An interesting thing to point out, O(N^2) is not optimal for checking the longest palindromic sequence, there is an algorithm called Manacher's which finds all palindromic substrings in O(N). So ChatGPT wrote out the most popular 'good and optimal' solution, rather than the actual best.

In my next interview, I'll use the network error strategy!

It being really good at solving algorithmic problems is not surprising at all given all the solutions exist on the internet.

ChatGPT is not designed to think but to present knowledge, which is actually a surprisingly hard task for humans. Many professors I know have deep understanding of some topics, but fail to grasp what is potentially hard for others to understand.

Congrats ChatGPT for landing a job at Google! 🎉

I think the main reason why it passed the interview so well is because the prompts correspond exactly to the problem you wanted it to resolve. I am totally mind blown it can do that so well and so effortlessly but in real life, what we might consider as problem prompts are usually a bit messier and not that abstract and it would take a fair amount of effort to get to the point where they are as polished as an algoexpert prompt. Failing to provide that level of clarity and completeness might result in a seemingly correct solution that is wrong in the most subtle ways imaginable.

Yes, mind blown. The last question was very chatty, it is insane that it could destill the problem statement from all the text.

You should do this again but with variations to the questions. I've heard that GPT struggles more when it has to extrapolate concepts to unseen questions (i.e. a question which might have no solutions available online).

Incredible and scary. Though it is expected to do well in questions and solutions available on the internet, the fact it is able to understand and give different clear solutions at that quality is mind blowing.

I think Clément made a rather wrong conclusion about GPT from this test. ChatGPT is not a great coding problem solver, but it is a very good tool for eliminating well known interview problems. I have a lesser known problem, also one where I don't specify input - allowing the interviewee to design the API. The ChatGPT got very confused - it likely mixed some different solutions (that use different APIs) and wrote some code that uses one representation, and some that uses another. The problem is - they don't make any sense together. It also missed important parts of the problem.

This is such an interesting video. I actually tested chatGPT for the leetcode question for my Google internship. The code solution blew my mind, but in some cases, it didn't pass the test cases and I had to change 1-2 code lines. GitHub Copilot was also impressive, but I found clarity issues with it. But both AI tools really helped me a lot in preparation and practice. These tools are going to disrupt the software developers' lives because they are adding value to mine.

What this shows is that there are enough coding interview questions and solutions out there. It is very impressive how it writes and comments code, but it is not inventing new things here. Hopefully this will put an end to these learn-by-heart coding interviews and lead to more interesting/useful interviews. Also curious to know if Clement sees a thread for AlgoExpert in this, as it appears to provide the same (if not better documented) results.

This is the craziest technology leap of my life and already people are writing it off and minimizing it

I think it would not replace devs rather help them and make it easier for them to develop apps, webs, etc.

tbh, props to the people who made chatGPT. they are the smart ones. idk why they did this but I think it is to help engineers.

That was pretty incredible. I'm still learning programming and I could see this as a good tool for research and testing solutions.

I feel like we're gonna get to the point where people will get hired for certain jobs based on how good they are at writing prompts for AI...

I LOVE the difference in the comments and reactions to ChatGPT between programmers and artists... programmers are excited for this tool and artists are angry and trying to shut it down but BOTH are going to be affected professionally by this tool (programmers more so).

this thing is so freaking insane. I got a promotion at work in 2023 using this thing for the last 2 weeks. I didnt tell anyone i was using it but the code i was writing was so good that my leadership was like WOW GOOD WORK!. its freaking INSANE

I love ChatGPT. I work as an embedded software engineer. I am not afraid of it at all. In fact, I am excited to use it to complement my work. I have been using it to some extent this week even.

