All things data, check out SingleStore: tiny.one/DGOpgdT
Outline
Is AI all hype or is it really as profound as fire? In this video, we go on a journey to find the top reasons why AI might fail and not live up to expectations.
Get ready as we review 4 solid arguments on why AI could be a total dud:
* AI Hallucinations - Models make up fake info confidently - how can we trust them?
* Too Complex to Work - These giant models are too unpredictable - debug that!
* Models are too complex - Are models hitting a plateau? Is this as smart as they get?
* Prompting Problems - Is it just too hard for humans to talk to AIs?
After reviewing all the counterarguments, where do we land? Is AI still worth the hype? Watch to find out and join the journey to see how AI plays out. The future is unknown but exciting.
Overview
0:00 - AI Hype
1:08 - Hallucinations Stop Progress
2:35 - Models Are Too Complex
4:08 - SingleStore Sponsor
5:37 - Models Won't Get Better
6:54 - Advanced Prompting Is Too Hard
8:06 - Conclusion
Companies Mentioned:
- getmetal.io/
Sponsors that help support the channel:
- SingleStore (All In One Database): tiny.one/DGOpgdT
Greg’s Info:
- Twitter: tiny.one/GJYrcT8
- Newsletter: tiny.one/fPCn0po
- Website: tiny.one/Xmuqpnt
- LinkedIn: tiny.one/wgHhZ0m
- Work with me: tiny.one/M9Pe9Jg
- Contact Me: Twitter DM, LinkedIn Message, or contact@dataindependent.com
What are other reasons why AI might not work? It's a Steelman argument, so have fun with it!
You also asked the question in a bit of a different way at the start of the video. Why won't AI work in the workforce. I think this question is easier to answer. I think the future of white collar jobs will be gone and AI DAOs that operate as quasi charities for UBI payments will be the future. I don't see many people other than David Shapiro talking about post-economics.
LLMs are useful when text generation needs to be automated in a way that is mindless but produces output that looks like something a person would write. That's what they were made to do, and they're really good at it. If you need to crank out a first draft of a common type of document (e.g. a job ad or a resume or a customer feedback form) and you're ready and willing to put in some time editing what it cranks out, use an LLM! But if truth is important, or internal logical consistency is important, or if keeping track of pieces of information is important, or if the task at hand requires real human understanding of what's going on, then watch out. LLMs have no concept of truth, they're lousy at logical consistency, they're lousy at retaining and applying facts, and they don't understand squat. This is why the "hallucination" problem is not going to be solved without a radical technological change. One of the reasons LLMs work so well is that they're unconstrainted by notions of truth, logistical consistency, coherence with the real world, etc. You can't just take an algorithm for cranking out text that looks kinda like the text it was trained on and then add in a constraint of "don't say things that are false". It doesn't know what that means. An LLM will not replace any human being whose job requires understanding human interaction. This includes things as seemingly simple as scheduling appointments or answering calls made to a general customer service line and forwarding them to the appropriate department. This stuff requires understanding language, something that LLMs can't do but are great at faking. So the answer to "why might AI not work?" is "we try to get it to do stuff that it can't do." And that's a hard problem. It's really tempting to imagine it can do stuff that it can't do, considering that it's been optimized to sound as convincing as possible no matter what.
I think this argument is limiting AI to be only about large language models (LLMs) such as ChatGPT and others, which is obviously not the case. AI is a bigger world than LLMs, which are only one of the many applications of artificial intelligence. AI encompasses a variety of fields and disciplines other than natural language processing, such as computer vision, robotics, machine learning, knowledge representation, reasoning, automation, and more. AI is a multidimensional and interdisciplinary domain that has already been applied in a diversity of industries, including medicine, healthcare, manufacturing, security, and others, with plenty of success stories.
I think most commentators on your videos don't seem to understand you're actually are showing how IT DOES actually succeed... It comes down to people not understanding the thing/how to use the thing.
Great video Greg! Very well summarized.
Ha thanks Stephen!
How expensive they are to run! I'm interested in distilled and smaller versions of models for this reason.
When i calculated Parameter Efficiency (Watt / Parameter) against humans lowest estimates became one of the main reasons why AI can't compete for very very long time. Now if you compare AI against human networks using other networks and deployables as tools, difference is incomprehensible.
Hallucinations strongly remind me of a child trying to tell you a lie when all they wanted to do was make the parent happy. It's not that child's bad it's just the child hasn't learned that honesty is the best trust builder. The child is scared you want like them.😢😅😂
They make it to large. I have a brain injury and my brain rewired itself. There are many parts of the brain that are capable of doing things, but not everything. I rewired to things most people can't do, but there are things that you do that I can't. If you want more from AI make them smaller and more focused on a task. Combine them with one in charge. Langchain is a good start I think. Like I can't see picture in my head, but I can describe things in words, but I can't see them. It's kind of like ChatGPT.
That’s great news for people building and selling business. « Simplification » is widening the gap of know how.
A good definition would be necessary, since it's not a binary issue. It does work, you asking chatGPT and it understanding your question already proved that. Case closed. How far it will go, how fast, and if the technology will make these companies as profit as one might anticipate are all questions on their own. The other thing is that LLMs and deep learning are not all of AI, and not all of what AI will be. I hurts that it's necessary to explain that, but it's great to know how "limited" people are and conclude that there are great opportunities for that reason.
👍
The output of LLMs is deterministic when you set the temperature to zero.
Why not value for ordinary people ? Do business really need more power over us ? We are Billions of people needing a defense against 'business' intrusion into our lives - daily. Why not use AI to help people instead ??
There are tons of open source models that perform pretty good. You can build the solution. Or at least one that fits your particular use case.
The wheel wasn’t a great invention until someone invented the axle. AI won’t be great until we have quantum computers. But it’s off to a good start.