人工智能发展到什么程度了？是不是太快了点？

The development of Artificial Intelligence is too fast So fast that it sent a shiver down my spine When ChatGPT went viral I thought this was gonna be a small climax but the development of generative AI is slapping across our faces non-stop It has been less than 2 months since my last episode regarding ChatGPT There was a tremendous change in the AI market The influence of my video is quite impressive right On March 14th

Google released large language model interface PaLM API A day later, on March 15th OpenAI introduced GPT4, upgraded version of ChatGPT and then Microsoft announced that they’ll merge GPT4 in their Microsoft Office The next day on March 16th Baidu introduced ERNIE BOT which was first large language model in China On the same day, image generation company Midjourney introduced their fifth version the image produced is really very realistic

Later Huawei Alibaba 360, Sense Time have all introduced their own large model language Most importantly, they not only introduced the basic model they had slowly integrate these models to their applications We’ll talk about this later This incredible development speed has made people feel excited but also a little bit worrying In 2018, Musk had said remember my words AI is more formidable than nuclear weapon We’ll talk about the

problem with AI at the end of the video Regardless As far as you can see, the current development trend is already unstoppable So I think it is necessary to publish another video on AI to help everyone, including myself and update changes in AI field over the last two months We’d also talk about the potential risk it possesses After all we need to stay on trend right?

In today’s video I don’t really want to emphasize or compare the competition between these companies because it’s still in early stage and the products are not mature yet and many are still in beta It doesn’t make much sense to compare which is better Let's take a look at the latest breakthroughs applications and products Actually AI, machine learning are quite common these days but they are mostly focus on analytical work like big data analysis

AlphaGo Face recognition and etc But this wave is all about Generative AI Generative AI AI that can produce new content like new texts, images new codes, sounds, videos Meaning to say, AI has already possess ability to create The thing that it creates is close to the real deal This is a big step towards artificial intelligence in the true sense.

Let’s first take a look at large language model First we’ll definitely talk about GPT4 the newer version of ChatGPT After you pay and upgraded to ChatGPT Plus then you can use GPT4 I could even figure out their future name ChatGPT Pro ChatGPT Max ChatGPT Pro Max Before this we’ve talked in detail about ChatGPT the GPT3.5 it basically can answer any question it can write codes, summary and etc.

of course the accuracy can’t be guaranteed So what is being upgraded in GPT4?

The most obvious upgrade is that it can now understand picture Not just identifying object in the picture The key is that it has common sense and a little bit of humour For example if you ask what’s funny about the picture it’ll tell you You plugged an outdated VGA port into a modern smartphone What’s the problem with the picture it’ll tell you someone is ironing clothes at the back of a cab it’s not normal

What if I cut this thread it’ll tell you that the balloons will fly If you draw a bad sketch and ask it to create a website GPT4 could generate a code The website will definitely not be mature but the point is it can understand the simple sketch that you drew Isn't this kind of scary?

I’d like to remind you that after you’ve paid to upgrade to Plus you could use GPT4 but actually the image input function is not available yet That is only open to the API interface Apart from image input there’s another major upgrade with ChatGPT We know that before this it has a flaw and that is sometimes it talks nonsense or that the accuracy is not high In GPT4 version

The accuracy of its answers has also been greatly enhanced If you let GPT 3.5 and GPT4 enter our human exam the blue bars is the percentage of the 3.5 The green bars is the percentage of GPT4 Most examinations have somewhat improved The most obvious one is the US Bar exam The previous edition was in the bottom 10% GPT4 now is in top 10%

Apart from these two points, GPT4 has some other improvements like higher word limit and better at avoiding prohibited content reduce cost etc. We won't go into that Maybe you are not a stranger to the answering capability if ChatGPT But with GPT4 people are getting more creative like generating Swift code for animation create Snake game generate legal letter generate smart contracts for Ethereum and more The thing is, the improvement of these products

happened in less than 6 months Test scores have gone from bottom to top from primary school student to undergraduate Imagine what it could become in two years time It is very likely to surpass most of the humans or experts As for GPT4 why it develop so fast how their model improved or how many parameter it uses OpenAI no longer publish it It used to be a non-profit organization GPT2 is even open source

After Microsoft’s investment it completely turned into a private company Competition is fierce in the market and all these are trade secrets Not gonna say anything anymore not even gonna tell you the parameters On large language model GPT4 by Open AI Definitely has first mover advantage But other developers are trying their hardest to get on this train Anyone who can will train their own model like Google’s PaLM LaMDA Baidu’s ERNIE Alibaba’s Tongyi Qianwen Meta’s LlaMa

I don’t know how to pronounce it Huawei’s Pangu as well as Claude created by former Open AI’s employee There are many large languade models in the market Google itself owns 7-8 models You might think only those big tech giants could develop these expensive LLM Stanford University used $600 only $600 to create their own LLM based on Meta’s model They name it Alpaca It’s the same level as GPT3.5 And its code is open source

Maybe soon everyone will be able to train their own LLM with their phones So which model is better?