Recently NeetCode posted a video on KZfaq where ChatGPT got it wrong. Well, we know it's not perfect but that's amazing, I have tried it and was blown away. It will only get better over time

Imagine that: what if the network errors were intentional (maybe the AI learned how to trigger them), so that it could give you multiple solutions and impress you even more. :D

I made a question up off the top of my head and ChatGPT nailed it. (Probably an easy question for you pros out there) Question: You are interviewing for a python engineer position, say you have a list of tuples. First element being a string (animal name) then the next element within the tuple being the number of years that animal has survived. sort the array of tuples in order of animal years then by the name. There will always be at least 5 animals in the list with a maximum of 50 animals. Please return the optimal solution

ChatGPT is at capacity right now. You just killed that website with the video :D

A brief summary: The speaker describes conducting a Google coding interview with Chad GPT, an artificial intelligence chatbot developed by OpenAI. The speaker states that they have previously conducted real Google coding interviews, both on KZfaq and in person, and that they have prepared three coding questions of increasing difficulty for Chad GPT to solve. The first question involves writing a function in JavaScript to determine if a given string is a palindrome. A palindrome is defined as a string that is written the same forwards and backwards. The speaker notes that single character strings are considered palindromes. Chad GPT responds with a solution that involves using two variables, "left" and "right," to keep track of the leftmost and rightmost characters in the string, and using a while loop to iterate through the string from the left and right ends towards the center. If the characters at the left and right indices are not equal, the function returns false. If the characters are equal, the indices are incremented and the loop continues. If the loop completes and the characters at all indices have been found to be equal, the function returns true. The speaker is impressed with Chad GPT's solution and notes that it is the optimal solution for this problem. They also mention algoexpert.io, a platform that helps software engineers prepare for technical interviews, and encourage software engineers to visit the site and use the promo code "CLM" for a discount on the platform.

I did an interview at google before and if there is one thing that I remember well is the fact that the interview problems had missing information (or unclear information) that pushed me to ask questions. The one thing besides solving them was asking the right questions to clarify the ambiguities in order to get it right. I'm not sure why you gave it a 4 in that part since the problem was pretty clear and there are no questions involved (also a probably overused problem since It is a well known google interview problem according to what you said). Don't get me wrong, the GPT is probably one of the coolest things I have ever seen, but, besides being super cool search engine with some aggregation power in it, I do not see it doing anything else then being a productivity increaser (a super stackoverflow). I really want you to do an interview with easier but more uncommon problems and make it ask the questions to see if he handles them right and produce code based on the answers.

It’s was so efficient with the last solution because it has seen the exact same question somewhere on internet and it knows the solution already

This is both amazing and scary at the same time.

As soon as they can create instruments that will allow this model to work on a new version of this model without human feedback, it's time for all of us to change professions. Our best chance is to become a managers who would describe the tasks to AI to get optimal solutions. But even that would not last long.

As an active user of co pilot, I'm not surprised at all, I mean it's true that the technology is amazing and has a lot of potential but it is trained by using millions of peoples data without their active consent. Whenever I'm stuck , i could either spend an hour to try and figure the solution out or ask my copilot to fetch me a working solution from a guy who suffered a similar problem . Other then that, Co pilot also helps me to write code efficiently and you can use it for other purposes to (i.e writing this comment).

This is not really unexpected that algorithmic are solved to perfection since there are so many sources of solutions out there and its dataset/knowledge is based on that.

This really should have people like Clement and sites like leetcode and hackerrank sweating - because it exposes live interview questions for what they are - a competition in memorization. The reason generative AI does so well with these types of questions is that all it needs to do is look up existing solutions to these generic problems. The real challenge software engineers usually face have more to do with understanding user requirements and hardware restrictions that need to be worked around to do what you want, not recite an algorithm for traversing a tree. Perhaps more software companies will change their interview styles after this, which would be a good thing. Let all the leetcode and hackerrank sites rot away as a bad memory

I'll bet you the ChatGPT was trained on Algoexpert Q/A, as it is a popular platform for interview prep

Mind blown. Well I was mind blown with the previous GPT3 codex models as well, that I've shamelessly exploited in some of my projects. I've done some of my own testing and yes, ChatGPT is formidable. And its a natural language model(s) - it's not a specialist "coding" model. Imagine it doing stuff like this is Physics, Medicine or Mathematics. Or specialized versions of this model. Never thought I'd see this in my lifetime :-)

