“Hey, Copilot, I’m late to the meeting. Did anyone mention me?” “Yes, John said ‘Surprise! Jeremy is late again. Man, that guy needs to go.’ and Karen said ‘Keep this between us, but Sue in HR has already got the paperwork ready and they are just waiting on him to finish his current task so we don’t get screwed over.’”
Goddamn this sounds familiar. The plight of Jeremys everywhere, I suspect. It's not our fault our time keeping is shot - YOU try going through school with a name like this and then see how keen YOU are to show up at the same time as the rest of the group....
Thanks Luke, great info as usual. I'd be so dissapointed if I had forked out $20 for CoPilot and it wasn't able to analyse my data and actually cost me time rather than saved me time. It's going to be great - just not yet. Very glad to see it will be available for Mac though.
6:12 the reason why it didn’t generate a response, it’s likely due to you overflowing its context window. For LLMs such as GPT-4, the context window is similar to a computer’s RAM -in the sense that they act similarly-. This context window is measured in “tokens”. GPT-4 has 8K, 32K and 128K tokens. Not sure which model you’re using with co-pilot but 8K is about 20 pages of text, 32K is about 75 pages of text and 128K is about 300 pages of text. That’s for one whole interaction. And the models at any given case can only generate about 4K tokens worth of a response, that also counts against your context window.
I wonder if we can put to a test with a large messy data set. Can Copilot clean a data set faster than you could clean it in something like R or Python? If the inputs to Copilot are short and concise, then cleaning data has just become way easier and would cut 80% of our workload in half.
@@jazzmann1984 I agree, I think co-pilot should be faster, if you're writing the scripts from scratch. BUT! It requires both, prompt engineering and LLMs knowledge and best practices.
it wouldn't send the data in the context. It requires it being a table so it can send the table metadata and a sample (a few rows), then the 'calculations' per say are done via pivot tables
Changed forever? Not quite. I’ve been using it for months and I can assure you it is completely overhyped. I’m sure it’ll improve over time, but has a way to go.
@@rp26101 it depends on what type of work you do. The more technical your work is the more valuable it can be for you. If you don't do technical work you won't see it as valuable
Interesting insights on co-pilot! The potential for automating tasks in Microsoft apps is promising. However, the data limitations in Excel and the need for cloud storage are notable drawbacks. Looking forward to seeing how co-pilot evolves in future versions. 👍
Good to have confirmation it's not just me that found the excel functionality very limited. It seems like a dream come true, on paper, but then I found I could get whatever insights I want faster manually most of the time.
I mean, you have a data with 40k rows you might as well use Python or R to make your analysis. Excel and any sheets app is only capable of performing well until certain point.
It has pretty long way to go annnndddd that's the amount of time we data nerds have to established ourselves coz once the copilot is mature enough with no surprise there will be job cut for these small small works.
I'm inclined to wait for the early adopters to "train" this into a pretty rigid mode of perceiving and communicating the data world. And then I'll get better at what it fails at :-)
@@vuyilemagwaza3236 It does mean you will have to be good at writing algorithms for macros & to be skilled enough to spot errors/flaws made by AI programs. This is a huge step up from the accepted skill set & knowledge base of most employees.
9:36 the data issues is context window overflow due to bad prompting. You can use huge data sets as long as you indicate in your prompt what the AI should be looking for. If you tell it “column H and plot it against Column A” likely you’ll get a response. Also, to improve the output of these models, you need to improve your input. Your prompts seem vague and poor, gotta refine those to get better answers. Hit me up if you need help with this, I can help you navigate through it.
On this, it needs refinement, what type of product has most of its user use it "the wrong way", when natural language is supposed to be intuitive, I know that this is just the beginning but I can assure you that within a couple of years, prompt engineering won't be a thing anymore
@@steveloco1170 I disagree. Natural Language is not intuitive, it is natural to us, but it plague with vagueness, reason why our laws need judges and lawyers to interpret/discuss them. I agree that for high level tasks, such as searches online, prompt engineering will not be as necessary as it will still be for tasks such as processing vast amount of data. The main reason is due to restrictions on context window. I don't think a chatbot will be able to digest huge datasets to later understand vague prompts from users, because users that typically deal with these type of datasets are skilled individuals like myself. Thus, it wouldn't be economically viable to spend time and resources in creating such product that scientist and engineers are able to circumvent easily with proper prompt engineering practices. I agree that this models will get more capable over time, but what you suggest is technically a model with a context window with orders of magnitude greater than 300 pages worth of text (>128K tokens), although powerful models will come, datasets are much more bigger than what these models will accomplish any time in the upcoming 5 years. Of course, this is an educated guess, and I could be wrong. I still remain skeptical. And today, this product is not working as it should for this guy because of poor prompt engineering practices; adding the lack of knowledge of the technology behind.
According to my subscription information, I have copilot but it's not available in any of the programs - be it the browser version or in the desktop version. That being said, with regards to the limitations on data size, I'm not sure that needs to be a big hurdle. I'm 99% confident I'd get into trouble if I used it on the data we have at work (even with the stipulations about data protection etc.), so I'd be using co-pilot on dummy data anyway, and using it to aid in quickly getting good visualisations out of data that REALLY should not be in a spreadsheet to begin with (but in a database).