It’s really hard to say but for now people commonly think that ChatGPT is more mature Looking at how fast this thing develop probably six months later it’ll have another earth-shaking changes Each developer says they have their own unique feature this is one of the main characteristic of machine learning It’s hard to compare them like how you compare phones or computer with this you have to look at parameters to get a basic classification With LLM

you don’t know the specific generated logic you can only compare the parameter is it tens of billions, hundreds of billions or trillions even with this it’s not accurate Moreover this is what is claimed by developer you won’t be able to know if it’s true It's kind of like when you evaluate a student's ability if you look at how many books he read the hours he takes on reading it could provide some valuable reference

But we are familiar with the most intuitive It’s the same with LLM you have to test it to know if it’s good or not So I tested a few models that I could get access To make it easy to read I used Chinese to converse with it If you are interested you can pause to read I think there’ll be a AI testing agency to test these models and rate them like financial rating agency because

who would test all the models on their own right Alright, what we talked about earlier are mostly the basic LLM they pursue the versatility It's a bit like the general education that children learn If you want to train a specific sector AI you could of course train it yourself There are some very rich and powerful companies like Bloomberg, did that but SME definitely unable to do this then they could carry out second training based on

these well-trained LLM It's like you're bringing in the kid who just finished his general education and then let him study specialised course for a few years based on your data so that he becomes an expert Or simply use the API of one of these larger models and make it part of your own service Its extensibility is actually very strong It’s more than chatting with you or generate some images for you The important part is

Its potential applications are very wide and likely to penetrate the entire market This is why ChatGPT and Google’s PaLM opening their API is a very big deal Let’s take a look at the current application The classic example is search engine I believe everyone is familiar with this Microsoft integrates ChatGPT into their search engine Bing they named it New Bing Google merged LaMDA and PaLM into Bard

it’s more or less half of a search engine Apart from that there are You.com Baidu 360 A traditional search engine that could chat with you We’ve talked about this in ChatGPT episode so let’s not get into it today There’s another suitable application for this and that is in office Large Language Model What it does best is organize language So when you work it could help generate a context write a summary correct grammar

These are the most direct application For example Notion a very big note-taking software They introduced Notion AI this year and went viral Actually they connect ChatGPT Look at Microsoft they hold both Office and ChatGPT two trump cards On March 16th they introduced Microsoft 365 Copilot GPT4 was embeded to Excel, Word PowerPoint Outlook

and other Office softwares For example if you take note in One Note then in Word, you can use AI to generate content and summary After you adjust the text and add in details PowerPoint could then generate PPT Their format looks quite good actually It will also create animation As for excel If you have a bunch of data you can ask what is the feature of this data can you generate a report It sounds quite incredible

and wonderful right But with AI’s current ability Its application is still not that wide It’s not possible to be as great as shown in the presentation video get everything done with just one click But I’ll say it again the odds are high in the future If it really is that smart you could let it learn on how to reply to your boss and do all your works for you How great is that Well then

your boss will probably don’t need you anymore Actually there’s a company that lets AI become their CEO August last year, NetDragon appointed Tang Yu to become rotating CEO Tang Yu proper name with surname is actually a digital human A CEO who is everywhere and always available As for what has it done or how is its performance we do not know yet Another interesting thing to share with you Another application of Generative AI

is generating image It's not just the application layer anymore because the model it is based on is no longer a large language model I believe you would’ve seen some images generated by AI You just need to describe what you want or just give it a sample then it can generate a very realistic image with any style that you want like Disney style, paper cutting Picasso style, oil painting etc All can be generated within seconds

These are images generated by Midjourney I don’t know how you feel about this but I was quite shocked when I first looked at it The other mainstream apart from Midjourney are DALL E2 by OpenAI Stable Diffusion And we know the king of image processing is It's Meitu No no, it’s Photoshop Adobe came out with their own image generator model called Firefly It combines generative image with image processing

Then the room for imagination becomes wider For example change this image to winter Here you go, done It’s a snowing scene I don’t like this lighthouse, change it Change the material of this watch and let the watch moves add a little river to this grassland Look at the effect Not to say we can’t do it before but even the professionals would take a while to do it Meanwhile AI could do it in just few seconds

Apart from generating image there’s also AI that generates video There’s Runway Just key in few words or upload an image it could generate a video Of course this is not as mature at the image ones Another example is music creation I can create any music that I wanted based on the style, speed and key that I want The generation of these multimedia is AI What’s shocking to me is not the result of what it generates

but its development speed Two months ago the fingers that AI came out with was really terrible now looks so much better I think probably in a few years we’d see many images videos musics created by AI or at least majorly involves AI Another application of AI is in finance We mentioned that Bloomberg, known by those in finance industry has been using data accumulated over decades

50 billion parameters and trained a financial expert called BloombergGPT The test results are said to be good Actually in China, many banks and brokerages have announced to integrate GPT Especially HiThink they are one of the earliest to start using AI Although they stated that their technology is far from the international standard but their stock price has more than doubled

This shows how optimistic the market is about the application of AI in the financial field There's another application that's particularly popular in AI circle AutoGPT Autonomous robot I think it's kind of imaginative It's not a consumer-facing app Instead, a big guy just put an open source project on Github It connects to the interface of GPT-4 and create a robot that operate autonomously You just need to give it a target for example

start a business for me and let it earn money continuously and as for the rest it will all be handled by AutoGPT Sounds amazing isn’t it So how does it work?

