OpenAi o1 is Amazing - It will have a HUGE Impact on EVERYONE in Tech
Вставка
- Опубліковано 17 вер 2024
- OpenAI has surprised the AI community with its latest model, 'o1,' which has the ability to "think" before responding. OpenAI claims that it even surpasses human performance on PhD-level science questions.
In this video I briefly introduce you to o1 and then test it with some programming questions!
---
Twitter: / garyexplains
Instagram: / garyexplains
#garyexplains
Very appropriate that the snake game was written in python.
Indeed.
i can just imagine Boeing engineers cutting and pasting code into their new plane. :-)
thereby improving their safety record
They have been doing this for decades! Cut and pasting known good code is common practice!
@@johnjay6370 , reduce, reuse, recycle...every programmers does it. the problem comes when you integrate them together...sometimes you need to modify them slightly, other times you don't fully understand the codes so you treat it as a black box. cost cutting put pressure on engineers limiting their time and efforts.
@@test40323 It will be a problem and it is most likely already happening! The reality is that AI will cause engineers to get lazy and soon and I mean soon, programming will be all "BLACK BOX". I see in 5 years most programming jobs will not be programming like we have it today, but more of directing and reviewing. That is until we find that AI does a better job than people! And i am 100% sure it will and this will happen soon. Another thing is that the market will push us towards AI programming as the number of computer programming degrees goes down. The question i ask is to people "would you let your kids go into a programming degree?". 5 years ago i totally would, today i would tell them to look into other fields of study.
Anyone else talking about Open AI, generative AI, and I'd have scrolled on by. Thank you Gary for not turning your channel into a typical UA-cam cash grab for ad revenue.
That is a high compliment, thank you.
Great ai scam always causes additional scams in youtube.
No thank you Joseph. Just thank you for being you 🤮
It's just sending multiple requests to the LLM - you could do exactly the same yourself. Except that you'd know you were sending, say, 5 requests and would therefore expect to be charged for 5 requests/responses. o1 hides that from you and you are therefore susceptible to multiple additional costs without realising. In the EU, this is probably illegal and could be construed as a scam and price-gouging. Do this in Openrouter, etc., compare with another LLM of similar charges and add up the actual costs. This has been checked and confirmed by an AI writers' group. Why do you think they are massively limiting the number of requests in ChatGPT currently (for this LLM)?
maybe it's time we try our luck with that app idea we all had.
😢Are we there already?
@@seetsamolapo5600 we might be, we won't know unless we try..
I just tested it, give it network code to optimize - gpt o1 refactor was 3 times slower than original implementation.
It’s STILL just making shit up as it goes based on what token is likely to come next.
A bit like customers’ coding really.
That’s my experience too
Great video. I took it down a maths route and got broken down steps solving a quantum mechanics question and could question parts of the result. If you are doing research, it is going to be a new best friend.
Thank you Gary. I have read various papers on this but you really nailed it with your explanation.The chain of thought capability with RAG will be interesting to see how this evolves.
Be much more interesting to see how it would tackle the problem of writing the snake & space invaders game in VS COBOL II…
I tried mathematical expression with standard gpt for TCL and took 22+ tries with error resolving to get it working. looks like a real improvement :)
In "one" try?
@@ThePowerLover yes and gave error(s) back to "solve" the issue. after 22+ corrections it was okay and working
Currently this model does not allow uploading data or files for analysis, unless I am missing something. Will this change going forward?
That is correct, at the moment it isn't multi-modal. It will be eventually.
Being able to see its thought process is invaluable. I have ADHD and struggle to keep an organized train of thought most of the time, but reading how o1 thinks about my questions is already helping me understand considerations I miss. Going to work with this to not only help my work but help me think better as a human!
Don't know why there are so many complaints in your comments. This is miracle technology that I can't imagine taking that for granted. A lot of the tech community needs better perspective on how truly incredible it is that LLMs even exist, but that's just my grateful ass :p
They specifically say they are hiding the thought process though?
@@chrisjsewell The actual thought process, but o1 already does that for them, so there are at least two layers of "imprecisions".
o1, can you please write me an AI, it must be called o2 and be superior to you in every-way.....
Is it significant that the 'thought process' explanation is listed as 'I'm' (doing this) rather than saying 'This task requires'
At the same time, because "OpenAI instructs it", it says it doesn't have personal interests or emotions.
Very interesting and very nicely, well, explained!
It is a disappointment actually. A lot of programmers had been doing exactly that. We are actually waiting for a new model that is more powerful than current model and not just a well-scripted GPT-4o.
I used to achieve similar results in GPT 4 using prompts. I would ask it to go through it's chain of thought process, list down the steps etc and in the end Answer in most cases would be much better. Interesting to see it's doing this now by default.
