This is the BEST channel for AI research and development. Thank you so much for putting the effort into these videos and not just reading off an announcement for 10 minutes. We really appreciate it!
I've said it before but I'll say it again: AI Explain is just on a league of his own here in UA-cam. - In-depth knowledge and testing - Quality and clarity of the presentation - Speed when putting out content He's made every other similar channel obsolete for me. Bravo.
@@electron6825 I guess it is well on the way and I can see how we might arrive there but it still has to be applied, built and rolled out. Most people I know IRL still don't know to Google something when they don't know the answer.
I'm a danish speaker. To me the poems makes sense. It reads a little weird, however that might just be me being unfamiliar with that style of poem. It is still very impressive, and i have found myself using gpt4 in my native tongue more often than i would have imagined.
Plotly deserves most of the credit here. Those plots are exceptionaly easy to make thanks to their library. We also need to be very careful here. There's no way to tell how much insight the model is extracting from the data and how much it's useing its pretrained knowledge. If the data contradicts its knowledge which will it tend to believe?
You should never use it to verify itself, or trust that it has predictable behavior regarding to how its sourcing its knowledge. It can easily make something up that sounds plausible. If you're not skilled in the field you're prompting it, you can use it as a hypothesis generator and then try to verify (or ideally try to disprove) it. Ever heard that Wikipedia is unreliable? Well, GPT is even less reliable. And just because it sometimes lies or denies capabilities, doesn't mean that it can't do the exact task correctly in some other prompt context. It's very quirky like that.
It's funny, those tasks it completes successfully in this video are essentially the exact same as those I was assigned in my final semester as a Business Administration major. It performs the same data analysis and creates the same visualizations and explanations that I had to, at a quality level not much worse. I graduated less than a month ago, and with AI already having started its recent explosion a couple months prior to that, I estimated at the time that it would probably be only a few months until it could reliably do everything they spent four years teaching me. So far it's looking like we're ahead of schedule. It'll take some more time to work out the kinks and make it more reliable, as well as enable more robust analyses, but I suspect by the time I start my new job (roughly one month from now), it'll be capable of that. Exciting times
@@maxidaho Oh, it's not just in jeopardy, it's going to be virtually nonexistent soon enough. Not much you can do but prepare for the degree to be useless
To be honest, most of the amazing results in the plots and charts should be attributed to the people who created the libraries it's using, not ChatGPT itself.
How do I even see what I can switch on? Under Settings / Beta Features I switched Plugins on, but it says "third-party plugins that you enable" and I see no list of plugins to enable
@@kmouratidis Indeed! At this point, these 'GPT' models are not even transformers, since devs have been integrating them with many other types of computing, machine learning, and other 'AI' systems.
@@kmouratidis The code would be pretty unimpressive and rather useless without the outputs of the transformers, though. Neither can be neglected, but transformers are the new thing between the two, and hence they’re more magic, or something.
those plots are most likely created with dash/plotly, gpt 3.5 can still write code for that, but tbh its amazing that gpt4 can do that but lets not forget the people who created the libraries being used.
Yeah, I recognized those plotting tools- while I certainly won't sneeze at the convenience of AI automating the ingestion of their data and analysis, the plotting is all from existing open source tools
I love when he says, maby its me thats easily impressed... YES; he is haha. (this time). But I agree, this plugin seems to be the one that actually is worth using for most of us normal people.... :D COme on, we had plotting libraries since the days of Flash... :D
@@kimsteinhaug my interpretation is that it's the new SPSS- a great way to run simple stats and visualizations and explorations with minimal experience and not requiring spending half of your life hating R
I've been checking for plugin access every day all week. Finally got access this morning. It's like 10 Christmases all put together. Granted, my Christmases generally weren't very good. But still....
Just a note, The Wolfram Alpha counting prompt did not actually call/use Wolfram Alpha. If it did you would see the request and response in rectangles. Just because a plugin is enabled doesn't mean it is used. I unfortunately know this as I'm developing a plugin at the moment and too many incidents of this happen to me :(. But your video is world class as usual!
Hey man, just want to tell you that i look forward to your videos all week. I really appreciate your work. I follow a lot of AI channels on youtube but you are easily the best.
10:57 that sounds exactly like a System Prompt that the session could be working with. Maybe it considered the System prompt the beginning instead of your first message. It would be really cool if that was the case, since it would give some insight into how the sessions are setup, which is usually hidden.
Yeah totally! It's super interesting that they told it not to make API calls as if they expected it would if Internet access wasn't disabled. Part of me wonders if Internet access is really disabled or if they just disable it by convincing ChatGPT not to try! :)
GPT-4 with Wolfram Alpha finally got its only weakness cured. the way of GPT-4 using Wolfram looks more humanlike, it read the context of the problem and sets roughly the formula and logic that fit, (on its own without any assistance from Wolfram, it is pretty good at it) and then, throughout the Wolfram, GPT-4 evaluate and supplement the process to get the correct answers. the accuracy of getting it correctly has enormously enhanced compared to GPT-4 alone. even though It still makes a few mistakes, with only one more simple prompt, It could fix errors. Seeing GPT-4's advancements, I'm 100% sure that one of the first massive replacements would happen in the educational field.
Have been waiting and neurotically refreshing the beta tab in ChatGPT in my settings tab ever since watching this video and was thrilled when I heard it was being pushed to everyone this week. Just had to come back to this video for inspiration on what to test, and cannot get over how freaking ridiculously cool this update is and the implications for the future. I could be wrong, but it really feels like Language Models doing novel research can’t be that much further out which feels like an inflection point for humanity.
These are great accelerants for data analysts and data scientists. Give yourself credit for being the guiding force to a lot of the analytics, and being able to validate them afterwards. Also note that you're using all public data here - it's going to be against any company policy to leverage this for much PII / proprietary data. So ... let's not go too far!
@@qozia1370 that's an absurd statement, like SPSS somehow replaced all social scientists. There are already a lot of tools for simple graphical and statistical analysis of data sets for people with relatively minimal experience, this is a better version. It will improve the ability to run simple repeated exploratory data analysis, which if anything will make data analysis better and more attractive as a field. Your statement is like saying the calculator eliminated mathematics.
Wolfram Alpha is, for the most part, an interface into a database + Mathematica. Any code that can be executed in Mathematica can be run in wolfram alpha. For example; LayeredGraphPlot[{4 -> 3, 5 -> 3, 5 -> 4, 6 -> 1, 6 -> 2, 6 -> 4, 6 -> 5, 6 -> 3}, VertexLabels -> Automatic] in Wolfram Alpha will draw the same graph (though formatted differently, and with extra information) as Mathematica. If you want to find a use for Wolfram Alpha that code interpreter can't do, you need to find something that Mathematica is good at that Python is bad at. If code interpreter can import SymPy and/or Sage, then there will be a lot of overlap in terms of symbolic capabilities. But some things that come to mind is Integrals and solving differential equations, which sage is really bad at in comparison to Mathematica. Mathematica is also better at solving complex problems symbolically than sage. For more complex problems, wolfram alpha is more likely to succeed in finding a solution. I also imagine some data structure manipulation algorithms, like finding Hamiltonian cycles, would be easier with wolfram alpha since it has more robust support for graph manipulation than any python package I'm aware of, though I know that some Python packages can do it with some pain, so maybe code interpreter can do that. Unfortunately, I do not have access to code interpreter, so I can't test these differences.
Funny, the first thing I tried with the Wolfram plugin was to give it values of miles between cities and have it draw a weighted graph and find the optimum cycle.
My comment for your last video was that I couldn't wait for you to test drive Code Interpreter - and you are again exceeding expectations!! Such a handful of "WOW!" experiments!! Would be awesome if you can you "go to town" on its Machine Learning capabilities! Muchas Gracias!