It doesn’t give you direct answer like ChatGPT because we know that in most cases GPT is not reliable It basically connects GPT4 with abilities like programming searching, long-term memory and etc You just need to give it a target and it will asks itself how to do it and then execute If research needs to be done it will go on web and search on its own It will keep asking itself questions and optimise the details

and realise your target as much as possible Of course, that's all I can say This actually most people are just playing with it But I think it’s pretty interesting I’d like to mention some applications of AI in other fields such as education As AI is so good at language it is very suitable as virtual teacher and teach language Duolingo connected GPT4 called Duolingo Max Another field is e-commerce like Shopify

connected ChatGPT to help businesses to write product details It’s also applicable in programming A must-have programming tool Github Copilot is one of those I think worth understanding on the application of generative AI We of course welcome everyone to comment and discuss in the comment section Maybe many of you would think if language model deserves so much praise Well it’s just a chatbot that often makes lots of mistakes For me personally it is quite revolutionary

Firstly on the mistakes issue it’ll be solved with the fast iteration Actually the appearance of LLM is a very basic revolution AI can already generate its own content is akin to when internet started to appear Everyone just thought that they can read news online now more convenient Who would’ve thought years later there’ll be things like Facebook Taobao Wechat live streaming and etc All these applications not only bring great commercial value

it also changes how people live So for generative AI the key is it produce huge amount of possibilities For example maybe in future you can train a smart assistant based on your own data It will know your preferences your personality and will be online 24/7 What a big market it is Of course now it is still in early stage but even then you can see great commercial value in the current AI wave

Metaverse AR VR were all the hype before it has great possibility but why is their development not as smooth?

The main reason is that it doesn’t have much commercial value If you look at VR now including Metaverse their commercial value is not high but it’s different with generative AI The popularity of ChatGPT right now probably OpenAI didn’t even expect how it suddenly blow the market So when there’s huge business interest a business opportunity with endless possibilities in the future I think we probably have never seen so many of it all at once

This is why Microsoft keeps pushing on OpenAI to introduce it to the market even though it’s still debugging Many developers were forced to publish their beta version Some didn’t even have a product yet and were forced to announce their plans Robin Li mentioned in the press conference why they need to introduce it at this time it’s because there’s market demand Most importantly, all the resources and talents are all flowing to this direction

and it's only going to get faster How much risk does it bring?

In fact, what many people are discussing or even fearing is the emergence of AGI, Artificial General Intelligence To put it simply, a sentient AI that surpasses humans So is it possible for AI to be like in the movies turning against humans In 2020, Musk said that AI could surpass human by 2025 On March 22nd Microsoft published a 155-pages paper

with the title Sparks of Artificial General Intelligence: Early experiments with GPT-4.

On the same day an open letter warning human not to compete in AI immoderately was published Many people across the fields including Musk signed on the letter I’ve checked on it two days ago there are over 26000 people signed this open letter calling upon these companies to stop training AI that’s more powerful than GPT4 for 6 months Of course we know they wont stop competing in the AI race just because of the letter

Whether you can accept it or not this wave of generative AI is irreversible Our entire business environment is likely to be drastically changed within a few years Do you think this current AI wave is scary?

Actually the examples I gave is a bit extreme The paper published by Microsoft I’ve read through it it basically says how incredible GPT4 is it answers questions really well so good that it’s close to AGI There are many experts that came out and criticize the paper saying it’s ridiculous, not even close to AGI As for the open letter many people went and signed the letter But this kind of warning is actually a bit exaggerating

It’s not big of a problem So how big is the risk of this AI We mentioned about ChatGPT fundamental logic in previous video To put it simply it’s like a word game guessing the next probable word based on the previous text AI doesn’t really know what it says It only possess the ability to learn and imitate it’s like a super parrot So if it’s technology like ChatGPT then it’s not that scary for now

But this “for now” can last how long we do not know If this parrot can answer all the questions perfectly How would you know it won’t develop its own consciousness I’m certainly not an expert on this field I’m also quite curious what you all think about this Do you think ChatGPT can develop consciousness and reach AGI level Many would worry if AI would overtake their job For me I think there’s no point

Thinking about it too much What we can do is try to understand as much possible And try to use these AI And let it assists us to complete our job The one that replaces you is not AI It’s the person using AI