Excel remains quite capable of sinking even good computers in terms of compute. The idea that people will offset gigantic data sets, and I'm kind of stupidly assuming there is some processing basis like a cell is a token (or something akin to that) - so if you have 100,000 x 10 row sheet, its a way larger issue than what is normally pushed to AI. For small business, where its perhaps small sheets, this might work out well in the present. I've got my reservations about larger data, and what I see in the enterprise where you get really crazy linking across massive, massive sheets. Its in these really large areas where AI would probably give the best return, but equally IMHO the heaviest AI workloads.
It’s amazing how easy it was to find this information without copilot. I wonder if copilot is going to prevent us from actually looking up negative reviews of AI.
You'd be surprised. Now granted it probably isn't worth that in its current state. But companies are regularly dropping several hundred dollars a month on subscription services per member. Some of them get real expensive.
The new word has image creation capabilities.. will those be included in the copilot subscription or are there other ways to create images? I use image creation a lot to do sketches of landscapes and modify them at work
It’s not at all surprising that the data size is limited. It’s using GPT-4 under the hood, it’s almost certainly converting it to a CSV and just passing it to a fine tuned GPT-4 model. If I remember correctly, the newest GPT-4 can handle 32k tokens, with a token generally being 3/4ths of a word. Now, I haven’t checked the embeddings or anything, but I suspect that it treats most numbers as 1 token per number, or perhaps 2 number pairs. Something like that. Anything with a decimal point is going to get its own token for that (or maybe paired with a single number). So I’d guess a number like 23,462.56 would eat up something like 5 tokens in the best case and 9 in the worst. If it’s 5 tokens. I didn’t check exactly what your lines were, but one of them was decently long JSON. So, let’s say you have 4 cells with numbers in it, 1 cell with a label (let’s call that 2 tokens), and 1 cell with JSON (let’s say that’s about 50 tokens). That’d be 72 tokens per line. Let’s round down to 64 (for easy math and to be generous). That’d give you 512 lines at most that it could keep in context. Hey, didn’t you say you could only do about 500 lines? Well, there you go. That’s why. Additionally, LLMs tend to have issues “remembering” things from the earlier parts (except the very beginning) of a message when generating the next token. They seem to remember the very beginning, and then the latest X portion of the conversation, and if the conversation exceeds X, even if it’s within the context length, you’re going to get worse answers if they needed to reference information in the portion between begin + C and end - X (where C is a very small number and X is a number that’s a very significant fraction of the total context length, like maybe 10 and 5000 or so.) We don’t have unlimited context length and not all data is created equal in terms of token efficiency. Excel data is probably among the least efficient, whereas word is likely extremely efficient, thus, you can get a lot more out of it in word than excel. (Plus, excel documents tend to be very large more often than word files, and way more often than PowerPoint). If we can get longer context lengths in the future, this will be less of an issue. Also, to some degree, this is using a hammer to put a screw into wood. LLMs are not calculators. They make terrible calculators. This is going to be a weakness. It must be using code in the background to do the calculations, because the LLM by itself would fail nearly every time.
Hi, does identify if there are duplicated numbers? Example if you have a list of Orders and Orderlines. Does it sum up correct and identify that it needs a unique count?
Great video Luke! 00:00: Introduction to Microsoft's Co-pilot, an AI assistant powered by OpenAI. 00:18: Overview of Microsoft's keynote event and the announcement of Co-pilot. 00:36: Excel's capabilities with Co-pilot for ad hoc analysis and data visualization. 00:55: Using Co-pilot in Word for planning and incorporating images. 01:09: Creating PowerPoint presentations with Co-pilot using Word documents. 01:20: Drafting and summarizing emails with Co-pilot in Outlook. 01:47: Utilizing Co-pilot in Teams for meeting summaries and follow-ups. 02:00: The potential of Co-pilot to automate tasks and delegate work. 02:21: Introduction to Co-pilot Pro and its features. 03:13: Co-pilot for Microsoft 365 and its additional features for organizations. 03:52: Customization options in Co-pilot Pro and Co-pilot Studio. 04:25: Hands-on testing of Co-pilot in Microsoft Excel. 05:52: Testing Co-pilot with larger and smaller data sets in Excel. 07:14: Using Co-pilot in Word and PowerPoint for creating essays and presentations. 08:05: Final thoughts on the performance and limitations of Co-pilot in Excel.
Do you think that this is going to replace a lot of data analyst jobs, I just started learning and going through the google analytics certificate, but it seems that copilot just may do an entire data analyst job, not now of course but in the future.
No. Even if Co-Pilot could analyze and transform data way better than it can now, that wouldn't replace the core functionality of a skilled data analyst. Which is being able to work with people to make shitty, unorganized data more understandable and organized, and also creating platforms to distribute your findings and determine how data collection and entry is done. Generative AI is a useful tool, but we're reaching a ceiling to how much these machine learning models can actually respond based on large amounts of data and code. We're struggling to make learning models that can function well past a 128k token limit. Even then it's really pushing it. Remember, these machine learning models don't actually think, they just regurgitate information. Literally better-Google. To be useful in any tech field you need actual critical thinking skills, especially data analysis. Computers don't think. They just do.
