degree that one is going to be reliant on the abstraction without any understanding what is abstracted away.
Hustle culture nonsense aside, I find using whisper (AI) transcription instead of typing to be super efficient.
My workflow involves setting up a whisper server, downloading the Whispering(1) app on my computer, and binding it to a shortcut on my keyboard and mouse. Whenever I want to write something down, I just hit the shortcut, dictate and it transcribes instantly. With a Nvidia GPU (1070 in my case), transcription is nearly instantaneous. Although I haven’t set it up on my MacBook yet, I suspect it will be just as fast with Apple Silicon
Learning to leverage AI as it exists today takes effort, but is usually worth it.
However, ignoring the effort undercuts the distance these products have to go. Speech is great because there is just a single interface to integrate with. Obviously I'm biased to my employer's speech product, but I'm sure there are many.
Biggest thing for me was when I saw that the lem editor[0] posted on hacker news[1] was a small editor, which has 3 top level features: common lisp API, LSP support, and copilot support.
I've installed gptel[2] in emacs, and hacking up a few tools that really make it shine. Up next is figuring out voice + AI + emacs :)
I'm probably at 20 times a day, at least for random trivia. What would be great (and maybe it already exists) is a daily email report, formatted for Anki, that helps me remember the things I've asked ChatGPT.
> Publishing data-driven blog post analysis is a key part from formatting the data to analyzing it using R and then publishing charts. All of this is now predominantly handled with prompts to an AI.
> I can generate several hundred lines of code in 5-10 minutes. With the newer models, I expect this to collapse to 1-2 minutes.
It seems main usage is to make plots with R... Which is funny as the bar chart on the page can be done in Excel/Numbers/Sheets by entering a few numbers and headers.
You have no idea how many times you use it.
Maps? Recommendation systems (the algo)? Other optimizations?
BTW, to the best of my knowledge, dictation still has a high error rate.
There is only so much AI can do because it currently lacks in certain domain knowledge.
The worst one I ever had to fix was ESPN captions of commentary for some indoor motorcross thing with dirt bikes going around a track. First, the motorbike noise, but secondly, the commentators were using the (well known to fans) nicknames for all the riders, which the AI had no idea how to transcribe, no idea who they were, and were almost impossible for me to even Google.
Whisper has a feature where you can provide it with a text prompt before it starts transcribing to clue it up to additional vocabulary that it may need to know.
Like it or not "AI" has become synonymous with generative AI, when most people talk about how much they use AI they don't mean the YouTube algorithm or the neural network that handles autocorrect on their phone.
Whisper (OpenAI 2022) and other recent dictation models are rather good. Whisper automatically punctuates sentences and usually gets just a few words wrong in my experience.
Artificial Analysis says Whisper 3 has a 10% word error rate, although it typically does better than that in my experience. When I use Whisper 2 in the ChatGPT app it usually gets 2 or 3 words wrong per paragraph of prompt.
But that is really bad, no? I am not a user so please enlighten me on the ergonomics, but I understand that to mean I have to check and correct everything.
Probably because AI is a misnomer. Machine learning is a better name, but even than it is pretty lacking.
What is generally called AI in common speech is pretty much a class of non-linear statistical models which require some training to generate weights which are than used to fit the model. Most people that know anything about statistics knows this, so it is fine actually. Misnomers exists in all industries and all science, and we just deal with them.
I've been thinking that a lot lately - LLMs are really nice search engines, which is great, that's a product everyone needs.
Before Google, information is hard to find. After Google, websites are incentivized to bury information behind ads and bullshit. After LLMs, the information is juiced out like lemonade.
I wonder what the next step in the incentive landscape is?
I notice you keep popping up every time there's a chance to spread FUD and hate regarding AI tech, its developers and its users, with comments that are getting more and more intolerant and unhinged. I consider you a toxic, potentially dangerous extremist, someone we all should fight against.
That's laughable. When I say a few harsh words against AI, you label me an extremist (a word used by the power structure to be categorically dismissive), and when a tech company uses megajoules to train an AI program that pollutes the air we breath, you call it progress. You might consider me toxic, but I consider you delusional.
(I have not done the math:) If you factor in 20-30 years of education that each typical white collar human requires as an entry point into their field, and the energy that goes into that, and the energy that goes into the continued upkeep of that human, the billions of seconds you can shave off menial work by pushing it all to an increasingly efficient AI are probably easily worth it, from a purely ecological perspective.
>the billions of seconds you can shave off menial work by pushing it all to an increasingly efficient AI are probably easily worth it, from a purely ecological perspective.
For a fixed amount of work, yes, but capitalism is based on growth, we will just create more work to do.
He mentions that both the dictation model and the editing model run on his laptop, so I'm not sure this is actually using all that much resources (beyond what it took to train the models originally)
