Editor's note: At present, large models can help humans deal with almost all aspects of things, such as dialogue, writing articles, writing code, and so on. Under the trend of "hurricane" of large models, small partners who want to engage in the field of AI may hesitate: Is it too late to enter the field of AI now?
The author of this article combines his own transformation experience and research and judgment on the AI market, and expounds that it is never too late to enter the field of artificial intelligence. There's no time like the present .
This article also details how to quickly master the necessary AI knowledge and how to find your own career position in the AI era in the era of "AI permeates everything".
The following is the translation, Enjoy!
Author | swyx
Compile | Yue Yang
A developer friend of mine recently told me, "If I were 20 years old now, I'd drop everything in AI." But he's spent more than a decade learning expertise, networking, and good His reputation has reached the highest level in his current field. Therefore, he has to stay in the original field for now. Another older college friend was an executive at a public tech startup. He is very comfortable with his current job, has a nearly perfect resume, and his previous professional experience is also an enviable position. However, he's switching gears now because, as he told me, "life is short" and he doesn't want to end up thinking "what if then...".
In recent days, I've had conversations like the one above with friends in both technical and non-technical jobs. While I'd love for this article to focus on specific technology developments and share the latest technology that is rapidly evolving, I thought it was worth dedicating an issue to the topic of career transitions , since that's something I happen to be particularly qualified to discuss.
01 The turning point in your thirties
I still remember how scary it was when I first made a career transition at the age of 30. I had been in the financial industry for 6-7 years, I wanted to be in the financial industry since I was 16 years old, and I traveled all over the world. CEOs ask questions and help manage $1 billion in assets at one of the world's top hedge funds. It looks like I'm good, but I know that my heart is not satisfied, and this is not my ultimate goal. Making some endowment and pension numbers even bigger pales in comparison to creating something out of nothing. I decided to move from finance to software engineering (and developer relations) . Everyone knows what happened next [1].
Six or seven years later, I switched my career again. I think the shift from software engineering (SWE) to artificial intelligence (AI) is almost as big as the shift from finance to software engineering , they are only superficially similar (both computer fields), but require a lot of new knowledge and practice Experience can make work efficient. My transition strategy was the same as last time: I studied every night and weekend for the first six months as much as possible to ensure that I had a strong interest in the field I wanted to enter (Note 1, explanation at the end of the article, the same below), and I was able to Make meaningful progress, and then start to say goodbye/break the boat/go all out with the past and tell everyone that I entered this field[2](Note 2).
But that only applies to my own case. Everyone's situation may be different. I believe you can find out how to make a successful career change if you want to. This article is aimed at those who want to gain enough confidence to make a decision.
I think there is a lot of hidden ageism and sunk cost fallacy [3] (sunk cost fallacy) in the choice of technology careers. So here's my quick list of some reasons why you won't be barred from changing careers just because of age .
02 Even if you are very old, you should still enter the field of AI for the following reasons
2.1 The Great Potential/Development Speed of Artificial Intelligence
- Jeff Bezos quit his finance job at age 30 to start Amazon.
- He did this because in 1994 internet usage was growing 2300% a year.
- The promotion of general purpose technologies [4] (note 3) will take decades.
- Imagine if you could be a "back wave" in technology in 2000 or 2010, only to conclude that it was "too late" and not enter the Internet industry.
- Since January, the usage of ChatGPT has increased by 1000% [5] (Note 4).
2.2 The time needed to get started with AI is shorter than we thought
- If you do not enter the field of machine learning through the path of obtaining a Ph. The field of machine learning does interesting things.
- But at present, our learning path to master generative AI is getting easier and easier. (Note 5)
- Jeremy Howard's fast.ai course [7] from 2016 claims to allow students to enter the field of artificial intelligence in seven weeks. By 2022, he has led students to re-implement Stable Diffusion[8] through ten 90-minute courses. Suhail Doshi took the course in June 2022 and launched Playground.ai[9] in November.
- This was driven in part by the Transformer architecture introduced in 2017, which has since entered nearly every AI domain [10] and provides a powerful and flexible baseline from which prior architectural knowledge becomes available. dispensable . So there are no decades of research to study, just the last five years .
https://www.stateof.ai/2018
- Some readers asked about the math involved in AI. It's debatable whether AI "just uses matrix multiplication" [11], you can learn matrix multiplication in college linear algebra and calculus courses if you want, but my answer is that you don't have to, the current ones AI development frameworks (like Pytorch) can help you with any backpropagation and matrix manipulations.