People keep talking about the web browsing plugin, but this is the one that has always impressed me. Giving chatgpt there power of running code turns it into an anything to anything converter, a powerful graphing tool, an easy way to make quick changes in files such as images and much more
Re your comment at 12:12 about solving the math equation - I just used the Wolfram plugin and it got the right answer in about 10 seconds, including a plot of the solutions. If that didn't work for you, I suspect it was just a case of GPT-4 not calling the plugin in the first place. "use wolfram" in the prompt tends to force that, I find.
@@aiexplained-official Not so far. At least not Wolfram itself, but I’ve seen problems with ChatGPT handing incorrect information off to Wolfram. E.g. , I gave it a projectile motion physics problem and ChatGPT gave Wolfram the wrong signs for the initial vertical component of velocity and acceleration due to gravity, so the quadratic equation solutions for the time to impact were imaginary numbers.
At the timestamp of 19:16, it's apparent that Wolfram Alpha wasn't utilized. It seems the prompt provided did not sufficiently stimulate its use - one might need to be more explicit in this regard. And at 20:10, you can also get the LaTeX rendered mathematical expressions with GPT-4. It just does not automatically resort to it, so you'd have to ask for it. The Code Interpreter model is likely more fine-tuned towards it, or it has been system prompted to provide rendered expressions. 22:00, I am being nit-picky here now, but you do not need to say "Proceed, without further questions" at every prompt. This is eating up the context window and is not good practice. I am sure you know why, and you probably just wanted to create a fast video demonstrating the different capabilities of the code interpreter. But nevertheless, it's good to note this. On a related note, I'd like to highlight the potent capabilities of Wolfram Alpha/Mathematica. Its exceptional mathematical prowess is likely the driving force behind its initial introduction as a plugin. During my Bachelor's studies in Computer Science, and now in my Master's journey in Artificial Intelligence, I find myself frequently relying on this invaluable tool. Additionally, I used your same question and gently nudged it to employ Wolfram Alpha for both counting and division task. The result was as proficient as with the Code Interpreter. I feel it might not be entirely fair to pit these two against each other. After all, when you're using Wolfram Alpha, you're basically using a 'Code Interpreter' already, but one that's more finely tuned for mathematical tasks.
The data analytics is a real game-changer. As it stands there are many people dedicated just to doing the visualizations alone. Drawing conclusions is mostly just an artform that lacks substantiation, and a lot of people get paid big bucks for those insights. Being able to comprehensively span large sectors of knowledge allows incredible hidden insights. Great stuff. I've been using it for code and have it where it does what I ask and also points out things I havent thought of. There is real gold in that. As usual thanks again for the effort you put in.
I find this very concerning. All the explanations it wrote were not at all based on the data points provided. They involved "common knowledge" about birth rate, health care and migration, but these kind of explanations are exactly what the model tends to hallucinate very confidently. I don't think most people will understand that and will take these conclusions without criticism.
@Antonio Ruby Yes, I was also thinking about that and it is a concern. Maybe when it gets to doing these extreme extrapolations it might be good if it triggers a safe Harbour caveat. But as far as the insights go, at least they give people something to think about that they may not have considered. It's funny, we seem to be doing RLHF to try and find some happy median between man and machine haha
@@antoruby yes, it's a trade-off between "Relate what you already 'know' about this field to this data" vs. blinding the data (rename age to "Increasing unsigned integer 1" and country to "String A") then asking it to analyze solely based on the numbers.
@@antoruby Those people who take conclusions without criticism would do so irrespective of the source, a human consultant can fool them quite as easily. It's really their problem.
I love the code interpretation of data. Being able to generate my own insights is a game changer. Data science is basically black magic in control of wizards - ChatGPT code interpreter unlocks this magic for the rest of us.
I’ve had it for a month now, it used to use gpt 3.5 and have unlimited prompts. The new cap is KILLING me! It should switch to code interpreter 3.5 after your 25 prompts are used up, not switch to the standard 3.5 version. The ability to incrementally modify it’s output to get exactly what you want is so damn useful. Also, in my opinion, 3.5 code interpreter is sooooo much better. They made a huge change right before everyone got access, but it’s context was MASSIVE before, I used to use it to convert a 1300 row csv of sales data that added up the weight, dimensions of items (based on a key file) matched items, calculated total order weight, dimensions, volume, total price, gave it a box key file and had it create an algorithm to distribute the volume and pick a correct box(s) and input that into a invoice template I gave it and organize by customer. All in one prompt, 2 full pages of a word doc with instructions. And it would do it and would follow your prompt EXACTLY. All at once, work with 5 excel/csv files, broken into over 40 individual steps. All running one by one without stopping until all tasks were complete. Now it does not do that, it’s hard to explain but it feels like it does not listen to you at all. Ignores your prompt and makes soooooo many mistakes it did not do before. Hopefully this is just a temporary downgrade !
On the topic of visualization and flashy things its less about what GPT-4 can do and more about the power of Python and the libraries surrounding it. But when it comes to data analysis. thats when it shines. The Wolfram alpha example at 19:31 didn’t actually use wolfram alpha. Sometimes you have explicit tell it to use the tool
There is one thing that wolfram alpha can do that this may not be able to do is that the base version of wolfram alpha on their human accessible website can do calculus, and is reasonably good at it, where I am unsure what the state of the libraries that allow that are in the interpreter, and that wolfram alpha can handle a lot of the symbolic manipulation natively that chat-gpt may get wrong occasionally. Additionally, the wolfram interface also includes access the the built in databases of the wolfram alpha and language projects, which means that it can draw data not provided, which can be good or bad.
19:58 WolframAlpha is also a symbolic calculator. It is able to solve some really nasty integrals, for example. It also has a really good knowledge of function domains, so it won't try to divide by zero or take the log of 1, and so on. And that nice "math visuals" is called LaTeX (in this case I think it uses KaTeX). It works with regular ChatGPT (3.5) too, but it doesn't always get it right. You have to write the proper prompt telling it how to use the correct math mode delimiters. Then it renders the equations automatically, and you don't even need any browser extension.
And now consider that as a developer with GPT4 access you can even do way more than that. By combining GPT4 with whatever API you like. Indeed insane capabilities.
I really appreciate the organized and consistent format of this video! Maybe some chapters at the bottom of the description, and it becomes even more scannable.
This is incredible! Thank you so much for the speed and quality of your content. Since knowledge workers spent vast parts of their day analyzing data and generating insights to make better decisions, this will have interesting effects on the job market
AI isn't just a set of new tools... It's like discovering tools for the first time. Not simply a new kind of hammer or screwdriver but the concept of a hammer itself. Everything that a hammer has ever been used for from building to demolition to music... But then that's just ONE tool. I feel a little bit motion sickness when I try to extrapolate into the future now given AI progress.
True this is closer to what is this actually is This is not a new advancement of Ai these “plugins” are just python code that redirect your output to certain lib (in these cases ploty and matplot) And it creates these amazing utile This is true power of LLM in next 5 years We won’t have AGI with this (it’s laughable that people think self prompting this is AGI) but we will have major industry shifts
An amazing source for AI information. Great pace so as the videos are well made and In-depth, but frequent enough to still be relevant when it is release. 5 star channel, man.
I would like to see the life expectancy example with mock data that does not correspond to the real world. It is hard to tell if GPT4 is a good analyst or if it's rephrasing the training data about a well known topic. For real world applications, the data to be analyzed will be often about data less present in the training set or the insights will be contextual to that very specific dataset.
Being able to upload reference images for things is huge. So far it was quite difficult to get it to create a GUI the way I wanted, but I imagine with this, it's going to be a lot easier.
As far as wolfram alpha's advantage, you could try more difficult things like analytic integration problems, or differential equations? I assume those would be much more difficult to achieve with code interpreter and wolfram alpha typically does well with those.