That’s apparently the general idea, the “explain your steps” _is_ the reasoning. Also more and adaptive computing power.
I like Claude, I don't think these are major improvements. I think the new approach is just ENHANCING your prompt. Recently I needed a series of calendars and asked it to create a .SVG. I gave up. They had the years at the top, the months, but only a couple boxes, and only the first row had dates. It also forgot which months I asked and gave me a whole year. This was on Bard, Chat GPT and Claude. Claude made a very nice "artifact" calendar, but could not make the SVG so I could print it.
Claude is great as it can create and display charts/graphics(SVG or the other one) as well as html pages(even with javascript libraries) in the browser. With ChatGPT I have to copy and paste.
I use both by the way.
I would like to see how it performs to program for piecewise differential equations
I don't know if I should be excited or terrified...maybe I should ask o1?
''it takes time before it can say I am''
My favorite tech youtuber :)
Very kind of you!
Can't wait for everyone to stop pretending that this is new. GPT 4 could do all this
No it couldn't. I literally just tried the "write code in lua that can parse and evaluate a mathematical expression like (4+5)*2+1" and the code made by ChatGPT-4o doesn't run.
And just for completeness I tried it on Claude 3.5 as well, that code didn't work either.
A human made that snake game, if user asks for snake game here is the code saved in db, for its 'thought process' its just regular gpt being asked 'please list steps that you would take if you were building a snake game in python'
Yes, if I was pretending that I'd done it myself I'd probably give a list of actions to justify myself in order to avoid legal issues . I'd be interested to see how it copes with my recent experience identifying images with cows when trying (unsuccessfully) to set up an email address with my internet service provider which is seemingly impervious to any form of intelligence when it comes to answering customer queries (the chat-bot "says it all" & nothing).
And at the same time, you call some people "conspiracy theorists"...
Please bring back Speed Test G
All the coding demos on AI seem to be a total waste of time, it is just reproducing something that is readily available and not doing it very well. Get it to do something useful, see if it can actually do something productive. I have seen little evidence that any of these AI models can do anything innovative or even deal with anything more than the simplest specification.
Have you tried for yourself?
@@GaryExplains No, which is why it would be useful to see something closer to an actual development scenario, rather than a couple of pretty poor attempts at well known arcade games or an eval() recreation. Give it a specification for interface to a web service and get it to create an abstraction to interface directly with an existing system that has a well defined data structure. If it can do that I will be impressed and spend time and money using it.
🤷♂️
So much to unpack. It's not going to be putting any professional programmers out of a job any time soon. The gulf between human intelligence and computer plagiarism is massive. And will remain so. Reminds me of COBOL, write English text, computer does something - ideal for those that aren't programmers and want to create lifetime maintenance associates to make it work properly. Can it fix its own bugs at 3am when the mission critical app has just crashed. Please make me a system that will result in maybe a million lines of custom code and tens of thousands of persisted data structures. When will the programmers at openai no longer be needed. If I ask it to make me a clone of o1, with all the code and data I need, but much better, how do you think it will perform.
Some of the unearned enthusiasm needs to be tempered with reality so as not to create false expectations. Its obviously amazing how far machine learning has come over the past 40 years.
What do you think the timeline might be for it to be able to self-improve?
This model was designed to help those into science and engineering. It lacks some of the ability of 4 omni, for now.
Great video! Could you please break down the apple A18 pro chip vs the A17, also the same for s9 vs s10 chip in the apple watch! Thank you
Why do they all try to make a game? A snake game too...
I tried more than that, no?
@@GaryExplains never said you didn't. Just sick of seeing snake. They probably tune these models for snake now... and space invaders etc
Passes phd level. Suuuuuure.
I watched a video yesterday of a PhD in physics testing it and he was blown away by its calculations.
Maybe next time he won't ask it to make an atom bomb :-)
Gemini been beating that azz!
Is there anyway to upload eda tool user guide and command them to make script ?
Nikhona South Africans? 🇿🇦🇿🇦🇿🇦
Interesting. Thanks. Just wondering how the thinking time is established. Is it dependent on the number of users? I don't think there is so much 'thinking' involved of supercomputers. Maybe the timing number is just a gimmick..
The thinking time is directly related to the complexity of the problem not the hardware or the demand.
@@GaryExplains Thanks for the response. Let's say the 'thinking' time is 20 seconds as in one of your examples. What can a computer(park) not do in 20 seconds? I still have problems with it. Are the computers doing some internal reasoning for 20 seconds!? Note in the past Google told you how long a search action took. You got figures like 0.1 micro seconds (or less).. Also hope that not too many people ask complicated question at the same time. I think it's more gimmick (of course it takes some time).