- Of course, taking shortcuts doesn't turn you into a Ph.D. who can push technology forward. But looking at the careers of the top AI researchers, you can also get an idea of how long it took to reach the top level. Yi Tay has contributed to or led many of the latest LLM efforts at Google, but you might be surprised to know that he's only been doing his Ph.D. for about 3.3 years. Ashish Vaswani was only 3 years away from his doctoral degree when he published his Transformer thesis, and Alec Radford had just graduated 2 years after his undergraduate degree when he published his GPT and GPT-2 papers.
- Career trajectories like this will not happen in more mature fields such as physics, mathematics, and medicine, because their "FOOM (Fast Onset of Overwhelming Mastery)" years [12] have passed for centuries, and AI's "foom" is clearly happening.
- These words are meant to illustrate: this is still a very young field , and in 20 years, no one is likely to care that you feel like you are "late to the game".
2.3 In addition to becoming a professional researcher in the field of machine learning, there are many fields to choose from
- Prompt and large model capability research: Riley Goodside[13]'s career will change dramatically in 2022. By tweeting his GPT-3 usage skills, he went from Grindr's data scientist to the world's first senior prompt engineer [14] who also discovered and popularized [15] the important LLM security problem of " prompt injection ". Since then, many people have realized that finding interesting use cases for GPT-3 and GPT-4 is popular on social media.
- In the field of software engineering: Recently, Whisper.cpp and LLaMA.cpp have stimulated many people's interest in running large models on the user terminal [16]. I listened to Georgi Gerganov's interview on the Changelog [17] and learned that in September 2022 he called himself a "non-AI believer" and just ported Whisper to C++ for fun. LLaMA.cpp is growing faster than Stable Diffusion [18], which is already one of the fastest growing open source projects ever [19]. Although no model training was performed, Georgi's software engineering expertise made these underlying models more accessible. Harrison Chase's Langchain [20] has attracted a lot of attention by building the first prompt engineering framework for all developers, fusing prompt and software improvements into pretrained LLM models. A range of LLM tools from Guardrails to Nat.dev are helping to bridge the gap between these models from academia to commercial applications. ChatGPT itself is largely a UX innovation delivered with the GPT 3.5 series of models, which is good news for front-end/UI developers.
- Productization of AI technology: Speaking of Stable Diffusion, Emad Mostaque was a hedge fund manager until 2019[21], it seems that in addition to conducting "literature review of autism and biomolecular pathway analysis of neurotransmitters[22]" for his son No prior AI experience outside of research. But after he joined the EleutherAI community in 2020, he realized that something like Stable Diffusion might exist, and found Patrick and Robin[23] of the CompVis group[24] at Heidelberg University, and provided about $600,000 for training and delivery. Second or most important AI product in 2022. Nobody wants to scrutinize who did what, but it makes sense that a former hedge fund manager would be rewarded massively by spotting opportunities and applying financial (and organizational) leverage to ideas whose time had come . Nat Friedman has publicly stated that the overcapacity created by years of research [25] has not been digested by enough startups , and it seems that entrepreneurs like Dave Rogenmoser who are willing to jump on this train early will convert Jasper's ARR ( accounting rate of return) grew from zero to $75 million [26], and would reap a disproportionate return.
Both incumbents and startups across verticals are embracing AI, suggesting that the future will be "AI permeates everything" , so understanding the underlying models may be a means to an end (leveraging them) rather than an end in itself (training models or thinking about security and awareness) . Think a little less about yourself and your potential future direction, not "change careers and study AI", but "learn how to use it" in an area you are already interested in or proficient in.
My last age-related appeal is general — challenging yourself is good for the brain. It is generally believed that neuroplasticity ceases after the age of 25, but this is controversial [27]. There is a broader consensus that continuous learning helps to build the cognitive reserve that helps stave off malignant neurodegenerative diseases like dementia and Alzheimer's disease.
Are you dealing with any kind of challenge like understanding AI and figuring out how to apply it in real-world applications?