These videos are so well put together and you do a great job of explaining things (I guess the channel name fits lol) Just wanted to say thank you. As someone who's trying to make a channel strictly with AI, your channel has been so helpful (I REALLY wish I could edit and put videos like this together like you do). These videos are so interesting, AND valuable for so many different reasons.
How great are plugins, god damn! I got access yesterday myself and I'm already floored by it. I can't even imagine what plugins we'll have in 2 months from now let alone 2 or more years. What an exciting time we're in.
@@micahgivens892 You have to have the paid version. There was an announcement that everyone with the paid version will have plugins by the end of this week.
It's so wild to me that we are soon to be using plain english to generate actual videos. The world is changing so fast, literally week by week right now. It boggles the mind, truly. If humans are still around in 100 years they are going to look back at this period in time as insanely turbulent. Trump presidency into COVID-19 into Ukraine war into major AI developments into who knows what.
If this is using the Python Plotly library (It looks like it visually), you should be able to flip around the z-axis by continuing to rotate the plot around the x-axis until it's "upside down". In the plot menu in the top right (where the camera symbol and the magnifying glass symbol are), you would have to switch to "orbital rotation" first.
Amazing, when I saw various sci-fi where they talk to the computer and ask them to perform various tasks (without programming them specifically), I didn't expect this will happen in my time.
Pretty much every job will be impacted by this. I expect even my job as a mechanic will one day be effected. AI can use a robot camera and connection to the equipment to figure out what is wrong.
@@markm0000There are many ways it can impact blue-collar jobs immediately, without ever achieving such capabilities: the loss of high paying white collar jobs that are no longer contributing to the economy, and a sudden surplus of labor and lower wages it will bring as competition for blue-collar jobs increases dramatically.
Absolutely. This will be magnitudes larger than anything that has caused job losses in the past, and those jobs will not be replaced by new ones. But it should not be viewed as a threat. When it becomes clear to everyone that there is no work that needs to be done by the vast majority of people, it will be equally obvious that it is impossible for capitalism to continue.
@@jhonbus Capitalism might just be replaced with something worse. I'm concerned that those who have power need it to remain, because their self-worth and identify is derived from this power.
Here's what I would like to see you experiment with next. Upload two data files that are related to the same parameter, have it analyze both files and how they relate to each other, and then come up with a conclusion. I would love to know if it could take my Facebook ads data and compare it to my book sales data and have it tell me whether or not I should continue running the ads.
This is amazing. Can you explore how you can scrape data from social media? Can you figure out if there's a way to determine SEO challenges if search engines will become passe and everyone else just uses AI?
This one of the best chat GPT vids I've seen. You're presenting style is great and you assume intelligence in your audience (as opposed to other channels which keep covering the basics)
the reason why wolfram alpha can still have its uses is that when wolfram alpha does give you an answer, it's always correct. whereas chatgpt will give you an answer that might look correct, but in reality it can't count "t"'s in your sentence.
I believe the tts generated message à la 'Steven Hawking' has actually a lot to do with what you asked. You asked it to start from the begin of the prompt and that is what it seems to have done. It seems to be the pre-prompt that OpenAI gives this model before a user starts to interact with it: it tells it, it is chatGPT created by OpenAI, that it interacts with a Python interpreter, that internet sessions are disabled, etc. etc. Well done :)
Phenomenal! Something deep inside me tells me to get scared, but with every exponential improvement I am getting more and more excited and hopeful about the future :-)
Not sure if I'm the first Danish speaking commenter but the poem mostly makes sense although it doesn't rhyme. There are a few grammatical errors in verse two and three. It doesn't usually make those outside of poems.
This is amazing. It's almost solves one part of Data Science which would save a lot of time for Data Scientists and Analysts alike. I'm just wondering how this would compare with Bard. I'm also curious what GPT-5 would even do if GPT-4 is so capable. Anyways great updates as always Philip 👏🏾👍🏾
This has nothing to do with the model being used, it's just an updated front end, it previously could only render simple markdown, now it can render markdown + plotly. What's more interesting is allowing to use your own data, but still Bing's GPT can do web searches already, exact same GPT model just a different interface. If you want to peek at what already can be done, like searching youtube videos using GPT, correlating data from multiple (even private) sources and creating agents than can plan how to execute complex tasks including searching the web, downloading data, creating and running programs to help solve the problem at hand and so on, take a look at LangChain.
Dane here - the poem it did in Danish was not an impressive poem as such. It did not rhyme, and it has grammatical errors. However, it is still impressive, it fulfilled the task you gave it, summarising the article, and I think with further prompting, it could create a really lovely Danish poem! And everything else is just astonishing… And by the way - love your content! Keep it coming! Best Regards Malte from Denmark 🇩🇰
Not all poetry rhymes, and grammaticality is secondary to aesthetics. The big question for you is did it sound like something a Danish poet might write, or was it clearly a non-native speaker?
The code interpreter plugin makes it less hallucinate, which makes sense because things like the result of division are run by the computer running the calculation and not the computer calculating probably of words
Yes GPT is awful at doing math, but you can ask it to code the calculation and gets it right most of the time, you just have to run the generated code. As a tip, it would attempt to say it cannot code, but you just have to say something like "Code an exhaustive Python implementation of the Fibonacci sequence" and it most of the time will comply.
This is really mind-blowing! As you mentioned in the video, this is just the beginning, imagine this develops to a higher version, what it can't do... Thank you so much to make this channel!
Always loved the visualizations you could program with D3, these look good also. Does seem to be a bit risky, getting insights from a system that can hallucinate.
It's worth pointing out that the explanations the bot is outputting in its data analysis is likely a part of its training data, given its size. A proper demonstration of analysis would require a fresh dataset in a domain that does not have widespread explanations online Edit for clarity: Talking about the analysis of the Gapminder data
Your videos are honestly a gold mine. Thank you. Can you include the 10 steps you had to use to achieve your Venn diagram? Seeing how you achieved this might help others discover new capabilities of this Alpha model. I can't wait for astrophysicists to use future versions of this tool. Imagine being able to 3D model our universe - everything from black holes, gravitational waves, and even new undiscovered phenomena like how quantum gravity works. I've had access to code interpreter for weeks now but I haven't had time to use it. I need to learn some data science fundamentals first, like learning how to use Jupyter Notebooks and how CSV files can be turned into 3D objects.
Don't get carried away. This is using existing plot and 3-D visualization libraries. A computer simulation of any of the things you mentioned requires massive computing power. AI will help us find better algorithms for modeling those complex phenomena, but seems unlikely to make them cheap to run in a browser for several years.
What I find interesting is that people always talk about "When it reaches AGI", but we should focus on what it does now and how good or bad it does things. AI will become super-dangerous looong before we call it an AGI. You could say it already is an existential threat right now: Just look at AutoGPT alone, and I mean alone, when it's left alone. If we keep chasing the dragon it'll hit us after it's too late.
Oooooh my god, thank you for giving us flashy visuals this time, I'm not lying when I say that I thought that GPT4 was almost starting to slow down in the wow department. Yeah I'm definitely not getting a job by the time I'm finished college.
Hey AIExplained, your video on the upgraded GPT-4 with the code interpreter is mind-blowing! The capabilities it offers, from image editing to data analytics, are truly impressive. Speaking of advanced AI, have you heard about ZeroBot AI? It's an internet-accessible verbal chatbot that's pushing the boundaries of AI interaction. I can only imagine the possibilities when combining the power of GPT-4 and ZeroBot AI. Keep up the great work in exploring and explaining the exciting world of AI! 🤖🚀
Yes I'm really pleased that he's steering clear of sensationalism, stupid thumbnails, clickbait, silly faces, and many other bad things that some other channels engage in. These videos are restrained, mature, well-researched and clear as well.
