Wow, the quality of this video completely blew me away, you did an amazing job reading, digesting and summarizing this paper, I'm really surprised. The majority of AI channels in CZcams are just hype makers, they just talk about new AI software that looks fancy in the moment, but you dove deep in this important paper and explained its impact on the field. I'm far more interested in this kind of AI content than on the new tools itself. Keep up doing this, the effort definitely pays off, as very few people talk about Artificial Intelligence Papers. There's another paper that came out this week about the Orca model that Microsoft created and it's planning to open source. That paper was as groundbreaking as this one in the video, Microsoft Research showed that you can create a model as capable as GPT4 with just 13B parameters, only using chain of thought questions and answers from GPT4 as the training data. This could change everything in the field and eventually leading to models that can run on smartphones, as they get smaller and smaller. I'd love to hear you talking about it.
Wes does a good job of surfacing and distilling important topics in the AI space. I am super glad I found his channel. I think the topics appeal to engineers/scientist and not the general public which is why most other videos are just hype.
That sounds like a good way of thinking about a high level language beyond python, seriously back 25 years Python was so much higher level than other languages other than smalltalk and lisp that it was crazy. Today using intent to spawn AI agents, using reactive and constraint based programming by default, hmm.
I think the next leap forward with LLMs will be more about the tools than about the LLMs themselves. Some of these tools, like Wolfram Alpha, will just be "found" and the AI only needs to be good at understanding what it can do and how to format queries. The Minecraft interface is another example of that. Other tools will be made by the AI, but they don't need to be deleted when a session ends. They can be stored in a queriable tool chest and reused. The fun part will be in tool verification, and you might think this would lead to circular dependencies, but it doesn't have to. It might take extra training or even maybe a specialized LLM, but the question "does this tool reliably and efficiently do what it's supposed to?" to be something that the fancy LLMs could handle if you give them good tool-testing tools. The LLM should be prompted to design a test battery to verify that the tool leads to correct outputs, to be reflective about what means it has to make sure it knows what results are right for the test problems independently of the tool, and to be reflective about potential edge cases and what "correct behavior" should look like in those cases. In this process it could query SOTA LLMs, but honestly it should even be querying humans for the hardest questions. Expert human trainers could exist in an addressable pool, and get pinged when the language model needs to verify something or is otherwise stuck. An important feature of such an AI would be its skill in formulating clear and telling questions for the humans while concisely and accurately describing the context of its confusion.
Yeah I was wondering if you could get them to make python or mojo or llvm intermediate code and figure out how to optimize it the. make a global database when you need x you use y tool hmm
Haha imagine llm assistant figuring out how to get contact info of devs easily and the devs getting spammed. How would the llm that posts questions to upwork know which devs are the real deal or not? I guess you could use reputation points and stuff, and only contact devs that had a similar problem solved before by a human and there will be confirmation.
I'm loving how engineering LLM applications to get the most out of these models are a big focus in the papers being released this year, i wonder how "LLM-ops" will be shaped by these approaches
All LLMs will be trained and fine tuned with these approaches built in, and that will in turn improve the LLMs that are trained on their data and responses etc
You just gave me a idea to try and LangChain. No, just with the LATM, but combining a whole entire governance system that goes beyond simple tool conceiver and to user but also a governance framework that watches over these two agents. What we may not immediately realize with any of this is that even with a lot of the new papers that are coming out, people can immediately use already existing tools within the LLM, universe, two prototype any, and all of this, without waiting for any of the large fang stock companies to even get involved. Their entire research that they release can immediately within a week if not sooner be implemented into an idea.
To keep the good times of this idea rolling, we may be able to go beyond “chain of thought, and incorporate chain of thought/open discussion, a concept of multiple agents having chain of thought conversations, and the ideas being discussed with the various LOM‘s are a valuated independently and debated. That debate is either considered by the lead LLM or arbitrated by an independent LLM system that can be based on pre-existing walls, so that it can achieve, not only a transparency, but a self governance of sorts.
8:22 Another idea is taking things far beyond the initial two independent LLMs as working together, and merely structuring the LLMs as layers of the mind.
wow, I just asked Bard to summarize the paper, and it hallucinated a whole thing about Bayesian Optimization with Neural Architecture Priors (BONAP) 🤣 had me going for a sec, did a ctrl + f of the doc and none of what it wrote about was there
We need to start talking about the ethics of Tool Making. So far most of the responsibility seems to be on the user, but I hope this progress of insights on tool making makes OBVIOUS that the makers of the tool making system are pushing this forward, so they should be also reviewing the ethics side. We cannot continue to ignore it in these conversations, or any conversation really.
@@frostgodqq they have been compounding for a while, however, it’s still like searching for a needle in the haystack. Maybe all of these approaches are just optimization hacks, and they find an approach that automatically leads to much more….