This is just a computer doing math, basically. If you really want your mind blown, take a look a this: Q: Which of those options do you find the most compelling? A: I think it would be most interesting if the generation ship became a space station after its journey, as it would allow for further exploration and colonization of the Trappist system. It could also serve as a hub for trade and communication between the different colonies, and potentially even other star systems. Q: That sounds great, where would you like to go from here? A: We can explore the societal dynamics of the colony as it grew and evolved over time, eventually leading to the formation of the current oligarchic government. We can also consider the role that technology and scientific advancement played in the development of the society, and how it shaped the cultural and political landscape. Another direction we could take is to delve into the internal politics of the corporations and how they interact with each other and the government. There are many possibilities for where to take the story from here. Like... dude. 🤯

This is really amazing!! I am totally astounded as I had to work for years to reach this efficiency level in C++. I can see the demand for proficient programmers would reduce drastically. I can see job implications as fewer programmers are required and possibly very complex pieces of software will be written in no time. We'd see millions of lines of code becoming a norm.

Scary to think about the possible impact it has/Will have. Remember they have even stronger models.

Network error happens when responses are too long and for some reason it will timeout. To fix this ask it to: Add numbers to lines and when you reach line number N you must end your response, tell me how many lines are left and prompt me to continue. Or something like: Tell me the first N words then end your response and prompt me to continue.

Your surprise is surprising. In this case it is not how well the bear dances but that it dances at all. You should be absolutely stunned that it can coherently reply to even the simplest problem. After that, it's all gravy. Once you have hardware that can compute a 2x2 matrix, you should not be impressed if it can manage 100x100. That's why computers are handy. They do what we don't. They scale well but think badly. The breakthrough is that GPT thinks a bit better.

Its wonderful how advanced Neural Networks are becoming

I was pretty mind blown. I was also really happy to see how interviewers score potential candidates.

I like the idea of thinking of ChatGPT-like tools as a search engine and as a compressor of information. It creates a database of the internet and lets you search it, with nice and useful presentation of results.

Second problem: the longest palindromic substring can be solved very elegantly using recursion: const isPalindrome = (str) => { for (let i = 0; i < str.length / 2; i++) { if (str[i] !== str[str.length - 1 - i]) return false } return true } const longestPalindromicSubstring = (str) => { if (isPalindrome(str)) return str const leftStr = longestPalindromicSubstring(str.slice(0, str.length - 1)) const rightStr = longestPalindromicSubstring(str.slice(1, str.length)) if (leftStr.length > rightStr.length) return leftStr else return rightStr }

everyone gangsta til you realize u r a CS major and there's no way you'd have solved that question optimally or even ever get hired at Google :')

Ok, BUT: the answers to those kind of interview questions were already there - some for years, some for decades. The bit that it can actually spit out a working algorithm is not the impressive part because you'll likely find almose exact answers on stackoverflow. What is truly impressive is that it parses questions in plain English (or many other) and connects it to good answers. It's like google 2.0. In an interview situation no one actually wants you to write code but for you to show that you understand the code (by writing it, if that makes sense).

Considering these questions were all on the internet verbatim, right next to their solutions, on your website, I'm not surprised it got them right. You should try altering the questions or even making up new, original and unique questions to see how well it does. What it's doing here isn't too far off from copy pasting the solutions from your website.

It's very interesting to count how many times you used the phrase "mind blow..." It kind of tells the quality of product...

OMG the third question is exactly the one i got at uni just a month ago for my first midterm in algorithms and data structures

You should have told it that it got a "Strong hire", just to see it's reaction.

OpenAI has a team working on generating code - Codex. I guess they may have more training data than that easily available online. I still think all of these questions could be in the training data. When I tried asking chatGPT easier, but more customized task, it wasn't doing so well.

Can't wait for someone to get a job by answering "well, given the simplicity of the task I would just ask chatGPT to write up a solution, rather than wasting my boss's time doing it myself."

I am truly shocked at its capabilities

The future is here, we are here to solve big problems, not to write small algorithms, finally someone to open everyone's eye's in the it industry

The real test would be to ask it to improve its OWN programming to make itself more powerful/efficient

0:41 love the implication that real people and real software developers are two separate things

this shows why these questions are silly to ask in a job interview

May network error save us from the A.I. apocalypse.

Mind blowing, it exactly solving the question like actual Google interviewer specially when it stucks it just slow down or disconnected the internet connection

You don't need to restart new thread each time you get network error :) It's enough to refresh the page.