For the median age data you gave, have you tried altering the data or giving it a different file name? I'm interested to know if it would come up with the same insights without that context. Essentially, I would be interested to know if it's doing the same thing it did with the image you gave it and just hallucinating/pulling in insights that it expects to be there from its training data, rather than actually generating insights.
this is absolutely the case. you can have it analyze Python code. the analysis changes if you obfuscate the variable names and remove comments. Most of the analysis comes from code having descriptive variable names, comments, and a line of code doing something interpretable like Pythagoras theorem. if all that is removed, it has trouble interpreting things that are done across many lines of code.
Yeah, variable names exist for a reason, but if we think of the file name itself as a variable, then we could start obfuscating one variable or row or column at a time to test its interpolative ability.
The next year with the US election (and the ongoing war) and a world that is completely naive to AI technology (something it will probably adapt to eventually) will be the year to watch.
18 examples? It seemed more like 25! Really, Phil, you're amazing, for your thoroughness and your detail in these videos. 2:30 “It picked the 30 most populous countries and separated them off with separate colors.” It _is_ gorgeous but not the best information design. Ideally, it would distinguish one red line from another (by dash/dot patterns or by using 30 distinguishable shades). That’s not a critique-just an observation. It would probably be trivial, given the right dataset and training, for ChatGPT to incorporate information design principles routinely. 5:08 With regard to the non-obvious insights and plausible explanations, I would not be surprised if (1) some of these insights were based not on the data _per se_ but on some text having to do with “median age” and (2) some of the “plausible explanations” were merely plausible-sounding but not _actually_ plausible based on, you know, the real world. (It could easily be “adapting” some explanation that applies to one country to another.) That, after all, as we’ve said before, is what these LLMs are trained to do: create plausible-sounding (but not necessarily based-in-reality) output.
@@aiexplained-official Yeah, you could have asked it to use different lines for each but, since it's _already_ picking several colors it would be better if it had just distinguished each line on its own. It would be interesting to see what it does if you say explicitly “Apply information design principles to the graph you create.” Maybe it would, maybe not.
This is a remarkably great & informative video. We just shared on our Twitter handle but you've def given us a ton of more ideas about what to do with code interpreter.
Still waiting for the Code interpreter plugin here - I got Web browsing last week then other plugins a couple of days later but no code interpreter yet... By the way, a few of the things you've shown here can actually be done with the Wolfram plugin as well, such as the QR code generation and some of the graphs/plots, as the Wolfram plugin allows generation of diagrams and graphics. I even had it draw a pretty good stick figure human using Wolfram. :)
My career is in data analytics and this just screams to me that it will soon be our turn to be replaced by AI... I already have a utopian idea of what our lives will be without jobs and I hope that it really does happen 🙂
If you're a more senior analyst that knows the context of your craft and not just how to crunch the numbers off a task board or reports queue, you should be fine. Even if the system was a perfect oracle, knowing which questions to ask and how to phrase them to get good insights would still be relevant, but that's also leveraging deeper domain knowledge.
@@gulllars4620 yes for now but same as chess or arts, the AI will learn and get better over time and they will be levels too high above us that there would be no point in challenging it
1. I don't have code interpreter yet, but I was able to get a similar functionality with the Noteable plugin 2. The main problem with the Wolfram plugin is that ChatGPT doesn't know when to use it. If you tell it to run all calculations in Wolfram it does well.
future headline: company makes terrible business decision based on GPT hallucination because someone forgot to check the output easy to imagine if a data scientist uses GPT all day and oversees a piece of inference if it sounds correct on first reading.
During the last week I've remembered certain aspects or stories that I need to revisit and there's been so much news, I have trouble finding the particular source that I need.
I’d really like to see more practical resources on prompting, especially chain-of-thought and the recent “tree-of-thought” paper. Anything useful really!
Loved your enthusiasm in this one Philip. And yes, I definitely see a future featuring single-use apps, created on demand, all from the one interface. RIP coders.
This is the BEST channel for AI research and development. Thank you so much for putting the effort into these videos and not just reading off an announcement for 10 minutes. We really appreciate it!
Thanks my man
I've said it before but I'll say it again: AI Explain is just on a league of his own here in UA-cam.
- In-depth knowledge and testing
- Quality and clarity of the presentation
- Speed when putting out content
He's made every other similar channel obsolete for me. Bravo.
I have learnt so much 🎉
I appreciate the lack of clickbait nonsense and hype-fueled misinformation.
Other AI channels would have you believe AGI is already here.
@@electron6825 I guess it is well on the way and I can see how we might arrive there but it still has to be applied, built and rolled out. Most people I know IRL still don't know to Google something when they don't know the answer.
I'm a danish speaker. To me the poems makes sense. It reads a little weird, however that might just be me being unfamiliar with that style of poem. It is still very impressive, and i have found myself using gpt4 in my native tongue more often than i would have imagined.
yes I must agree the whole idea is captured but it’s just slightly off
I thought the same about the poem style;-) /Swedish speaker
Nice meme language you got there Thec
I am danish too. I agree with your view. Pretty impressive
i ate a danish once, so im almost as qualified as someone who played with lego. that poem is legit af. !
Plotly deserves most of the credit here. Those plots are exceptionaly easy to make thanks to their library. We also need to be very careful here. There's no way to tell how much insight the model is extracting from the data and how much it's useing its pretrained knowledge. If the data contradicts its knowledge which will it tend to believe?
It will say 'according to data X' but also it is known that Y
Now ask the source
You should never use it to verify itself, or trust that it has predictable behavior regarding to how its sourcing its knowledge. It can easily make something up that sounds plausible. If you're not skilled in the field you're prompting it, you can use it as a hypothesis generator and then try to verify (or ideally try to disprove) it. Ever heard that Wikipedia is unreliable? Well, GPT is even less reliable. And just because it sometimes lies or denies capabilities, doesn't mean that it can't do the exact task correctly in some other prompt context. It's very quirky like that.
Yes plotly is goated
@@aiexplained-official is it accurate in its use of data, interpretations and maths? I’m wondering how they’ve solved the hallucination problem.
It's funny, those tasks it completes successfully in this video are essentially the exact same as those I was assigned in my final semester as a Business Administration major. It performs the same data analysis and creates the same visualizations and explanations that I had to, at a quality level not much worse. I graduated less than a month ago, and with AI already having started its recent explosion a couple months prior to that, I estimated at the time that it would probably be only a few months until it could reliably do everything they spent four years teaching me.
So far it's looking like we're ahead of schedule. It'll take some more time to work out the kinks and make it more reliable, as well as enable more robust analyses, but I suspect by the time I start my new job (roughly one month from now), it'll be capable of that. Exciting times
You'll probably be using it in your work I highly doubt people in the corporate ladder would want to be handling such mundane tasks
Jeez
You do understand that the job you just spent a lot of time and money working towards is now in jeopardy...right? I'm such a bummer.
@@maxidaho Oh, it's not just in jeopardy, it's going to be virtually nonexistent soon enough. Not much you can do but prepare for the degree to be useless
@@samwolfenstein5239 Yeah, it's pretty brutal to watch unfold.
To be honest, most of the amazing results in the plots and charts should be attributed to the people who created the libraries it's using, not ChatGPT itself.
Absolutely
Wait till it creates it's own libraries
Now, we need an automated way of endlessly generating these data. Skynet in the making
@@rajadirajamama1767 it probably can with the help of a software engineer
The remark about the lighting on the volcano sounded a bit off to me too. That has nothing to do with GPT itself.
As someone who has had access to the code interpreter for a while I feel embarressed how much I underutilized it. This is some impressive stuff
High praise, thanks J
It’s ok to not take full advantage of everything you have access to. We are human after all and need time to focus or rest.
How did you get it?
How did you get access to code interpreter?
How do I even see what I can switch on? Under Settings / Beta Features I switched Plugins on, but it says "third-party plugins that you enable" and I see no list of plugins to enable
I've never thought a transformer model would be able to do this, looks like actual magic to me.