@@joshuaduplaa9033 what do you think about data entry? I think if copilot doesn't take over data analyst jobs, data entry jobs will definitely be replaced
Not even close. If you're a data analyst worth your salt, you will do 100x better job than any AI will do for at least 5 years, probably more like 15 depending how the pace of AI research goes. Irrespectively, most of those tools are unlikely to be open to Excel users anyway.
Luke, with the announcement of the Copilot+ Laptops that run the Snapdragon X Elite chip with a high end NPU I'm wondering what the performance would be on that test which didn't perform well on your VM. Knowing that Microsoft typically releases software updates that will take advantage of new hardware features before that hardware is even announced, I'm going to assume those new laptops will smoke that test now. Also. given VM support for NPU is still early days, odds are you will have to run on physical machine to see any NPU goodness.
Great; a much-needed video - thanks! The lack of data security with the pro version is a dealbreaker for me. Shame, because it looks useful. The reason for the limit in the number of records that the AI could handle is almost certainly because the non-business version has a limit on the number of tokens that can be submitted, versus the business version - same with the $20 paid version of ChatGPT vs the enterprise version of ChatGPT
I haven't used copilot yet, and not sure if I can use copilot to its max potential, but I can already see a lot of features that really take off a lot of stress at work (I'm actually rooting for it because of the meetings summary feature, that would help a lot....)
Windows 12 is currently rumoured to require 16GB minimum for system RAM, alongside the requirement of a 15th gen INTEL CPU which will come with an embedded NPU for using AI features such as Copilot natively on your system and in the windows shell.
CP also can't analyse several excel docs at the same time or several sheets in the same workbook - never mind a large single sheet with thousands of rows. Very, very limited. Copilot in Excel honestly seems (right now) like it might as well just have been made into an improved pivottable/chart wizard.
Ask it to label the horizonal axis of the histogram with the midpoint of the bins rather than the range… i spent 20 min the other day trying to do this and couldnt figure that out. Curious if copilot can do it.
Great vid again, so nice to see the latest news explained like this, this way you know what to subscribe to or better wait a while (or forget about it) but hoping copilot is gonna work because of it’s MS connection
Do you know if i purchase the pro or business plan will i be able to use it across my personal and two business accounts in Microsoft 365. Or is the best way to just use it on my personal account and import data from the corporate accounts would rather not do this.
Hi Luke, one question please. May I know if using Excel for personal plan would be enough for data analyst. Or do you recommend the business standard. Thank you
In the short term, tools like copilot on office look cool, however, in the long run, copilot will hamper employees ability to perform critical thinking, do out side of box thinking and problem solving, leading to cookie cutter solution to the most important business problems since those solutions will be built on top of ai suggested started point.
I tried it by the way and right now Copilot is saying one can not edit the pivot tables, or charts made from the pivot tables, so you have to ask it to edit the original data set ( Excel table you created in the beginning). Copilot implies it one may have the ability in the future to edit the pivot tables and charts.
Signed up for it then cancelled when I found that even with it enabled it was not showing up in Excel, Word, or Outlook. Even Copilot Pro itself couldn't tell me what was wrong. I gave up after twenty minutes of searching. Too bad I didn't see your video yesterday! But in any case, it seems like waiting until version 2 might not be a bad idea.
Sounds like (for now) you'll still likely need to do your own data analysis and insights work, but Copilot can at least save you time building out the deck and white paper.
Just had my first conversation with co-pilot. It feels just like talking to Microsoft. Which to say I'm now really frustrated and re-considering my life choices. How do they ruin everything they touch??!
Copilot can barely respond to simple questions let alone do any of this. It just stops working. This is currently the problem: these apps aren't reliable or trustworthy, so they actually make everything take longer if you care about your work.
+1 to the comments about simple cleaning up data. I'd like to know if you can dump in a bunch of things like name,address,email,phone on a single column and have it break those into separate columns without error. In simple data, that is fairly easy, but when some names have three words, rather than two, and the names are space-delineated, but the addresses are comma-delineated, it gets painfully time-consuming to do a simple task. I might have to get a single license for myself, to try the things I ask my office manager to do and see if it can help her with this annoying stuff.
I work for a 5000 it-worker company. So co-pilot would run us 1.800.000$ a year? On top of the normal licenses from MS. Yeah that seems like such a good idea.
M365 Chat and Teams is useful. Also the MacOS apps have copilot as well. I run Parallels too but outside running it in edge. Or using edge browser on my iPhone, the Mac Apps have functioned the same.
Wait so, I expect that if I tell copilot "Can you write every 50 cells a series of text starting with Lot 1 in cell A2 so that Lot 2 is in A52, Lot 3 in A102 and so on and so forth" it will correctly help me sort my lots better. No chance?
Great video Luke! I heard you have a MacBook Pro in your video. I have one as well and run into issues getting PyCharm working correctly with selenium webdriver. Do you know of any places I can look for getting a MacBook Pro set up for data science tasks? Thanks a bunch!!