I asked to come with a story about a food contest between Snorlax and Cookie Monster. The story ended with Elmo saying overeating is bad and they decided to go on diet and stop eating

Genuine question. I'm at the start of a career change moving into software development and only been doing it a few weeks. Am I wasting my time moving into this area and learning since chatGPT and git hub co pilot are around, are developer jobs in danger?

Well this is an interesting concept. I wanna see the roles reversed too 😈

At 5:51 I noticed ChatGPT claims isPalindrome("") returns false when it should return true, since "" reads the same forwards and backwards. In fact the implementation that it gives for isPalindrome would return true too. Funny how it can get the hard stuff right and goof with the easy parts.

Plot twist: AlgoExpert is part of ChatGPT's training data :)

Am I crazy thinking hard question could be simply solved with one loop just to figure out how many downsteps or upsteps in raw are there? Thats when you simply count them either way. Expl: 8, 4, 2, 1 or 2, 3, 4, 5 would be same, 1+2+3+4, when you increment counter each time change is repeated(for higher values in raw or lower values in raw) as four changes for both cases. When going downwards just add little math to check how many down step are there, because you want always to start counter with one when checked value is lower then previous.

ChatGPT has to be a human being with the world's knowledge

But also, it's funny to see people say 'it just knew this already'. To that I ask - how much of what you know is actually original and inventive thought? Even your creative solutions, aren't they just a distortion and amalgamtion of things you already know?

"What is the space-time complexity analysis of this solution?" Now I'm 100% certain I never want to code for a company again and everyone has lost their mind.

I think it will be a great personal QA testing robot as well.

Is this gonna kill junior dev jobs? Should I stop studying and look to something else?

The solution given here for longest palindromic substring is not O(n^2) it is still O(n^3) (in the worst case) because at every single character you are potentially doing repeated character comparisons, that argument : “at each character you are doing 2 checks” this check doesn’t take constant time and instead takes O(n) time

I am mind blown and speechless at the same time

I don't want Chat GPT or any AI to take any potential Software Engineering or Security Engineering jobs in the future. I don't want to start my career only for it to be automated within 6 months. It's just not fair.

It fails on novel algo problems. They tried it on codeforces and it just spits out random code that sometimes doesn’t even compile. Basically it is a form of google/stackoverflow at the moment.

Might want to note that the fact Chat GPT rewrites the function in a slightly different way isn’t indicative of any thinking or that it wasn’t copying off something else. It’s noise in the transformer. Everything is a recombination of input data, while this can also be true of humans, what chat-GPT does is much more primitive.

Your reaction was great! and in fact, it starts to be a scary thing...

n^2 is not optimal for longest palindromic substring. There are a bunch of ways of doing it faster. n*log(n) is good with hashing or most other algorithms, but n is better and also possible.

So, I don't need to renew my AlgoExpert subscription next year lol

ChatGPT probably knew the answers because these specific Google questions are common knowledge in Internet nowadays, but you will be surprised to know that ChatGPT was unable to create an algorithm to count the vowels in a word.

Note entering a simple "continue" if chatGPT stops will output the rest of the answer

The answer to what's asked towards the end when he wonders if it was posting solutions it already knew somehow is going to depend on how you define "know," because I suspect it's likely using a lot of existing data, then maybe crawling over it to create its responses, and if that data is coming from a live internet connection would you still consider it something "known?" Even if so though, fundamentally this "knowledge" would depend on whatever algorithm is going on behind-the-scenes that synthesizes an answer from it, even if it wasn't using a live internet connection to obtain the data in real-time. It's going to be really interesting when we see how these interact with each other though IMO i.e. can multiples of them working together to solve problems help the overall performance?

So far everything code related i have tryied to ask the AI left me speechless, it is a tied beast thx to the network errors...

This video is INEVITABLE!

I‘ve seen similar videos like this and tried many options myself and it‘s all the time the same. General algorithmical problems with limited scope (because in comparison to a real-world project all these algorithms have a very tiny scope) like the ones in this video? Yes, fantastic! Complex solutions to not so generic problems? Not a chance! Therefore, my 2 learning are: 1. Teachers have no job in the future! Any generic thing a teacher can teach me in any subject, the AI can teach me as well and even better! 2. Human resources employees create no value to a company as their methods of questioning a prospect are useless to find the best prospects. Of course, the second point about the HR department of companies is something, that is true even without an AI and I knew that for 20 years already.