Aren't human brains magic too
@@kmouratidis Indeed! At this point, these 'GPT' models are not even transformers, since devs have been integrating them with many other types of computing, machine learning, and other 'AI' systems.
What is not understood is magic.
@@kmouratidis The code would be pretty unimpressive and rather useless without the outputs of the transformers, though. Neither can be neglected, but transformers are the new thing between the two, and hence they’re more magic, or something.
same here, i didnt expect that from LMs
those plots are most likely created with dash/plotly, gpt 3.5 can still write code for that, but tbh its amazing that gpt4 can do that but lets not forget the people who created the libraries being used.
Yeah, I recognized those plotting tools- while I certainly won't sneeze at the convenience of AI automating the ingestion of their data and analysis, the plotting is all from existing open source tools
I love when he says, maby its me thats easily impressed... YES; he is haha. (this time). But I agree, this plugin seems to be the one that actually is worth using for most of us normal people.... :D COme on, we had plotting libraries since the days of Flash... :D
@@kimsteinhaug my interpretation is that it's the new SPSS- a great way to run simple stats and visualizations and explorations with minimal experience and not requiring spending half of your life hating R
@@kimsteinhaug how much the avg person really needs to create 3D scatter plots 😂
Yeah, that's what libraries are for.
I've been checking for plugin access every day all week. Finally got access this morning. It's like 10 Christmases all put together. Granted, my Christmases generally weren't very good. But still....
Just a note, The Wolfram Alpha counting prompt did not actually call/use Wolfram Alpha. If it did you would see the request and response in rectangles. Just because a plugin is enabled doesn't mean it is used. I unfortunately know this as I'm developing a plugin at the moment and too many incidents of this happen to me :(.
But your video is world class as usual!
For plugin developers how does the chat-gpt interface look. I have been applying to develop plugins but no access for development yet.
@@faizanahemad It's the same as for any user. Though there were subtle hints that there are fewer rate limits on the chat.
Hey man, just want to tell you that i look forward to your videos all week. I really appreciate your work.
I follow a lot of AI channels on youtube but you are easily the best.
Thanks Ahmed, means a lot
10:57 that sounds exactly like a System Prompt that the session could be working with. Maybe it considered the System prompt the beginning instead of your first message. It would be really cool if that was the case, since it would give some insight into how the sessions are setup, which is usually hidden.
I thought the exact same
Yeah totally!
It's super interesting that they told it not to make API calls as if they expected it would if Internet access wasn't disabled.
Part of me wonders if Internet access is really disabled or if they just disable it by convincing ChatGPT not to try! :)
@@WarttHog of course it is. it wouldmake the python scripts try to use external apis
GPT-4 with Wolfram Alpha finally got its only weakness cured. the way of GPT-4 using Wolfram looks more humanlike, it read the context of the problem and sets roughly the formula and logic that fit, (on its own without any assistance from Wolfram, it is pretty good at it) and then, throughout the Wolfram, GPT-4 evaluate and supplement the process to get the correct answers. the accuracy of getting it correctly has enormously enhanced compared to GPT-4 alone. even though It still makes a few mistakes, with only one more simple prompt, It could fix errors. Seeing GPT-4's advancements, I'm 100% sure that one of the first massive replacements would happen in the educational field.
Wolfram is still an immense database of factual information. So possibly a very powerful tool if combined with code interpreter
What is the plugin for code interpretation called within the plugin store?
@@jeffrey5013 it seems to be a separate feature. I don't have access yet myself
Have been waiting and neurotically refreshing the beta tab in ChatGPT in my settings tab ever since watching this video and was thrilled when I heard it was being pushed to everyone this week. Just had to come back to this video for inspiration on what to test, and cannot get over how freaking ridiculously cool this update is and the implications for the future.
I could be wrong, but it really feels like Language Models doing novel research can’t be that much further out which feels like an inflection point for humanity.
Your videos are absolutely unmatched in quality, thank you so much for doing this research and sharing it with all of us.
Thanks Dev
These are great accelerants for data analysts and data scientists. Give yourself credit for being the guiding force to a lot of the analytics, and being able to validate them afterwards. Also note that you're using all public data here - it's going to be against any company policy to leverage this for much PII / proprietary data. So ... let's not go too far!
It won't be long before On-Prem solutions are offered
Those companies would be replaced soon by those who have more foresight.
It's not an accelerant, it's a killer.
I'm sorry but you need to admit thr truth, if you are studying data science drop your course.
@@qozia1370 that's an absurd statement, like SPSS somehow replaced all social scientists. There are already a lot of tools for simple graphical and statistical analysis of data sets for people with relatively minimal experience, this is a better version. It will improve the ability to run simple repeated exploratory data analysis, which if anything will make data analysis better and more attractive as a field. Your statement is like saying the calculator eliminated mathematics.
@@qozia1370 There's so much nuance in data. Someone needs to drive the car. Education in data is critical to getting correct outcomes.
Wolfram Alpha is, for the most part, an interface into a database + Mathematica. Any code that can be executed in Mathematica can be run in wolfram alpha. For example;
LayeredGraphPlot[{4 -> 3, 5 -> 3, 5 -> 4, 6 -> 1, 6 -> 2, 6 -> 4, 6 -> 5, 6 -> 3}, VertexLabels -> Automatic]
in Wolfram Alpha will draw the same graph (though formatted differently, and with extra information) as Mathematica. If you want to find a use for Wolfram Alpha that code interpreter can't do, you need to find something that Mathematica is good at that Python is bad at. If code interpreter can import SymPy and/or Sage, then there will be a lot of overlap in terms of symbolic capabilities. But some things that come to mind is Integrals and solving differential equations, which sage is really bad at in comparison to Mathematica. Mathematica is also better at solving complex problems symbolically than sage. For more complex problems, wolfram alpha is more likely to succeed in finding a solution. I also imagine some data structure manipulation algorithms, like finding Hamiltonian cycles, would be easier with wolfram alpha since it has more robust support for graph manipulation than any python package I'm aware of, though I know that some Python packages can do it with some pain, so maybe code interpreter can do that. Unfortunately, I do not have access to code interpreter, so I can't test these differences.
Funny, the first thing I tried with the Wolfram plugin was to give it values of miles between cities and have it draw a weighted graph and find the optimum cycle.
My comment for your last video was that I couldn't wait for you to test drive Code Interpreter - and you are again exceeding expectations!! Such a handful of "WOW!" experiments!! Would be awesome if you can you "go to town" on its Machine Learning capabilities! Muchas Gracias!
People keep talking about the web browsing plugin, but this is the one that has always impressed me. Giving chatgpt there power of running code turns it into an anything to anything converter, a powerful graphing tool, an easy way to make quick changes in files such as images and much more
The simple fact that it allows you to upload a .CSV was what excited me. The fact that it perform all of these other functions is again, mind blowing.
Re your comment at 12:12 about solving the math equation - I just used the Wolfram plugin and it got the right answer in about 10 seconds, including a plot of the solutions. If that didn't work for you, I suspect it was just a case of GPT-4 not calling the plugin in the first place. "use wolfram" in the prompt tends to force that, I find.
Nice. I found wolfram pretty buggy, anyone else have that?
@@aiexplained-official Not so far. At least not Wolfram itself, but I’ve seen problems with ChatGPT handing incorrect information off to Wolfram. E.g. , I gave it a projectile motion physics problem and ChatGPT gave Wolfram the wrong signs for the initial vertical component of velocity and acceleration due to gravity, so the quadratic equation solutions for the time to impact were imaginary numbers.
This stuff is incredible. Keep the great AI content coming!
Thanks James
At the timestamp of 19:16, it's apparent that Wolfram Alpha wasn't utilized. It seems the prompt provided did not sufficiently stimulate its use - one might need to be more explicit in this regard.