Copilot presentations at next year's Microsoft keynote gonna be: "Let me demo to you how I sleep on this hammock while Copilot earns my six-figure salary."
The problem with uploading files to OneDrive or a SharePoint server under Microsoft's control is that it violates contract clauses which restrict disclosure of intellectual property or other confidential data to third parties. Given that Microsoft employs foreign nationals, uploading certain types of data (much of which is commercial) can result in a violation of U.S. export control laws and regulations (which carry penalties of up to 20 years imprisonment and a $1,000,000 fine).
Not sure what version of excel and copilot your using (maybe a future one that will change excel forever...) but been using it at work and its very weak
I tried using Excel to chart my spending in different categories based on my credit card history ( about 200 rows). It kinda worked but also couldn't quite finish it. It would not ignore one category of data like I asked it to. It was making charts for me many times but kept including the category I wanted to exclude. Then after about 15 attempts, the response was " I am a large language model and can not make visual representations" >.... HUH? It was making charts , but then stopped. Is that Copilot telling me there is a limit to my attempts or do I need to get out my tin hat? Overall it works nicely, and I assume it and I will improve over time.
That's the future! It's really interesting how AI has changed the world just in the last year... I remember how everything started with the hype around Dall-e at the end of 2022, then Chat-GPT and GPT-3 were released to the public. From there it only took a couple weeks until everyone was using them. GPT-4 hit the market really hard and literally changed everything. Now after literally only one year with these large language models out everyone and even my grandmother know of theses AIs. To think that I can sit down and talk to my 82 year old not in the remotest techsavy grandmother about what AI has done to our world in the last year alone is astonishing. To think that my grandparents literally grew up in a small village in the middle of nowhere with literlly no electricty and a donkey as their main vehicle is really putting stuff in the frame. For them they went from middle ages to scifi in a whole lifetime. To think that in their lifetime they started without the concept of computers let alone microchips and to get to a point where you can literally have a computer talk to you is making me excited at to where the world will be when I am their age.
Great video which yeah even enhanced search I mean Copilot on Bing/Edge still has issues and needs to be lead sometimes so I'm not surprised at it's ProFormance in this video that said... I normally use word for writing, notes and gamedev and powerpoint for timelines and excel once for a card game... So yeah it sounds good especly for $20 a month and I get some bonuses to the other Open ai bots to.
“Hey, Copilot, I’m late to the meeting. Did anyone mention me?”
“Yes, John said ‘Surprise! Jeremy is late again. Man, that guy needs to go.’ and Karen said ‘Keep this between us, but Sue in HR has already got the paperwork ready and they are just waiting on him to finish his current task so we don’t get screwed over.’”
Lmaoo
Already experienced this with zooms version. Gonna be a long, sometimes painful learning curve.
😂 😂😂😂😂
😂😂
Goddamn this sounds familiar. The plight of Jeremys everywhere, I suspect.
It's not our fault our time keeping is shot - YOU try going through school with a name like this and then see how keen YOU are to show up at the same time as the rest of the group....
"Excel has forever changed"
Video : "i'm disappointed it doesn't work"
Thanks Luke, great info as usual. I'd be so dissapointed if I had forked out $20 for CoPilot and it wasn't able to analyse my data and actually cost me time rather than saved me time. It's going to be great - just not yet. Very glad to see it will be available for Mac though.
Yep, keeping my eye on it so I’ll let yall know when it does get there!
I had the same experience with Chat GPT4. I have the impression like there is still way to go before they are actually useful for data analysis.
It's not going to be great. It will continue to suck, as the technology does not do anything useful and fundamentally cannot.
6:12 the reason why it didn’t generate a response, it’s likely due to you overflowing its context window. For LLMs such as GPT-4, the context window is similar to a computer’s RAM -in the sense that they act similarly-. This context window is measured in “tokens”. GPT-4 has 8K, 32K and 128K tokens. Not sure which model you’re using with co-pilot but 8K is about 20 pages of text, 32K is about 75 pages of text and 128K is about 300 pages of text. That’s for one whole interaction. And the models at any given case can only generate about 4K tokens worth of a response, that also counts against your context window.
I wonder if we can put to a test with a large messy data set. Can Copilot clean a data set faster than you could clean it in something like R or Python? If the inputs to Copilot are short and concise, then cleaning data has just become way easier and would cut 80% of our workload in half.
@@jazzmann1984 I agree, I think co-pilot should be faster, if you're writing the scripts from scratch. BUT! It requires both, prompt engineering and LLMs knowledge and best practices.
it wouldn't send the data in the context. It requires it being a table so it can send the table metadata and a sample (a few rows), then the 'calculations' per say are done via pivot tables
I’ve had the same experience using Copilot in Excel. Copilot in Teams meetings is probably the most capable at this time.
Changed forever? Not quite. I’ve been using it for months and I can assure you it is completely overhyped. I’m sure it’ll improve over time, but has a way to go.
AI influencers tend to overhype themselves
It's just clickbait. You are new to this
It's not over hyped. You have no idea what this thing can really do!
@@bibebibebibe , give me some examples. I’ve been using it and don’t see the value.