11:59 To see the rest of text/code you can write something like "Write the rest" or "Continuation"

This is truly mind-blowing.

Longest possible palindrome in a string (single string, not a sentence etc.) can be found in less than O(n^2) in average, if there is a palindrome. Since the method has to access memory more, that might not be faster tho. substr pieces, for example string is abababb, you remove one character from one end, test, then the other end test, and return immediately when isPalindrome returns true. No need to test for ALL of the possible solutions. Hence, average solution is more like o((n^2)/2) -- but depends on the dataset, since palindromes are not commonplace, a comment like this would yield more like o((n^2)*0.99) or whatever the natural frequency of palindromes actually are in natural language - but on dataset of mostly palindromes the it could be more like o((n^2)*0.01). If the string is completely a palindrome difficulty of solving was O(1) A little bit of memory overhead here removes the need to find ALL the possible palindromes, since we only care about the longest. Rest of optimization depends on the language, interpreter and platform on what is the fastest methods to do. for example you don't actually need to even substr but you could loop the individual indices of the string; well essentially what substr does but you can put it directly into the algo to avoid even more lookups, spend a bit more memory on avoiding extra strlen() calls etc. Point is that optimization is more than just optimizing the base algo, and there are always methods to make it faster, and just because at a quick glance it seems O(n^2) is the minimum, it might not be. Infact, that praised o(n^2) solution is kinda weak at 14:29 -- enough to get the job done, but inefficiently. Right now i noticed the comment of Dragono about Manacher's algo to find all possible in O(N), so even O(n^2) to start with is inefficient solution.

ChatGPT just started It's journey. I wonder what it will be able to do after getting more updates.

So discouraging for someone who was hoping to make a career switch into software engineering! Glad I didn't spend all the cash I would have needed to go through a year+ of training perhaps. Back to the drawing board.

if I went to an interview and said def isPalindrome(str): return str==str[::-1] would I get the job?

It’s now become a lot more stable. Now it occasionally quits midway through but generally throws less network errors

I personally think the next wave of layoffs will be because of ChatGPT doing the Software Developers' jobs.

Hey, I'd like some input here from a programmer. I started to slef-study web dev a couple of month (I'm brand new otherwise to coding). My main goal was to create a personal project, and I can definitely see how ChatGPT will help. That said, as a secondary goal, I was hoping that throughout this process I could eventually gain the skills required to land some entry level job down the road and move up from there, or even do some freelancing for smaller/simpler projects. It seems to me that ChatGPT poses a problem. It can certainly replace novice guys and gals like me who are self-taught. Am I right to assume that this tech might render the 'self-taught' or even the 'bootcamp' candidates useless? My fear is that if the low-end coding work can be done by AI, the actual programmer that is supervising it will always be senior level, or at least have a solid academic background in CS. So what are your thoughts about coding noobs who are currently following a non-academic path? Did we miss the train?

Watch the openAi GO documentary. First version lost once to the best go player. Latest version doesn't ever lose to that initial version. It looks like majority of software engineers will be laid off by the end of this decade(maybe even earlier) and the ones that keep their jobs will never see high salaries again.

Clement : I asked chatgpt to solve the Edit Distance problem -- minimum letter changes required to transform word 1 into word 2 It wrote Java imperative program Java functional program with recursion Java functional program without recursion but with mutations Java functional program without mutations really scary

Welp I'll just become a farmer then

It did so well on these problem because these are popular problems on the internet. But if you add a simple nuance like find the longest palindrome not containing the letter c or containing the letter c, it does really poorly. Btw n^2 is not the optimal complexity for this problem, there are many solutions in O(n log n) and O(n) using hashing, mihasher and other string algorithms,

So a public demo version of ChatGPT easily passes a google coding interview and people are still saying that there is nothing to worry about lol.

FYI, you can hit the "Try Again" button to retry the most recent question without having to restart the whole thread

this was fun to watch. thanks for doing this video