I am not sure if you are being sarcastic because I don't know how people view IEEE "digital badges", but anything from MOOCs on LinkedIn stopped being valuable a long time ago, if it ever was.
Oh that was very much sarcasm. I have no idea why I'd pay for five hours of recorded videos from two guys from Samsung when I can get top-tier academic and industry content for free. I'm happy to pay for good education but nothing about this training has $240 of value to me.
I like how the correct, optimal suggestion is being downvoted by people who've never tried it, or who last tried it in 2022.
Suggest going through the papers (or the subset that interests you) listed at https://news.ycombinator.com/item?id=48822131 with a GPT/Claude/Gemini chat window open. Supplement with the Karpathy 'Zero to Hero' video series if it suits your preferred learning style. That will get you where you want to be, in terms of ML knowledge. It won't get you a job at Anthropic, but neither will a paid IEEE course.
This simply isn't true. Given that the whole promise of AI is accessibility, there isn't an obstacle being raised by other people adopting it faster. You can always pick up the latest trick quickly, and if you struggle at all, an LLM can explain. There is no evidence of people falling behind.
The idea that you need specialised knowledge to compete with the tool that is designed to let you do things without specialised knowledge is trivially nonsensical.
Admittedly part of the hype around AI is that you won't have to need to know what you're doing in order to use it. But there's a massive downside if you believe that and it turns out to be wrong.
As with any powerful but potentially-hazardous tool that is not perfectly reliable, I personally prefer to know what I'm doing and how the tool works. That's why I still have 10 fingers, 2 eyes, and a job.
I agree for every other tool, but AI's unique proposition--not part of the hype, the whole thing--is that it can do the thinking for you. If AI becomes what everyone claims, then you can't fall behind: you can always ask it to catch you up.
Everyone knows how to use AI; it's designed to be accessible. The people who think it's a special unique skill are the ones behind every vibe-coded disaster.
Don't people get better results by putting LLMs inside the right kind of loops and using the right kind of prompts? If I go to Claude and prompt "make me a website" it certainly doesn't produce as good of a result as other people with more sophisticated processes get.
I haven’t seen loops work for anything beyond simple tasks. And for complex tasks the loop itself is so complex half the time it would just fail and need you to intervene. Why would you go through all that effort of engineering a loop when a descriptive prompt followed by inspection is more than enough. It’s like idle hands looking for something to do.
And Claude or gpt can make websites just fine. So can the open weight models.
LLM training courses may have some valuable tips and tricks behind them, but the platforms change so often and no two personalized LLMs look the same. It feels like prompting isn't a science you can capture with step-by-step tutorials, but rather it's an art form you compose. Can start from the same place and get two completely different outcomes.
Search for the DOI string that is associated with virtually every paper (which often but not always looks like a URL) and paste it into scihub.ru. If you are primarily interested in older papers (pre-2022 or so) this will be the path of least resistance, or at least minimum loop area.
>>Relying on such LLMs without understanding their internal logic creates a significant reliability risk. To build tools that work consistently, developers must understand the core principles that govern how the models process information and generate results. By mastering how a model processes information and how its internal settings influence the result, developers can move away from a trial-and-error approach toward a more precise one to ensure the AI tool handles complex data reliably.
This is staggering bullshitp. In what way does understanding a transformer allow you to solve the core problem of LLM's that no frontier lab has managed to resolve?
>>To fix the problem, retrieval-augmented generation (RAG) forces AI to look up information in a trusted source such as a company’s database.
This also is bullshit. Yes, RAG helps and reduces errors, but NOOOO! it does not fix hallucinations...
>>Prioritizing data security. When using AI with proprietary code, security is a major concern. Engineers must learn how to set up “private” instances of the models to ensure that sensitive company data stays within a secure cloud environment and is not used to train public versions.
This is somewhat true, but really the motive is providing a soverign instance that cannot be withdrawn for arbitary reasons. Fundamentally the big providers are not going to steal your data, they may change the license to allow them to use it in the future, but then all their big customers will leave. So, they won't be able to, probably. What might well happen (and has happened) is that the USA might withdraw access with no notice leaving you high and dry.
I want to learn to build a real LLM so I looked at https://allenai.org/olmo where there are instructions and ingredients. But, unfortunately I can't afford the required compute resource so I will have to wait for a bit I guess.
Personally, I think understanding deeply how a transformer works helps a lot to understand what's probably the result of specific choices in the RL process vs what's architecture. A lot of the "We asked 30 LLMs and they all said the same thing" type analyses of how LLMs work often bump into what's being prioritized in the name of alignment right now, as opposed to architectural insights.
Why pay for this when Stanford has a playlist of a free course on LLMs: https://youtube.com/playlist?list=PLoROMvodv4rOCXd21gf0CF4xr...
Saved, thank you.
Here is the linked course in the article: https://iln.ieee.org/public/contentdetails.aspx?id=B570F53B5...
$240 (non member price) for a 5 hour course.
Did I read that right? Or is it more 5 hours of instructional videos?
Either way, it doesn't seem to include grading or other help etc.
Yes but you get a digital badge with it, so that's nice.
I am not sure if you are being sarcastic because I don't know how people view IEEE "digital badges", but anything from MOOCs on LinkedIn stopped being valuable a long time ago, if it ever was.