@@rp26101 it depends on what type of work you do. The more technical your work is the more valuable it can be for you. If you don't do technical work you won't see it as valuable
Interesting insights on co-pilot! The potential for automating tasks in Microsoft apps is promising. However, the data limitations in Excel and the need for cloud storage are notable drawbacks. Looking forward to seeing how co-pilot evolves in future versions. 👍
Good to have confirmation it's not just me that found the excel functionality very limited. It seems like a dream come true, on paper, but then I found I could get whatever insights I want faster manually most of the time.
I mean, you have a data with 40k rows you might as well use Python or R to make your analysis. Excel and any sheets app is only capable of performing well until certain point.
Great video as usual - all my questions answered without having to spend hours playing around myself.
Glad I could help David 🙌🏼
That was an AWESOME video, practical, engaging with great real world data and humor, thanks man! !!
The first half of the video provides a good overview of the functionality Copilot provides across the various Microsoft apps.
It has pretty long way to go annnndddd that's the amount of time we data nerds have to established ourselves coz once the copilot is mature enough with no surprise there will be job cut for these small small works.
I was just thinking about that, we can use this time to prep ourselves and upskill.
@@LemonadeRock haha. Thanks for worrying about me but i am a machine learning engineer and had been in service from past 6 yrs😂
I'm inclined to wait for the early adopters to "train" this into a pretty rigid mode of perceiving and communicating the data world. And then I'll get better at what it fails at :-)
all the data still has to be standardized for this to work...maybe in 10 years we can start worrying
@@vuyilemagwaza3236 It does mean you will have to be good at writing algorithms for macros & to be skilled enough to spot errors/flaws made by AI programs. This is a huge step up from the accepted skill set & knowledge base of most employees.
This vid had me cracking up. Thank you for this summary.
Just like GTP4, with a lot of absurd limitations!
Great video!
Thank you for this, not enough videos showing capabilities and drawbacks out there.
9:36 the data issues is context window overflow due to bad prompting. You can use huge data sets as long as you indicate in your prompt what the AI should be looking for. If you tell it “column H and plot it against Column A” likely you’ll get a response. Also, to improve the output of these models, you need to improve your input. Your prompts seem vague and poor, gotta refine those to get better answers. Hit me up if you need help with this, I can help you navigate through it.
On this, it needs refinement, what type of product has most of its user use it "the wrong way", when natural language is supposed to be intuitive, I know that this is just the beginning but I can assure you that within a couple of years, prompt engineering won't be a thing anymore
@@steveloco1170 I disagree. Natural Language is not intuitive, it is natural to us, but it plague with vagueness, reason why our laws need judges and lawyers to interpret/discuss them.
I agree that for high level tasks, such as searches online, prompt engineering will not be as necessary as it will still be for tasks such as processing vast amount of data.
The main reason is due to restrictions on context window. I don't think a chatbot will be able to digest huge datasets to later understand vague prompts from users, because users that typically deal with these type of datasets are skilled individuals like myself. Thus, it wouldn't be economically viable to spend time and resources in creating such product that scientist and engineers are able to circumvent easily with proper prompt engineering practices.
I agree that this models will get more capable over time, but what you suggest is technically a model with a context window with orders of magnitude greater than 300 pages worth of text (>128K tokens), although powerful models will come, datasets are much more bigger than what these models will accomplish any time in the upcoming 5 years.
Of course, this is an educated guess, and I could be wrong. I still remain skeptical. And today, this product is not working as it should for this guy because of poor prompt engineering practices; adding the lack of knowledge of the technology behind.
How can people "hit you up" to get some help?
Appreciate the honest and transparent work. Saves me a lot of time.
Thank you for the analysis Luke, wanted to try it out myself and this saved me a lot of time.
Thanks for sharing, very interesting.. the concern is about have sensitive data out there in the web
According to my subscription information, I have copilot but it's not available in any of the programs - be it the browser version or in the desktop version.
That being said, with regards to the limitations on data size, I'm not sure that needs to be a big hurdle. I'm 99% confident I'd get into trouble if I used it on the data we have at work (even with the stipulations about data protection etc.), so I'd be using co-pilot on dummy data anyway, and using it to aid in quickly getting good visualisations out of data that REALLY should not be in a spreadsheet to begin with (but in a database).
Excel remains quite capable of sinking even good computers in terms of compute. The idea that people will offset gigantic data sets, and I'm kind of stupidly assuming there is some processing basis like a cell is a token (or something akin to that) - so if you have 100,000 x 10 row sheet, its a way larger issue than what is normally pushed to AI. For small business, where its perhaps small sheets, this might work out well in the present.
I've got my reservations about larger data, and what I see in the enterprise where you get really crazy linking across massive, massive sheets. Its in these really large areas where AI would probably give the best return, but equally IMHO the heaviest AI workloads.
Can it clean your data?
Great insights! Interesting to see the limits of this software. Seems like the context windows is still pretty small 😢
yeah unfortunately too small 😔 (but it is preview at least for Excel)
It’s amazing how easy it was to find this information without copilot. I wonder if copilot is going to prevent us from actually looking up negative reviews of AI.