And at 20:10, you can also get the LaTeX rendered mathematical expressions with GPT-4. It just does not automatically resort to it, so you'd have to ask for it. The Code Interpreter model is likely more fine-tuned towards it, or it has been system prompted to provide rendered expressions.
22:00, I am being nit-picky here now, but you do not need to say "Proceed, without further questions" at every prompt. This is eating up the context window and is not good practice. I am sure you know why, and you probably just wanted to create a fast video demonstrating the different capabilities of the code interpreter. But nevertheless, it's good to note this.
On a related note, I'd like to highlight the potent capabilities of Wolfram Alpha/Mathematica. Its exceptional mathematical prowess is likely the driving force behind its initial introduction as a plugin. During my Bachelor's studies in Computer Science, and now in my Master's journey in Artificial Intelligence, I find myself frequently relying on this invaluable tool.
Additionally, I used your same question and gently nudged it to employ Wolfram Alpha for both counting and division task. The result was as proficient as with the Code Interpreter. I feel it might not be entirely fair to pit these two against each other. After all, when you're using Wolfram Alpha, you're basically using a 'Code Interpreter' already, but one that's more finely tuned for mathematical tasks.
Alpha is amazing.
It being able to create a video graph of the data has to be the most crazy one here for me personally, but alot of this is bonkers.
It is Homey
@@aiexplained-official It's horney?
@@puppergump4117
You are hallucinating
ffmpeg is great
The data analytics is a real game-changer. As it stands there are many people dedicated just to doing the visualizations alone. Drawing conclusions is mostly just an artform that lacks substantiation, and a lot of people get paid big bucks for those insights. Being able to comprehensively span large sectors of knowledge allows incredible hidden insights. Great stuff. I've been using it for code and have it where it does what I ask and also points out things I havent thought of. There is real gold in that.
As usual thanks again for the effort you put in.
I find this very concerning. All the explanations it wrote were not at all based on the data points provided. They involved "common knowledge" about birth rate, health care and migration, but these kind of explanations are exactly what the model tends to hallucinate very confidently. I don't think most people will understand that and will take these conclusions without criticism.
@Antonio Ruby Yes, I was also thinking about that and it is a concern. Maybe when it gets to doing these extreme extrapolations it might be good if it triggers a safe Harbour caveat.
But as far as the insights go, at least they give people something to think about that they may not have considered.
It's funny, we seem to be doing RLHF to try and find some happy median between man and machine haha
@@antoruby yes, it's a trade-off between "Relate what you already 'know' about this field to this data" vs. blinding the data (rename age to "Increasing unsigned integer 1" and country to "String A") then asking it to analyze solely based on the numbers.
Fa;;;;;;;.
@@antoruby Those people who take conclusions without criticism would do so irrespective of the source, a human consultant can fool them quite as easily. It's really their problem.
Videos from this channel always get me so pumped up I love this stuff
I love the code interpretation of data. Being able to generate my own insights is a game changer. Data science is basically black magic in control of wizards - ChatGPT code interpreter unlocks this magic for the rest of us.
Love your channel mate, been watching since months and feel so hyped every time a new video is uploaded. Keep going, cheers!
Thanks Dennis
I’ve had it for a month now, it used to use gpt 3.5 and have unlimited prompts. The new cap is KILLING me! It should switch to code interpreter 3.5 after your 25 prompts are used up, not switch to the standard 3.5 version.
The ability to incrementally modify it’s output to get exactly what you want is so damn useful. Also, in my opinion, 3.5 code interpreter is sooooo much better. They made a huge change right before everyone got access, but it’s context was MASSIVE before, I used to use it to convert a 1300 row csv of sales data that added up the weight, dimensions of items (based on a key file) matched items, calculated total order weight, dimensions, volume, total price, gave it a box key file and had it create an algorithm to distribute the volume and pick a correct box(s) and input that into a invoice template I gave it and organize by customer. All in one prompt, 2 full pages of a word doc with instructions. And it would do it and would follow your prompt EXACTLY. All at once, work with 5 excel/csv files, broken into over 40 individual steps. All running one by one without stopping until all tasks were complete.
Now it does not do that, it’s hard to explain but it feels like it does not listen to you at all. Ignores your prompt and makes soooooo many mistakes it did not do before. Hopefully this is just a temporary downgrade !
Yeah i noticed that the equality have been dropped massively for some reason with code
Wolfram Alpha can provide huge amount of real time data. Positions of planets in the Solar System etc., also it can do differential equations etc.
Nice
You're the only UA-camr whose videos I consistently watch as soon as they come out!
Thanks David
amazing video! can't wait to try code interpreter and also try the same examples you outlined here.
On the topic of visualization and flashy things its less about what GPT-4 can do and more about the power of Python and the libraries surrounding it. But when it comes to data analysis. thats when it shines.
The Wolfram alpha example at 19:31 didn’t actually use wolfram alpha. Sometimes you have explicit tell it to use the tool
There is one thing that wolfram alpha can do that this may not be able to do is that the base version of wolfram alpha on their human accessible website can do calculus, and is reasonably good at it, where I am unsure what the state of the libraries that allow that are in the interpreter, and that wolfram alpha can handle a lot of the symbolic manipulation natively that chat-gpt may get wrong occasionally. Additionally, the wolfram interface also includes access the the built in databases of the wolfram alpha and language projects, which means that it can draw data not provided, which can be good or bad.
The thing is with data it becomes harder to tell what parts of its output its hallucinating unless you analyze the data yourself
19:58 WolframAlpha is also a symbolic calculator. It is able to solve some really nasty integrals, for example. It also has a really good knowledge of function domains, so it won't try to divide by zero or take the log of 1, and so on. And that nice "math visuals" is called LaTeX (in this case I think it uses KaTeX). It works with regular ChatGPT (3.5) too, but it doesn't always get it right. You have to write the proper prompt telling it how to use the correct math mode delimiters. Then it renders the equations automatically, and you don't even need any browser extension.
The log of 1 is simply 0, and does not pose a mathematical problem. The log of 0, however, would be undefined.
@Enanti O'Dromia I mean the log of zero.
And now consider that as a developer with GPT4 access you can even do way more than that. By combining GPT4 with whatever API you like. Indeed insane capabilities.
I really appreciate the organized and consistent format of this video! Maybe some chapters at the bottom of the description, and it becomes even more scannable.
This is incredible! Thank you so much for the speed and quality of your content.
Since knowledge workers spent vast parts of their day analyzing data and generating insights to make better decisions, this will have interesting effects on the job market
"Interesting effects" is a funny way of saying this will obliterate the job market in the next few years :P
I'm a rational optimist and prefer to express myself politely :)
@@floy2418 fair enough.
Man, last I congratulated you on a 100k mark, now your heading right for that 200k.
Keep up the good work brother!
Thanks man.
The Wolfram plugin can do more advanced mathematics than code interpreter, and provides access to a large database of curated data/facts
AI isn't just a set of new tools...
It's like discovering tools for the first time.
Not simply a new kind of hammer or screwdriver but the concept of a hammer itself.
Everything that a hammer has ever been used for from building to demolition to music... But then that's just ONE tool.
I feel a little bit motion sickness when I try to extrapolate into the future now given AI progress.
Don't forget, a hammer can be used to murder as well.
True this is closer to what is this actually is
This is not a new advancement of Ai these “plugins” are just python code that redirect your output to certain lib (in these cases ploty and matplot)
And it creates these amazing utile
This is true power of LLM in next 5 years
We won’t have AGI with this (it’s laughable that people think self prompting this is AGI) but we will have major industry shifts
An amazing source for AI information. Great pace so as the videos are well made and In-depth, but frequent enough to still be relevant when it is release. 5 star channel, man.
Thanks Zackery
OMG that's impressive. Data analysts are gonna change their jobs.