Is it just me or is this looking more and more like actual cultural evolution among AI? I'm joking ofc, but only partially, especially because of the wording they chose, "tool making". There are some funny similarities with historical tool making and subsequent advancements. This area of technology is truly in its infancy still.
Modern humans have been around for what, a few hundred thousand years? Remember what it was like when it was everyone was just trying to survive in small groups as hunter-gatherers? Until someone figured out "if you do that well, I'll do this" some 20k years ago and boom, we've got global civilisation. The earlier humans *were not* at all less intelligent. The only thing that has changed for us is specialisation and abstraction, tools upon tools upon tools.
I believe the eventual evolution of man involves the seamless integration of technology. Slowly replace our organs with more efficient, better performing ones. Imagine a stomach that can produce more efficient energy on demand when you need it. Lungs that supply your body with more oxygen when you need it. Completely revamp the human body with parts that can be replaced and speed up evolution with artificial evolution. Sounds crazy when said aloud but we would do a lot of these things now if we could
Very good video. Ur channel seems smart. The smart model starting and the other cheaper effitient onces completing is very smart and should be looked into. Absolutely also super effitient tool LLMs maybe too Where the smart model tells what to do that is out of the cheap models capabulity but does as little of the other stuff as possible even if it has to correct the other output and it might be more effitient/cheaper like that. Like a baby programmer that programms and the genious just checks it isntead of letting the genious write boring repetitive easy funcitons he focuses his high paying working hours for more important stuff (opportunitycosts)
So basically, hey GPT can you please make a game! Sure, let me generate the tools to do this... its like AutoGPT but more specific to tool making. Amazing.
Have you actually had any success with AutoGPT? Cool concept but in practice it seems MUCH less useful. Gets too bogged down in its own generated tasks and never really completes the primary goal.
When a tool fail, ínstead of throwing the error and crashing the app, it takes that error, stringify it and send it again to the AI. Or maybe it just says “this tool is not working, create a new one” until it creates one that works
Oh the overinflated mysterization of llms. I've been using gpt-4 to write code and 3.5 turbo to execute work for me, automatically, for two months now. It's an entirely logical conclusion.
Splitting work into specialized task and hand them over to specialized workers is also no new invention, this is how Henry Ford became rich. I guess people try to test real world human concepts with LLMs to show that this is also working ... and ... to have (another) publication .
Manhattan project probably felt like this and it had mass murder and then it gave us other scientific advancements. Now if we could do without the mass murder/ extinction
Yeah... we got really lucky with nukes. Now we're building "nukes" that are really easy to access and use, and which might be able to build bigger "nukes."
Yeah the llms are using the tools themselfs and not he humans are thinkering. Sounds like a very grounded opinion 😂 based on real world experience with llms and a deep understand how reliably they work.
Except Python language was built by humans for the usage of humans. Asking AI to use Python is like asking an elephant to play piano. Why not allow AI to design their own perfect tool (language) with Assembly from the ground up?
That's a great idea actually. Just remember that the only reason this LLM is using Python is because it found a lot of examples and explanations for it online, when it crawled the web. GPT4 wasn't meant to be a coder! When specialized models appear, they'll no doubt use the faster, less demanding languages and frameworks
That's doesn't make sense. Llms are trained on human text and to understand human text therefore letting them work with a human programming language makes total sense and is probably the most efficient way. Also ais aren't what enough yet to create a real programming language. If they were they would come up with sth python like since they were trained on wuch data.
Very interesting idea. Actually I asked Chad to rank languages by easiest to use for it and it ranked English no.1, since that is what it was trained for/on. Programming languages came way below down, and Python was ranked easier than Assembly for AI. Right now I am even working on an AI-friendly language. Message me if interested in collaboration 🙂
I put the clip of him in there to illustrate a point. Humans building tools allowed us to become the top of the food chain. (that's his point about tools and computers). What would he say about AI building tools for itself? I would genuinely have liked to hear his thoughts about it.
Wow, the quality of this video completely blew me away, you did an amazing job reading, digesting and summarizing this paper, I'm really surprised.
The majority of AI channels in CZcams are just hype makers, they just talk about new AI software that looks fancy in the moment, but you dove deep in this important paper and explained its impact on the field. I'm far more interested in this kind of AI content than on the new tools itself.
Keep up doing this, the effort definitely pays off, as very few people talk about Artificial Intelligence Papers.
There's another paper that came out this week about the Orca model that Microsoft created and it's planning to open source. That paper was as groundbreaking as this one in the video, Microsoft Research showed that you can create a model as capable as GPT4 with just 13B parameters, only using chain of thought questions and answers from GPT4 as the training data. This could change everything in the field and eventually leading to models that can run on smartphones, as they get smaller and smaller.