Is a subscription service. I dont think lot of companies will pay monthly 30 dollar per user?
You'd be surprised. Now granted it probably isn't worth that in its current state. But companies are regularly dropping several hundred dollars a month on subscription services per member. Some of them get real expensive.
Hey, Copilot, are you logging all of the data that passes through your entire system of software?
Thanks for making this video. You basically helped me see I don't need CoPilot right now. But maybe when they improve it later. :)
On large data sets You need to create a dimensional model using Power BI first. Which you can ask CoPilot to do…
is there any example? thanks , i am interested in this issue
1 million records in excel is also way too much in general for excel. You really shouldn’t go above 500,000
I would've almost subscribed it to trim my excel data, if I hadn't watched this video. I really appreciate your review!
Well I'm officially obsolete, all hail our new Excel overlord Copilot!
Thanks Luke. Have been watching your videos. Very good work.
I don't think a million rows is what MS have in mind for the 'average user'. Especially if not some standard financial spreadsheet.
Great video! I wanted to ask how you went about collecting,organizing and maintaining the job data.
The new word has image creation capabilities.. will those be included in the copilot subscription or are there other ways to create images? I use image creation a lot to do sketches of landscapes and modify them at work
It’s not at all surprising that the data size is limited. It’s using GPT-4 under the hood, it’s almost certainly converting it to a CSV and just passing it to a fine tuned GPT-4 model. If I remember correctly, the newest GPT-4 can handle 32k tokens, with a token generally being 3/4ths of a word.
Now, I haven’t checked the embeddings or anything, but I suspect that it treats most numbers as 1 token per number, or perhaps 2 number pairs. Something like that. Anything with a decimal point is going to get its own token for that (or maybe paired with a single number). So I’d guess a number like 23,462.56 would eat up something like 5 tokens in the best case and 9 in the worst. If it’s 5 tokens. I didn’t check exactly what your lines were, but one of them was decently long JSON. So, let’s say you have 4 cells with numbers in it, 1 cell with a label (let’s call that 2 tokens), and 1 cell with JSON (let’s say that’s about 50 tokens). That’d be 72 tokens per line. Let’s round down to 64 (for easy math and to be generous).
That’d give you 512 lines at most that it could keep in context. Hey, didn’t you say you could only do about 500 lines? Well, there you go. That’s why.
Additionally, LLMs tend to have issues “remembering” things from the earlier parts (except the very beginning) of a message when generating the next token. They seem to remember the very beginning, and then the latest X portion of the conversation, and if the conversation exceeds X, even if it’s within the context length, you’re going to get worse answers if they needed to reference information in the portion between begin + C and end - X (where C is a very small number and X is a number that’s a very significant fraction of the total context length, like maybe 10 and 5000 or so.)
We don’t have unlimited context length and not all data is created equal in terms of token efficiency. Excel data is probably among the least efficient, whereas word is likely extremely efficient, thus, you can get a lot more out of it in word than excel. (Plus, excel documents tend to be very large more often than word files, and way more often than PowerPoint).
If we can get longer context lengths in the future, this will be less of an issue. Also, to some degree, this is using a hammer to put a screw into wood. LLMs are not calculators. They make terrible calculators. This is going to be a weakness. It must be using code in the background to do the calculations, because the LLM by itself would fail nearly every time.
There should be only one co pilot for all github & ms office apps
Hi, does identify if there are duplicated numbers? Example if you have a list of Orders and Orderlines. Does it sum up correct and identify that it needs a unique count?
Great video Luke!
00:00: Introduction to Microsoft's Co-pilot, an AI assistant powered by OpenAI.
00:18: Overview of Microsoft's keynote event and the announcement of Co-pilot.
00:36: Excel's capabilities with Co-pilot for ad hoc analysis and data visualization.
00:55: Using Co-pilot in Word for planning and incorporating images.
01:09: Creating PowerPoint presentations with Co-pilot using Word documents.
01:20: Drafting and summarizing emails with Co-pilot in Outlook.
01:47: Utilizing Co-pilot in Teams for meeting summaries and follow-ups.
02:00: The potential of Co-pilot to automate tasks and delegate work.
02:21: Introduction to Co-pilot Pro and its features.
03:13: Co-pilot for Microsoft 365 and its additional features for organizations.
03:52: Customization options in Co-pilot Pro and Co-pilot Studio.
04:25: Hands-on testing of Co-pilot in Microsoft Excel.
05:52: Testing Co-pilot with larger and smaller data sets in Excel.
07:14: Using Co-pilot in Word and PowerPoint for creating essays and presentations.
08:05: Final thoughts on the performance and limitations of Co-pilot in Excel.
Also, it puts the focus on the VALUE of the information. Mostly it's low. Very low.
Do you think that this is going to replace a lot of data analyst jobs, I just started learning and going through the google analytics certificate, but it seems that copilot just may do an entire data analyst job, not now of course but in the future.