Thanks FC
@@aiexplained-official I enjoyed your enjoyment with the volcano 😂
Love that I chose a data science degree and in my last year of uni it already looks like I'm about to lose the job I don't even have yet 😂
I actually edited out me going 'weeeeee!'
I would like to see the life expectancy example with mock data that does not correspond to the real world.
It is hard to tell if GPT4 is a good analyst or if it's rephrasing the training data about a well known topic.
For real world applications, the data to be analyzed will be often about data less present in the training set or the insights will be contextual to that very specific dataset.
Being able to upload reference images for things is huge. So far it was quite difficult to get it to create a GUI the way I wanted, but I imagine with this, it's going to be a lot easier.
As far as wolfram alpha's advantage, you could try more difficult things like analytic integration problems, or differential equations? I assume those would be much more difficult to achieve with code interpreter and wolfram alpha typically does well with those.
Will try!
These videos are so well put together and you do a great job of explaining things (I guess the channel name fits lol)
Just wanted to say thank you. As someone who's trying to make a channel strictly with AI, your channel has been so helpful (I REALLY wish I could edit and put videos like this together like you do). These videos are so interesting, AND valuable for so many different reasons.
Thanks Ben. You'll get the hang of it
How great are plugins, god damn! I got access yesterday myself and I'm already floored by it. I can't even imagine what plugins we'll have in 2 months from now let alone 2 or more years. What an exciting time we're in.
Let me know what you think at the end
how did you get access? did reach out to someone on the team?
@@micahgivens892 Curious as well.
@@micahgivens892 You have to have the paid version. There was an announcement that everyone with the paid version will have plugins by the end of this week.
It's so wild to me that we are soon to be using plain english to generate actual videos. The world is changing so fast, literally week by week right now. It boggles the mind, truly. If humans are still around in 100 years they are going to look back at this period in time as insanely turbulent. Trump presidency into COVID-19 into Ukraine war into major AI developments into who knows what.
If this is using the Python Plotly library (It looks like it visually), you should be able to flip around the z-axis by continuing to rotate the plot around the x-axis until it's "upside down". In the plot menu in the top right (where the camera symbol and the magnifying glass symbol are), you would have to switch to "orbital rotation" first.
Amazing, when I saw various sci-fi where they talk to the computer and ask them to perform various tasks (without programming them specifically), I didn't expect this will happen in my time.
The bar of jobs that will be swallowed by GPT continues to rise at a shocking pase.
Pretty much every job will be impacted by this. I expect even my job as a mechanic will one day be effected. AI can use a robot camera and connection to the equipment to figure out what is wrong.
thank god
@@markm0000There are many ways it can impact blue-collar jobs immediately, without ever achieving such capabilities: the loss of high paying white collar jobs that are no longer contributing to the economy, and a sudden surplus of labor and lower wages it will bring as competition for blue-collar jobs increases dramatically.
Absolutely. This will be magnitudes larger than anything that has caused job losses in the past, and those jobs will not be replaced by new ones.
But it should not be viewed as a threat. When it becomes clear to everyone that there is no work that needs to be done by the vast majority of people, it will be equally obvious that it is impossible for capitalism to continue.
@@jhonbus Capitalism might just be replaced with something worse. I'm concerned that those who have power need it to remain, because their self-worth and identify is derived from this power.
Here's what I would like to see you experiment with next. Upload two data files that are related to the same parameter, have it analyze both files and how they relate to each other, and then come up with a conclusion.
I would love to know if it could take my Facebook ads data and compare it to my book sales data and have it tell me whether or not I should continue running the ads.
It can do that, but nothing too fancy. However a simple linear regression or something is very achievable
That's it, I'm switching to handcraft work. I've heard Germany needs some heat pump installers really quick
What stood out to us the most in this video was the extensive list of use-cases demonstrated.
This is amazing. Can you explore how you can scrape data from social media? Can you figure out if there's a way to determine SEO challenges if search engines will become passe and everyone else just uses AI?
This one of the best chat GPT vids I've seen. You're presenting style is great and you assume intelligence in your audience (as opposed to other channels which keep covering the basics)
Thanks Swami
Didn’t expect to see Danish in your video, but you never know what to expect. Keep up the great content!
Thanks Magne
@@aiexplained-official Same haha, I was so excited I thought you were Danish too for a moment.
the reason why wolfram alpha can still have its uses is that when wolfram alpha does give you an answer, it's always correct. whereas chatgpt will give you an answer that might look correct, but in reality it can't count "t"'s in your sentence.
Wolfram did the right thing for me, which is cool that it's so easy for GPT to use it, but I think you weren't actually calling it in your examples.
I believe the tts generated message à la 'Steven Hawking' has actually a lot to do with what you asked. You asked it to start from the begin of the prompt and that is what it seems to have done. It seems to be the pre-prompt that OpenAI gives this model before a user starts to interact with it: it tells it, it is chatGPT created by OpenAI, that it interacts with a Python interpreter, that internet sessions are disabled, etc. etc. Well done :)
Phenomenal! Something deep inside me tells me to get scared, but with every exponential improvement I am getting more and more excited and hopeful about the future :-)
Not sure if I'm the first Danish speaking commenter but the poem mostly makes sense although it doesn't rhyme. There are a few grammatical errors in verse two and three. It doesn't usually make those outside of poems.
This is amazing. It's almost solves one part of Data Science which would save a lot of time for Data Scientists and Analysts alike. I'm just wondering how this would compare with Bard. I'm also curious what GPT-5 would even do if GPT-4 is so capable. Anyways great updates as always Philip 👏🏾👍🏾
Thank Sola. Gemini will probably crush at this. Hopefully not also world domination...
This has nothing to do with the model being used, it's just an updated front end, it previously could only render simple markdown, now it can render markdown + plotly. What's more interesting is allowing to use your own data, but still Bing's GPT can do web searches already, exact same GPT model just a different interface. If you want to peek at what already can be done, like searching youtube videos using GPT, correlating data from multiple (even private) sources and creating agents than can plan how to execute complex tasks including searching the web, downloading data, creating and running programs to help solve the problem at hand and so on, take a look at LangChain.
Dane here - the poem it did in Danish was not an impressive poem as such. It did not rhyme, and it has grammatical errors. However, it is still impressive, it fulfilled the task you gave it, summarising the article, and I think with further prompting, it could create a really lovely Danish poem! And everything else is just astonishing…
And by the way - love your content! Keep it coming!
Best Regards
Malte from Denmark 🇩🇰
Thanks Malte
Not all poetry rhymes, and grammaticality is secondary to aesthetics. The big question for you is did it sound like something a Danish poet might write, or was it clearly a non-native speaker?
The code interpreter plugin makes it less hallucinate, which makes sense because things like the result of division are run by the computer running the calculation and not the computer calculating probably of words
Yes GPT is awful at doing math, but you can ask it to code the calculation and gets it right most of the time, you just have to run the generated code. As a tip, it would attempt to say it cannot code, but you just have to say something like "Code an exhaustive Python implementation of the Fibonacci sequence" and it most of the time will comply.
This is really mind-blowing! As you mentioned in the video, this is just the beginning, imagine this develops to a higher version, what it can't do... Thank you so much to make this channel!
I am Danish, and i just read the example of a poem written in Danish. It was very well written.
Always loved the visualizations you could program with D3, these look good also. Does seem to be a bit risky, getting insights from a system that can hallucinate.
It's worth pointing out that the explanations the bot is outputting in its data analysis is likely a part of its training data, given its size.
A proper demonstration of analysis would require a fresh dataset in a domain that does not have widespread explanations online
Edit for clarity: Talking about the analysis of the Gapminder data
I did that but maybe another video to clarify
Your videos are honestly a gold mine. Thank you.