@@aamiddel8646 a search is just doing a pattern match within a usually indexed database. That's all it is. AI is nothing like that, so the comparison is totally wrong. If i were you i would investigate how AI and LLM works and then you will understand why it takes 20 second.
As I said, the thinking time is directly related to the complexity of the problem not the hardware or the demand. A more complex problem needs more iterations of "thought". Each iteration is likely an inference shot through an LLM, this continues until o1 has a strategy to solve the problem. However we don't actually know any of the details of how it actually works.
@@xeon2k8 Ofcourse i know that Google is a search and therefore different. I just gave it at an example how fast computers can do there thing. And to tell me to investigate how AI/LLM works is lowsy suggestion. If i knew this i would not foillow GaryExplains on an AI related item. But hey. You tell me how it works. And BTW than also explain to me how quantum computers work.. And about that is unrelated to hardware is not logical to me. If not does my old TRS-80 model 1 with 16K Level 1 then also takes 20 second for doing this job?
I gave it a programming problem that I received in an interview the other day and it made a meal of that.
Interesting. Could you share the question, or one similar if you can't share the original.
Please do share.
It was something like finding triplets in an array whose sum is divisible by a given number and that the triplets must be in increasing order of index. It took 60 seconds to come up with a solution which was incorrect, then when I pressed it took another 60 seconds and came up with a 200 line overly complex solution which I didn’t bother to check.
Something like this?
Given a 0-indexed integer array arr[] and an integer d, return the count number of triplets (i, j, k) such that i < j < k and (arr[i] + arr[j] + arr[k]) % d == 0.
I asked o1-mini the question above and it thought for just 2 seconds and gave a 70 line c++ program that worked as given 🤷♂️
clarifying "disallowed content" means it's now a political tool.
From where do you have this clarification?
Ask it to write a clone of Microsoft Word.
That’s ridiculous. It can’t output that much code and to even make an attempt, it would either need access to Word’s source code or some type of interface to iterate through the various features and capabilities that MS Word has.
Neat
As for writing codes for some specific and complicated tasks, i still don't believe AI can help me.
Have you tried? What specific tasks do you have in mind?
Complicated tasks can be broken down into set of simple tasks, and AI can definitely get set of simpler tasks if not a single complex tasks
Looks like a built-in RAG service
very different, retrieval augmented generation is the act of vectorizing the query, and doing a similarity search on it with an outside source (like a vector database or documents), to augment the prompt with, this is chain of thought on steroids; reasoning, and planning recursively feeding into itself as reasoning tokens to reach a final conclusion.
@@lifeofcode You are right, though what sometimes happen with RAG is that there's little to no augmentation, just repeated querying and breaking down a problem to multiple parts, or using other methods to rank or filter and then query.
👍
AI is creating an entire generation of incompetent lazy "programmers" . This is NOT a good thing.
That's what low level devs said when high level languages came along. I'm not disagreeing with you, just putting it out there that the new norm could replace the old norm quite rapidly and convincingly with maybe a few small hiccups along the way. This iteration of AI is here to stay and will only get more capable with time. It will grow with us as we grow with it in a sybiotic way.
You can dress it up how you like but it’s simply the application of computing to make computing easier for everyone, as has happened since the invention of the computer. Had exactly the same thought as the OP when I saw wysiwgy HTML editors for the first time in the mid 90s. I had this fairly rare and somewhat valuable knowledge: I knew HTML. Now any Joe could do what I could do.
Yeah the quality of output will go down. Just like it always does when “shortcuts” are introduced.
There is so much to build and so much time / money to do, Those type of tool can really make a difference on tech worker productivity. Also the code is not not modifiable and you can chose to use it or not when it matter.
Software devs will become obsolete.
Pivot to prompt engineering.
No. They will just be a different kind of busy.
Here we go again
Software devs will not be the first to become obsolete, they are maybe some of the last to become that. Plus they can always become data engineers, as 80% of an AI project is prepping the data.
@@MoroZ4300why would a company hire a software engineer when there is an abundance of data science engineers that are specialized in data analytics and can do a much better job more efficiently? Software engineers won’t just be replaced by AI, they will also be replaced by other specialists that are more suited for the job. Good software engineers (the top 5-10%) will still have a job and will continue to be handsomely paid, but the remaining 90% of empty brain leetcode bros will get the end they justly deserve (fingers crossed 🤞🏽)
An intelligent person does not try to answer a question that he or she knows it cannot answer. If it never comes back with a response which states that it cannot give an answer... [Continued on page 94] ps Can it cope with cursive? Can you ask it to "Spare me the details"? Is it free?