No. Even if Co-Pilot could analyze and transform data way better than it can now, that wouldn't replace the core functionality of a skilled data analyst. Which is being able to work with people to make shitty, unorganized data more understandable and organized, and also creating platforms to distribute your findings and determine how data collection and entry is done. Generative AI is a useful tool, but we're reaching a ceiling to how much these machine learning models can actually respond based on large amounts of data and code. We're struggling to make learning models that can function well past a 128k token limit. Even then it's really pushing it. Remember, these machine learning models don't actually think, they just regurgitate information. Literally better-Google. To be useful in any tech field you need actual critical thinking skills, especially data analysis. Computers don't think. They just do.
@@joshuaduplaa9033 what do you think about data entry? I think if copilot doesn't take over data analyst jobs, data entry jobs will definitely be replaced
Not even close. If you're a data analyst worth your salt, you will do 100x better job than any AI will do for at least 5 years, probably more like 15 depending how the pace of AI research goes. Irrespectively, most of those tools are unlikely to be open to Excel users anyway.
I just discovered your channel now and I'm really liking your style 🤓
Luke, with the announcement of the Copilot+ Laptops that run the Snapdragon X Elite chip with a high end NPU I'm wondering what the performance would be on that test which didn't perform well on your VM. Knowing that Microsoft typically releases software updates that will take advantage of new hardware features before that hardware is even announced, I'm going to assume those new laptops will smoke that test now.
Also. given VM support for NPU is still early days, odds are you will have to run on physical machine to see any NPU goodness.
Thank you! I've been so looking forward to seeing actual use case and not promo videos.
Great; a much-needed video - thanks! The lack of data security with the pro version is a dealbreaker for me. Shame, because it looks useful. The reason for the limit in the number of records that the AI could handle is almost certainly because the non-business version has a limit on the number of tokens that can be submitted, versus the business version - same with the $20 paid version of ChatGPT vs the enterprise version of ChatGPT
I haven't used copilot yet, and not sure if I can use copilot to its max potential, but I can already see a lot of features that really take off a lot of stress at work (I'm actually rooting for it because of the meetings summary feature, that would help a lot....)
This is revolutionary.
Is the copilot pro only available in Microsoft 365 apps or can someone having 2019 microsoft office use it as well?
Wonder if Copilot's preview version cannot cope not only with data amounts but if it might not work properly in a VM/Parallels 🤔
Windows 12 is currently rumoured to require 16GB minimum for system RAM, alongside the requirement of a 15th gen INTEL CPU which will come with an embedded NPU for using AI features such as Copilot natively on your system and in the windows shell.
CP also can't analyse several excel docs at the same time or several sheets in the same workbook - never mind a large single sheet with thousands of rows.
Very, very limited. Copilot in Excel honestly seems (right now) like it might as well just have been made into an improved pivottable/chart wizard.
Ask it to label the horizonal axis of the histogram with the midpoint of the bins rather than the range… i spent 20 min the other day trying to do this and couldnt figure that out. Curious if copilot can do it.
No more googling formulas, Copilot's got my back now.
Hey Copilot, Join the meeting and respond and contribute on my behalf. Do this for all future meetings...
Great vid again, so nice to see the latest news explained like this, this way you know what to subscribe to or better wait a while (or forget about it) but hoping copilot is gonna work because of it’s MS connection
Co pilot will make he a hero, and everyone else using it. The bar has been raised considerably.
Do you know if i purchase the pro or business plan will i be able to use it across my personal and two business accounts in Microsoft 365. Or is the best way to just use it on my personal account and import data from the corporate accounts would rather not do this.
Thank you, it was very objective and revealing.
Hi Luke, one question please. May I know if using Excel for personal plan would be enough for data analyst. Or do you recommend the business standard. Thank you
In the short term, tools like copilot on office look cool, however, in the long run, copilot will hamper employees ability to perform critical thinking, do out side of box thinking and problem solving, leading to cookie cutter solution to the most important business problems since those solutions will be built on top of ai suggested started point.
Loved the video...will keep an lookout for future Microsoft Copilot videos.
I'm now Subscribed ! 🎉 👏
I tried it by the way and right now Copilot is saying one can not edit the pivot tables, or charts made from the pivot tables, so you have to ask it to edit the original data set ( Excel table you created in the beginning). Copilot implies it one may have the ability in the future to edit the pivot tables and charts.
1:29 Yes, I do just love when nobody turns on their cameras
It looks like copilot needs a larger NPU & more memory to work with bigger datasets.
Signed up for it then cancelled when I found that even with it enabled it was not showing up in Excel, Word, or Outlook. Even Copilot Pro itself couldn't tell me what was wrong. I gave up after twenty minutes of searching. Too bad I didn't see your video yesterday! But in any case, it seems like waiting until version 2 might not be a bad idea.
Did Copilot recommend changing your shirt for this video?
Sounds like Clippy is still superior.
Copilot even dressed me 😳
I am sure it could better handle data sets with fewer text columns.
Why did you need to use parallels to run office when you had just said this is available in the mac version?
Sounds like (for now) you'll still likely need to do your own data analysis and insights work, but Copilot can at least save you time building out the deck and white paper.
Just had my first conversation with co-pilot. It feels just like talking to Microsoft. Which to say I'm now really frustrated and re-considering my life choices. How do they ruin everything they touch??!
Hey, thanks for sharing! How do I get access to copilot within microsoft 365?