03 How I Learned Artificial Intelligence
I have completed the fast.ai course content, but also continue to follow practitioners on my self-curated Twitter list [28] and put notes into my public GitHub AI repository [29] and Latent Space Discord [30]. I start reading most of the more important new papers the week they are published , and I also try to run or read the code of projects and products that get a lot of likes . Our upcoming "Fundamentals 101" series on the podcast, which includes AI fundamentals, has forced me to read more papers and learn the history of some things we take for granted today (Note 6).
https://github.com/sw-yx/ai-notes/blob/main/Resources/Good%20AI%20Podcasts%20and%20Newsletters.md
Notes:
-
In both career transitions, I didn't start from scratch - I was exposed to BASIC programming when I was 13, and when I was working as an options trader when I was 26, I wrote some extremely simple natural language processing code to parse brokers. Pricing for Vendors - I wish I could show you, but it's been so long that these are no longer available.
-
Publishing the content of the learning process in the community can achieve the fastest learning rate of human beings - L((PN)^2)! [31]
-
This AI wave is so big. Don't take my word for it, listen to Bill Gates [32] who says this is the most important technological advance since the GUI.
-
Winter is coming. On a certain day, this AI summer will end, and AI winter[33] will come again. What’s important to understand about this AI wave is that it’s likely to survive any winter, just as the internet industry only briefly paused after the 2001 recession.
-
Mandatory use of the term Generative AI would be unpleasant for us, as we all agree it is overhyped [34]…but no better alternative has yet been found.
-
Again, it's important to post learnings in the open community, for fear of damaging my personal reputation, I try to be as correct as possible, and puts extra pressure on me to make mistakes.
END
References
[1]https://learninpublic.org/
[2]https://www.swyx.io/learn-in-public
[3]https://thedecisionlab.com/biases/the-sunk-cost-fallacy
[4]https://en.wikipedia.org/wiki/General-purpose_technology
[5]https://twitter.com/swyx/status/1640561992472866816
[6]https://www.coursera.org/specializations/machine-learning-introduction
[7]https://www.fast.ai/posts/2016-10-08-course-background.html
[8]https://www.fast.ai/posts/part2-2022.html
[9]https://twitter.com/Suhail/status/1591813110230568963?ref=hackernoon.com
[10]https://twitter.com/karpathy/status/1468370605229547522
[11]https://twitter.com/search?q=%22just%20matrix%20multiplication%22&src=typed_query&f=top
[12]https://www.latent.space/p/ok-foomer
[13]https://www.linkedin.com/in/goodside/
[14]https://twitter.com/swyx/status/1616541173996482560?lang=en
[15]https://twitter.com/goodside/status/1617735459026915329
[16]https://news.ycombinator.com/item?id=35111646
[17]https://changelog.com/podcast/532#transcript-8
[18]https://twitter.com/ggerganov/status/1635636358126370817
[19]https://a16z.com/2022/11/16/creativity-as-an-app/#section–1
[20]https://langchain.com/
[21]https://en.wikipedia.org/wiki/Emad_Mustache
[22]https://twimlai.com/podcast/twimlai/stable-diffusion-generative-ai/
[23]https://research.runwayml.com/the-research-origins-of-stable-difussion
[24]https://github.com/CompVis
[25]https://stratechery.com/2022/an-interview-with-daniel-gross-and-nat-friedman-about-the-democratization-of-ai/
[26]https://techcrunch.com/2022/10/18/ai-content-platform-jasper-raises-125m-at-a-1-7b-valuation/
[27]https://www.goodtherapy.org/blog/change-is-a-choice-nurturing-neuroplasticity-in-your-life-0930154
[28]https://twitter.com/i/lists/1585430245762441216
[29]https://github.com/sw-yx/ai-notes/
[30]https://discord.gg/xJJMRaWCRt
[31]https://www.swyx.io/big-l-notation
[32]https://www.gatesnotes.com/The-Age-of-AI-Has-Begun
[33]https://en.wikipedia.org/wiki/AI_winter
[34]https://www.latent.space/p/why-prompt-engineering-and-generative
This article is authorized by the original author and compiled by Baihai IDP. If you need to reprint the translation, please contact us for authorization.
Original link :
https://www.latent.space/p/not-old