Continued on page 94... Where did you cut and paste this from?
Level 1 achieved
I think OpenAI was hoping for level 2!!!!
Level 2, actually.
Well, Im a bit surprised by the sensationalist headlines. Gary is normally more serious, tbh. I tested out o1 and it was quite mediocre, was slow and felt like a marginal upgrade. Coding tests was poor, not better than claude 3.5 and slightly worse(!?) then chat gpt 4. Maybe it's due to it being untrained and a 'preview', but it performed poorly. So, I'm not exactly sure what tests Gary really put it through. Just because it shows some path of thought, that is LLM at work. I'm holding off on passing judgement what type of 'upgrade' this will be as so far: not impressed.
Sorry you didn't like the headline. What tests did you perform that o1 couldn't do? Could you please share as I would like to try those myself. Currently I am working on a project with o1 and it is coding a project of over 700 lines of C that compiles every time. Very impressive.
@@GaryExplains Thank you for the reply. I feel it's unfinished. and don't perform much better than gpt4 and actually worse than Claude on coding specifically. I test out these tools for integration with work, as a 30+ years of engineering and code automation specialist for close to 20 years. So a bit experience in the field. I prompted basic things, like a browser based code, one was for a small game, and it generally ran, but not always. It was no better than claude, nor chat gpt 4 and the number of operations was many times higher, hence more expensive. When i expanded code base to 2000-3000 lines of code, it got confused and hallucinated, just like Chatgpt and Claude, which as you know, 2000-3000 lines of code is just a tiny hobby project, not professional assistance as you suggested. So its not meant to be mean - but as a fellow professional, I am just not liking the fact that corporate leaders seem to believe they can replace the vast majority of programmers - which is far far from the reality. Like I stated: 30+ years of engineering experience and developed my first neural network in 94 as part of a university grad program. So, a bit of experience. Just my two cents, didn't mean to be disrespectful, if it came across such, just wanted to point out that engineers - at least those with some experience, in my opinion, are still safe.
@@Ad434443you protest a little too much we think. Listing your credentials (we all have similar) is irrelevant if you don’t provide exact details of what didn’t work.
So. Give some concrete specific examples. Otherwise I call bull.
@Ad434443 Thanks for the follow-up. Reading your reply, I see that the issue is actually "I am just not liking the fact that corporate leaders seem to believe they can replace the vast majority of programmers - which is far far from the reality." And I agree with you on that. But I didn't imply or state anything like that in the video or in the headline or in the thumbnail. The reality is that o1 and its successors will have a huge influence on everyone in the tech industry, just like ChatGPT has had until now, but more so, and if OpenAI keeps up this pace of development, the impact on the tech industry will increase. Imagine where we will be in 5 years from now.
@@GaryExplains Maybe o1 likes you...
I might be totally wrong but I am starting to feel that AI is over the top as a buzzword, going down fast. Hyped too fast.
Yes. It is going down fast. 🤣
lol what's this clickbait title
What is clickbaity about the title?
@@GaryExplains don't play dumb, you know what you did
Honestly I don't. O1 will impact everyone in tech, just like ChatGPT 4 did, but more so. You don't think it will?
@@GaryExplains oh please, stop playing dumb. Or as you would put it to try to attract clicks: "STOP playing DUMB for EVERYONE in TECH!!"
I dont get it, why the clickbait title? Kinda didnt expect this kind of s**** from this channel.
Really? I am sorry about that. I see o1 as a huge advancement it will impact everyone in tech. I am working with it now to write a complex program, it is currently 730 lines of C code and it compiles almost every time but with the feature added I requested or bugs fixed.
@@GaryExplains Excellent ! Will love to see that through one of your next video. It's really fascinating as this approach seems to be the perfect match for humans to "enhance" their potential, their ideas, using AI assistant. The term "assistant" taking all his value here.
@@GaryExplains He has a different kind of expectations. Either he is too smart of a programmer or he doesn't understand a lot about computer programming.
I personally think it is not clickbait, as o1 is announcing a new generation of "LLM" that would be good in particular in the tech worker, and in the video gary show that it may be true.
Programmers are useless now.
Not quite, but the trend means coding will be less of an important skill compared to software engineering and system design.
@GaryExplains I hope this won't affect embedded programming, since it's tied to a particular hardware.
@@blancheherminewhy wouldn't it be affected? If you write your embedded program for your specific hardware using knowledge that is publicly accessible on the internet, the AI can do it
Just tested it out, I am a dev with 15 years experience. We are done guys. It's over and great while it lasted, thankfully I made a ton of money. Turn out the lights on the way out.