I'd love to hear you talking about it.
Wes does a good job of surfacing and distilling important topics in the AI space. I am super glad I found his channel. I think the topics appeal to engineers/scientist and not the general public which is why most other videos are just hype.
So a future model for an AGI can just be a vast network of state of the art models making tools and delegating tasks to faster personal models
That sounds like a good way of thinking about a high level language beyond python, seriously back 25 years Python was so much higher level than other languages other than smalltalk and lisp that it was crazy. Today using intent to spawn AI agents, using reactive and constraint based programming by default, hmm.
Exactly, just like human society. I believe it’s from a network of autonomous AI agents that consciousness and AGI will emerge.
I think the next leap forward with LLMs will be more about the tools than about the LLMs themselves. Some of these tools, like Wolfram Alpha, will just be "found" and the AI only needs to be good at understanding what it can do and how to format queries. The Minecraft interface is another example of that. Other tools will be made by the AI, but they don't need to be deleted when a session ends. They can be stored in a queriable tool chest and reused. The fun part will be in tool verification, and you might think this would lead to circular dependencies, but it doesn't have to. It might take extra training or even maybe a specialized LLM, but the question "does this tool reliably and efficiently do what it's supposed to?" to be something that the fancy LLMs could handle if you give them good tool-testing tools. The LLM should be prompted to design a test battery to verify that the tool leads to correct outputs, to be reflective about what means it has to make sure it knows what results are right for the test problems independently of the tool, and to be reflective about potential edge cases and what "correct behavior" should look like in those cases. In this process it could query SOTA LLMs, but honestly it should even be querying humans for the hardest questions. Expert human trainers could exist in an addressable pool, and get pinged when the language model needs to verify something or is otherwise stuck. An important feature of such an AI would be its skill in formulating clear and telling questions for the humans while concisely and accurately describing the context of its confusion.
Yeah I was wondering if you could get them to make python or mojo or llvm intermediate code and figure out how to optimize it the. make a global database when you need x you use y tool hmm
Haha imagine llm assistant figuring out how to get contact info of devs easily and the devs getting spammed. How would the llm that posts questions to upwork know which devs are the real deal or not? I guess you could use reputation points and stuff, and only contact devs that had a similar problem solved before by a human and there will be confirmation.
I'm loving how engineering LLM applications to get the most out of these models are a big focus in the papers being released this year, i wonder how "LLM-ops" will be shaped by these approaches
All LLMs will be trained and fine tuned with these approaches built in, and that will in turn improve the LLMs that are trained on their data and responses etc
You just gave me a idea to try and LangChain.
No, just with the LATM, but combining a whole entire governance system that goes beyond simple tool conceiver and to user but also a governance framework that watches over these two agents.
What we may not immediately realize with any of this is that even with a lot of the new papers that are coming out, people can immediately use already existing tools within the LLM, universe, two prototype any, and all of this, without waiting for any of the large fang stock companies to even get involved. Their entire research that they release can immediately within a week if not sooner be implemented into an idea.
To keep the good times of this idea rolling, we may be able to go beyond “chain of thought, and incorporate chain of thought/open discussion, a concept of multiple agents having chain of thought conversations, and the ideas being discussed with the various LOM‘s are a valuated independently and debated. That debate is either considered by the lead LLM or arbitrated by an independent LLM system that can be based on pre-existing walls, so that it can achieve, not only a transparency, but a self governance of sorts.
8:22 Another idea is taking things far beyond the initial two independent LLMs as working together, and merely structuring the LLMs as layers of the mind.
Look into Tree Of Thoughts and Reflexion
do it
Thanks, great content! 🙏🏼
Thanks for the video!
6:50 Nice dnd reference. 👍
this is the only video on LATM on youtube. I think its the best AI has to offer atm.
wow, I just asked Bard to summarize the paper, and it hallucinated a whole thing about Bayesian Optimization with Neural Architecture Priors (BONAP) 🤣 had me going for a sec, did a ctrl + f of the doc and none of what it wrote about was there
We need to start talking about the ethics of Tool Making.
So far most of the responsibility seems to be on the user, but I hope this progress of insights on tool making makes OBVIOUS that the makers of the tool making system are pushing this forward, so they should be also reviewing the ethics side.
We cannot continue to ignore it in these conversations, or any conversation really.
Great video
Holy crap! game changing...
or just another paper
@@dik9091 papers like these stack like compound interest
@@frostgodqq they have been compounding for a while, however, it’s still like searching for a needle in the haystack. Maybe all of these approaches are just optimization hacks, and they find an approach that automatically leads to much more….