Oh that was very much sarcasm. I have no idea why I'd pay for five hours of recorded videos from two guys from Samsung when I can get top-tier academic and industry content for free. I'm happy to pay for good education but nothing about this training has $240 of value to me.
Better spent on a $200 OpenAI or Anthropic sub and having their top models give you instant, personalized teaching.
I like how the correct, optimal suggestion is being downvoted by people who've never tried it, or who last tried it in 2022.
Suggest going through the papers (or the subset that interests you) listed at https://news.ycombinator.com/item?id=48822131 with a GPT/Claude/Gemini chat window open. Supplement with the Karpathy 'Zero to Hero' video series if it suits your preferred learning style. That will get you where you want to be, in terms of ML knowledge. It won't get you a job at Anthropic, but neither will a paid IEEE course.
The problem with LLM courses is that the topic is mostly alchemy and will not bring you much real enlightenment.
It's alchemy that works and if you don't know it you are left behind, so that's important.
> if you don't know it you are left behind
This simply isn't true. Given that the whole promise of AI is accessibility, there isn't an obstacle being raised by other people adopting it faster. You can always pick up the latest trick quickly, and if you struggle at all, an LLM can explain. There is no evidence of people falling behind.
The idea that you need specialised knowledge to compete with the tool that is designed to let you do things without specialised knowledge is trivially nonsensical.
Admittedly part of the hype around AI is that you won't have to need to know what you're doing in order to use it. But there's a massive downside if you believe that and it turns out to be wrong.
As with any powerful but potentially-hazardous tool that is not perfectly reliable, I personally prefer to know what I'm doing and how the tool works. That's why I still have 10 fingers, 2 eyes, and a job.
I agree for every other tool, but AI's unique proposition--not part of the hype, the whole thing--is that it can do the thinking for you. If AI becomes what everyone claims, then you can't fall behind: you can always ask it to catch you up.
The obstacle being raised is that people who know how to use AI are getting jobs instead of you.
Everyone knows how to use AI; it's designed to be accessible. The people who think it's a special unique skill are the ones behind every vibe-coded disaster.
There is no skill involved in using AI. Most of the stuff people do on top of frontier models is just cope. It’s literally the bitter lesson repeated.
Don't people get better results by putting LLMs inside the right kind of loops and using the right kind of prompts? If I go to Claude and prompt "make me a website" it certainly doesn't produce as good of a result as other people with more sophisticated processes get.
I haven’t seen loops work for anything beyond simple tasks. And for complex tasks the loop itself is so complex half the time it would just fail and need you to intervene. Why would you go through all that effort of engineering a loop when a descriptive prompt followed by inspection is more than enough. It’s like idle hands looking for something to do.
And Claude or gpt can make websites just fine. So can the open weight models.
only for a given value of works
"Gets shit done" is a good-enough definition of "works" in most economically-valuable contexts.
LLM training courses may have some valuable tips and tricks behind them, but the platforms change so often and no two personalized LLMs look the same. It feels like prompting isn't a science you can capture with step-by-step tutorials, but rather it's an art form you compose. Can start from the same place and get two completely different outcomes.
I didn’t realize IEEE had courses. I’m curious if anyone can comment on the general quality and if they have any good ones.
I really want to read a bunch of old papers from the IEE library but it looks very hard to get access to at any reasonable price for personal usage.
Anyone got any hints or tips for me?
Search for the DOI string that is associated with virtually every paper (which often but not always looks like a URL) and paste it into scihub.ru. If you are primarily interested in older papers (pre-2022 or so) this will be the path of least resistance, or at least minimum loop area.
>>Relying on such LLMs without understanding their internal logic creates a significant reliability risk. To build tools that work consistently, developers must understand the core principles that govern how the models process information and generate results. By mastering how a model processes information and how its internal settings influence the result, developers can move away from a trial-and-error approach toward a more precise one to ensure the AI tool handles complex data reliably.
This is staggering bullshitp. In what way does understanding a transformer allow you to solve the core problem of LLM's that no frontier lab has managed to resolve?
>>To fix the problem, retrieval-augmented generation (RAG) forces AI to look up information in a trusted source such as a company’s database.
This also is bullshit. Yes, RAG helps and reduces errors, but NOOOO! it does not fix hallucinations...
>>Prioritizing data security. When using AI with proprietary code, security is a major concern. Engineers must learn how to set up “private” instances of the models to ensure that sensitive company data stays within a secure cloud environment and is not used to train public versions.
This is somewhat true, but really the motive is providing a soverign instance that cannot be withdrawn for arbitary reasons. Fundamentally the big providers are not going to steal your data, they may change the license to allow them to use it in the future, but then all their big customers will leave. So, they won't be able to, probably. What might well happen (and has happened) is that the USA might withdraw access with no notice leaving you high and dry.
I want to learn to build a real LLM so I looked at https://allenai.org/olmo where there are instructions and ingredients. But, unfortunately I can't afford the required compute resource so I will have to wait for a bit I guess.
Personally, I think understanding deeply how a transformer works helps a lot to understand what's probably the result of specific choices in the RL process vs what's architecture. A lot of the "We asked 30 LLMs and they all said the same thing" type analyses of how LLMs work often bump into what's being prioritized in the name of alignment right now, as opposed to architectural insights.
[dead]