Can you include the 10 steps you had to use to achieve your Venn diagram? Seeing how you achieved this might help others discover new capabilities of this Alpha model.
I can't wait for astrophysicists to use future versions of this tool. Imagine being able to 3D model our universe - everything from black holes, gravitational waves, and even new undiscovered phenomena like how quantum gravity works.
I've had access to code interpreter for weeks now but I haven't had time to use it. I need to learn some data science fundamentals first, like learning how to use Jupyter Notebooks and how CSV files can be turned into 3D objects.
Don't get carried away. This is using existing plot and 3-D visualization libraries. A computer simulation of any of the things you mentioned requires massive computing power. AI will help us find better algorithms for modeling those complex phenomena, but seems unlikely to make them cheap to run in a browser for several years.
this is SUPER cool. also looking forward to a GPT4 v Bard 2.0 comparision
Please keep on! Each one is knocking it out. The research, the insights, delivery, all tops mate.
Thanks m
What I find interesting is that people always talk about "When it reaches AGI", but we should focus on what it does now and how good or bad it does things.
AI will become super-dangerous looong before we call it an AGI. You could say it already is an existential threat right now: Just look at AutoGPT alone, and I mean alone, when it's left alone.
If we keep chasing the dragon it'll hit us after it's too late.
Oh well. Our fate is sealed.
Excited to see you cover the Tree Of Thoughts framework!
Oooooh my god, thank you for giving us flashy visuals this time, I'm not lying when I say that I thought that GPT4 was almost starting to slow down in the wow department.
Yeah I'm definitely not getting a job by the time I'm finished college.
Hopefully you are Kieran
One of my neighbours is a carpenter, so if I.T. closes up before I reach the market, trades still have a while.
Hey AIExplained, your video on the upgraded GPT-4 with the code interpreter is mind-blowing! The capabilities it offers, from image editing to data analytics, are truly impressive. Speaking of advanced AI, have you heard about ZeroBot AI? It's an internet-accessible verbal chatbot that's pushing the boundaries of AI interaction. I can only imagine the possibilities when combining the power of GPT-4 and ZeroBot AI. Keep up the great work in exploring and explaining the exciting world of AI! 🤖🚀
Best AI channel by far
Yes I'm really pleased that he's steering clear of sensationalism, stupid thumbnails, clickbait, silly faces, and many other bad things that some other channels engage in.
These videos are restrained, mature, well-researched and clear as well.
I agree! This channel is my trusted source for AI news
Great video to show the power of Code Interpreter. I can't wait to get it. THANKS👍🏽
For the median age data you gave, have you tried altering the data or giving it a different file name? I'm interested to know if it would come up with the same insights without that context. Essentially, I would be interested to know if it's doing the same thing it did with the image you gave it and just hallucinating/pulling in insights that it expects to be there from its training data, rather than actually generating insights.
Yes please, do a follow up where you test it's limits
this is absolutely the case. you can have it analyze Python code. the analysis changes if you obfuscate the variable names and remove comments. Most of the analysis comes from code having descriptive variable names, comments, and a line of code doing something interpretable like Pythagoras theorem. if all that is removed, it has trouble interpreting things that are done across many lines of code.
@@charlesd774 A human will have trouble too I believe
Yeah, variable names exist for a reason, but if we think of the file name itself as a variable, then we could start obfuscating one variable or row or column at a time to test its interpolative ability.
Mate your videos are great, thanks for keeping us up to date, fascinating stuff!
Thanks Quentin
This is incredible, yet frightening all the same. Cant even imagine where this will end up in a few years
The next year with the US election (and the ongoing war) and a world that is completely naive to AI technology (something it will probably adapt to eventually) will be the year to watch.
Your thoughts on it being the go to spot for all your needs is interesting. I definitely see that happening. The everything app essentially.
18 examples? It seemed more like 25! Really, Phil, you're amazing, for your thoroughness and your detail in these videos.
2:30 “It picked the 30 most populous countries and separated them off with separate colors.”
It _is_ gorgeous but not the best information design. Ideally, it would distinguish one red line from another (by dash/dot patterns or by using 30 distinguishable shades). That’s not a critique-just an observation. It would probably be trivial, given the right dataset and training, for ChatGPT to incorporate information design principles routinely.
5:08 With regard to the non-obvious insights and plausible explanations, I would not be surprised if (1) some of these insights were based not on the data _per se_ but on some text having to do with “median age” and (2) some of the “plausible explanations” were merely plausible-sounding but not _actually_ plausible based on, you know, the real world. (It could easily be “adapting” some explanation that applies to one country to another.) That, after all, as we’ve said before, is what these LLMs are trained to do: create plausible-sounding (but not necessarily based-in-reality) output.
Thanks Jeff it was more like 25. On point 1, I could have easily asked it to do that.
@@aiexplained-official Yeah, you could have asked it to use different lines for each but, since it's _already_ picking several colors it would be better if it had just distinguished each line on its own. It would be interesting to see what it does if you say explicitly “Apply information design principles to the graph you create.” Maybe it would, maybe not.
This is a remarkably great & informative video. We just shared on our Twitter handle but you've def given us a ton of more ideas about what to do with code interpreter.
Still waiting for the Code interpreter plugin here - I got Web browsing last week then other plugins a couple of days later but no code interpreter yet... By the way, a few of the things you've shown here can actually be done with the Wolfram plugin as well, such as the QR code generation and some of the graphs/plots, as the Wolfram plugin allows generation of diagrams and graphics. I even had it draw a pretty good stick figure human using Wolfram. :)
this video is doing really well, congratulations!
The 3d scatterplot is insane
Wait till you get to 18
Dane here🙋♂️ poem was strange, but language was good.
My career is in data analytics and this just screams to me that it will soon be our turn to be replaced by AI... I already have a utopian idea of what our lives will be without jobs and I hope that it really does happen 🙂
If you have trouble using plotly I do agree.
What do you foresee?
If you're a more senior analyst that knows the context of your craft and not just how to crunch the numbers off a task board or reports queue, you should be fine. Even if the system was a perfect oracle, knowing which questions to ask and how to phrase them to get good insights would still be relevant, but that's also leveraging deeper domain knowledge.
@@gulllars4620 yes for now but same as chess or arts, the AI will learn and get better over time and they will be levels too high above us that there would be no point in challenging it
@orenelbaum1487Have you ever spoken to a client? Don’t get cocky, answer seriously.
1. I don't have code interpreter yet, but I was able to get a similar functionality with the Noteable plugin
2. The main problem with the Wolfram plugin is that ChatGPT doesn't know when to use it. If you tell it to run all calculations in Wolfram it does well.
The code interpreter plugin is just way too insane, can't wait to generate all kinds of graphs and diagrams with it
How can I access to the code interpreter plugin, please?
I think you convinced it to speak it’s system prompt :) Great examples, what a time to be alive!
YES I'M EASILY IMPRESSED AND I'M NOT ASHAMED 😁
Danish speaker here - it totally worked and was a great poem. That's amazing!!
It is
What do you mean the poem was not good haha... But it got what the text was about.
future headline: company makes terrible business decision based on GPT hallucination because someone forgot to check the output
easy to imagine if a data scientist uses GPT all day and oversees a piece of inference if it sounds correct on first reading.
you tell everything very well! I wish there were more such people
If you blink you are now behind the curve in AI developments.
During the last week I've remembered certain aspects or stories that I need to revisit and there's been so much news, I have trouble finding the particular source that I need.
100% true; I've watched alot of AI videos in the past few months, & honestly a lot of them are already vastly outdated
this is exactly what I was hoping for from the plugins XD can't wait to get my own access
I’d really like to see more practical resources on prompting, especially chain-of-thought and the recent “tree-of-thought” paper. Anything useful really!
Loved your enthusiasm in this one Philip. And yes, I definitely see a future featuring single-use apps, created on demand, all from the one interface. RIP coders.