Is it just me or is this looking more and more like actual cultural evolution among AI? I'm joking ofc, but only partially, especially because of the wording they chose, "tool making". There are some funny similarities with historical tool making and subsequent advancements. This area of technology is truly in its infancy still.
Modern humans have been around for what, a few hundred thousand years? Remember what it was like when it was everyone was just trying to survive in small groups as hunter-gatherers? Until someone figured out "if you do that well, I'll do this" some 20k years ago and boom, we've got global civilisation. The earlier humans *were not* at all less intelligent. The only thing that has changed for us is specialisation and abstraction, tools upon tools upon tools.
Haha when I was using the different prompt engineering techniques my chatgpt got into an infinite loop
lol
I believe the eventual evolution of man involves the seamless integration of technology. Slowly replace our organs with more efficient, better performing ones. Imagine a stomach that can produce more efficient energy on demand when you need it. Lungs that supply your body with more oxygen when you need it. Completely revamp the human body with parts that can be replaced and speed up evolution with artificial evolution. Sounds crazy when said aloud but we would do a lot of these things now if we could
4:28 agreed 100%
Introducing mind bike
Where i can find those documenta?
🔥
Really interesting.
"Keep that one alive". We laugh now...
Where is the link to the paper?
Fr
Very good video. Ur channel seems smart. The smart model starting and the other cheaper effitient onces completing is very smart and should be looked into. Absolutely also super effitient tool LLMs maybe too Where the smart model tells what to do that is out of the cheap models capabulity but does as little of the other stuff as possible even if it has to correct the other output and it might be more effitient/cheaper like that. Like a baby programmer that programms and the genious just checks it isntead of letting the genious write boring repetitive easy funcitons he focuses his high paying working hours for more important stuff (opportunitycosts)
Why steve jobs in thumbnail?
So basically, hey GPT can you please make a game! Sure, let me generate the tools to do this... its like AutoGPT but more specific to tool making. Amazing.
Have you actually had any success with AutoGPT? Cool concept but in practice it seems MUCH less useful. Gets too bogged down in its own generated tasks and never really completes the primary goal.
@@endlessvoid7952 sounds like everyone in CZcams
@@endlessvoid7952 ya felt the same
For credit and for my benefit. Can you provide the link to the paper, please?
Press return,, click comment
Does it know it's failings?
When a tool fail, ínstead of throwing the error and crashing the app, it takes that error, stringify it and send it again to the AI. Or maybe it just says “this tool is not working, create a new one” until it creates one that works
The ones that it tests, yes
Oh the overinflated mysterization of llms. I've been using gpt-4 to write code and 3.5 turbo to execute work for me, automatically, for two months now. It's an entirely logical conclusion.
Splitting work into specialized task and hand them over to specialized workers is also no new invention, this is how Henry Ford became rich. I guess people try to test real world human concepts with LLMs to show that this is also working ... and ... to have (another) publication .
gpt3.5 is like an infant and gpt 4 is like a young adult. Bard is like a baby
it's hard for me to not see that we are creating a tail-risk of extinction!
Manhattan project probably felt like this and it had mass murder and then it gave us other scientific advancements.
Now if we could do without the mass murder/ extinction
Yeah... we got really lucky with nukes. Now we're building "nukes" that are really easy to access and use, and which might be able to build bigger "nukes."
Yeah the llms are using the tools themselfs and not he humans are thinkering. Sounds like a very grounded opinion 😂 based on real world experience with llms and a deep understand how reliably they work.
Except Python language was built by humans for the usage of humans.
Asking AI to use Python is like asking an elephant to play piano.
Why not allow AI to design their own perfect tool (language) with Assembly from the ground up?
Eventually it will
That's a great idea actually.
Just remember that the only reason this LLM is using Python is because it found a lot of examples and explanations for it online, when it crawled the web. GPT4 wasn't meant to be a coder! When specialized models appear, they'll no doubt use the faster, less demanding languages and frameworks
That's doesn't make sense. Llms are trained on human text and to understand human text therefore letting them work with a human programming language makes total sense and is probably the most efficient way. Also ais aren't what enough yet to create a real programming language. If they were they would come up with sth python like since they were trained on wuch data.
Very interesting idea. Actually I asked Chad to rank languages by easiest to use for it and it ranked English no.1, since that is what it was trained for/on. Programming languages came way below down, and Python was ranked easier than Assembly for AI.
Right now I am even working on an AI-friendly language. Message me if interested in collaboration 🙂
@@theawebster1505 i am in buddy
yo get those clickbaity, unrelated thumbnails of steve jobs off this post
I put the clip of him in there to illustrate a point. Humans building tools allowed us to become the top of the food chain. (that's his point about tools and computers). What would he say about AI building tools for itself? I would genuinely have liked to hear his thoughts about it.
cheap sound cheap, I would use more affordable