Copilot can barely respond to simple questions let alone do any of this. It just stops working. This is currently the problem: these apps aren't reliable or trustworthy, so they actually make everything take longer if you care about your work.
+1 to the comments about simple cleaning up data. I'd like to know if you can dump in a bunch of things like name,address,email,phone on a single column and have it break those into separate columns without error. In simple data, that is fairly easy, but when some names have three words, rather than two, and the names are space-delineated, but the addresses are comma-delineated, it gets painfully time-consuming to do a simple task.
I might have to get a single license for myself, to try the things I ask my office manager to do and see if it can help her with this annoying stuff.
I work for a 5000 it-worker company. So co-pilot would run us 1.800.000$ a year? On top of the normal licenses from MS. Yeah that seems like such a good idea.
Exactly. It's still really expensive even for SMBs.
Would be great if you could update this with a full working version of copilot thank
M365 Chat and Teams is useful. Also the MacOS apps have copilot as well. I run Parallels too but outside running it in edge. Or using edge browser on my iPhone, the Mac Apps have functioned the same.
Thanks Luke for the video
Wait so, I expect that if I tell copilot "Can you write every 50 cells a series of text starting with Lot 1 in cell A2 so that Lot 2 is in A52, Lot 3 in A102 and so on and so forth" it will correctly help me sort my lots better. No chance?
Great video Luke! I heard you have a MacBook Pro in your video. I have one as well and run into issues getting PyCharm working correctly with selenium webdriver. Do you know of any places I can look for getting a MacBook Pro set up for data science tasks? Thanks a bunch!!
Do you need PyCharm at any cost?
Copilot presentations at next year's Microsoft keynote gonna be:
"Let me demo to you how I sleep on this hammock while Copilot earns my six-figure salary."
How does all this interface with Fabric? Ir is that going to be a totally different thing?
Curious what would happen if you put your same dataset into PowerBI.
Maybe if you ran your tests in an actual Windows computer it would work.
The problem with uploading files to OneDrive or a SharePoint server under Microsoft's control is that it violates contract clauses which restrict disclosure of intellectual property or other confidential data to third parties. Given that Microsoft employs foreign nationals, uploading certain types of data (much of which is commercial) can result in a violation of U.S. export control laws and regulations (which carry penalties of up to 20 years imprisonment and a $1,000,000 fine).
@LukeBarousse How does "PC, Mac, and iPad" lead to your claim "for the Windows only version"??
A PC is just hardware. Linux runs on PC's.
I am planning to learn DATA analyst, with AI and new features ,do you think it is good time to learn it ?
6:23 why are you using a virtual machine? Because you are using a subscription intended for a Windows machine? Why not just use you Mac and OS?
Wait, so if my files are on dropbox I'm doomed and cant use Copilot??
Next level will be..
Me staying at home while copilot goes to work for me.
Not sure what version of excel and copilot your using (maybe a future one that will change excel forever...) but been using it at work and its very weak
Tell me you didn't read my post...
Well, given the expectations this sucks. But is it worse than an "average" office worker?
Touché 👌
Do M365 users get higher priority for compute power?
I tried using Excel to chart my spending in different categories based on my credit card history ( about 200 rows). It kinda worked but also couldn't quite finish it. It would not ignore one category of data like I asked it to. It was making charts for me many times but kept including the category I wanted to exclude. Then after about 15 attempts, the response was " I am a large language model and can not make visual representations" >.... HUH? It was making charts , but then stopped. Is that Copilot telling me there is a limit to my attempts or do I need to get out my tin hat? Overall it works nicely, and I assume it and I will improve over time.
Mistral is good at spreadsheets too.
That's the future! It's really interesting how AI has changed the world just in the last year... I remember how everything started with the hype around Dall-e at the end of 2022, then Chat-GPT and GPT-3 were released to the public. From there it only took a couple weeks until everyone was using them. GPT-4 hit the market really hard and literally changed everything. Now after literally only one year with these large language models out everyone and even my grandmother know of theses AIs. To think that I can sit down and talk to my 82 year old not in the remotest techsavy grandmother about what AI has done to our world in the last year alone is astonishing. To think that my grandparents literally grew up in a small village in the middle of nowhere with literlly no electricty and a donkey as their main vehicle is really putting stuff in the frame. For them they went from middle ages to scifi in a whole lifetime. To think that in their lifetime they started without the concept of computers let alone microchips and to get to a point where you can literally have a computer talk to you is making me excited at to where the world will be when I am their age.
Now I can go for a quick nap during meetings
"Copilot is just thinkin, and thinkin'..." [...ad interjects...] lol! That's when I'd have gone to refresh my coffee irl, anyway! :-D
😂🤣
We all are going to be dumb and clicking copilots in the future
I'd hate to think just how much data is being handed over to Microsoft in this process, corporate data included
Great video which yeah even enhanced search I mean Copilot on Bing/Edge still has issues and needs to be lead sometimes so I'm not surprised at it's ProFormance in this video that said... I normally use word for writing, notes and gamedev and powerpoint for timelines and excel once for a card game... So yeah it sounds good especly for $20 a month and I get some bonuses to the other Open ai bots to.