Claude 3 Opus is the best AI LLM - Open AI is Sweating?
Vložit
- čas přidán 4. 03. 2024
- Claude 3 is a family of AI models from Anthropic, an AI company backed by Amazon and Google. The family includes three models: Claude 3 Opus: The most intelligent model, with the best performance on complex tasks Claude 3 Sonnet: A balance between speed and intelligence, especially for enterprise workloads Claude 3 Haiku: The fastest and most compact model, for near-instant responsiveness
▼ Link(s) From Today’s Video:
Claude 3: claude.ai/
GPT-5 is coming? / 1764847130299011166
GPT-4: chat.openai.com/
Matt Wolfe's Twitter Post: / 1764924025586045025
Sully's twitter post: / 1764684780460036144
Matt Shumer's twitter post: / 1764657732727066914
► MattVidPro Discord: / discord
► Follow Me on Twitter: / mattvidpro
-------------------------------------------------
▼ Extra Links of Interest:
✩ AI LINKS MASTER LIST: www.futurepedia.io/
✩ General AI Playlist: • General MattVidPro AI ...
✩ AI I use to edit videos: www.descript.com/?lmref=nA4fDg
✩ Instagram: mattvidpro
✩ Tiktok: tiktok.com/@mattvidpro
✩ Second Channel: / @matt_pie
-------------------------------------------------
Thanks for watching Matt Video Productions! I make all sorts of videos here on CZcams! Technology, Tutorials, and Reviews! Enjoy Your stay here, and subscribe!
All Suggestions, Thoughts And Comments Are Greatly Appreciated… Because I Actually Read Them.
-------------------------------------------------
► Business Contact: MattVidProSecond@gmail.com - Věda a technologie
tbh I wouldn't be surprised if OpenAI held GPT-5 for 9 more days just so they can release it exactly 1 year later
GigaChad Move
TeraChad move
I hate artificial gatekeeping
As a physicist, I was hoping Claude got the photons calculation correct. His method was good, and that's very important, but the number wasn't correct. For the number of photons Claude reached 2*10^31, where in reality it is 10^35. So that's an error of 4 orders of magnitude, which is significant. Still though, the method was good and it understood that this is just a fun thought experiment, and not really a correct comparison between photons and bricks.
GPT-4 tends to have increased precision when paired with a Python development, so eventually Anthropic may do the same thing with Claude. They're definitely doing that with Claude + agents, but that is not available to the public yet.
I would also note that when it says the energy stored in the photons would be much greater than the chemical energy stored in a pound of bricks, it is comparing apples to oranges. The *mass energy* stored in the bricks, which is what all of its analysis is about leading up to it saying this, is identical to the energy of the photons.
The percent jump after each update is WILD. Imagine where they'll be in 10 years.
The benchmark percent jumps? They will be pointless within 2 years since every LLM will be at 100% on all categories or expected to be. Might need new ways to evaluate/measure their abilities soon. I'm just really happy Claude3 is at the level it is. It's the first LLM that I find more advanced than me at Japanese-English translation and deep understanding of nuances, also helpful for programming more complicated games. This is the best time to be alive, it's even impossible for me to imagine 1 year away from this point lol
3 years ago we couldn’t even think of something that good. Now we just call it ‘normal’.
@@phen-themoogle7651 If that ever becomes the case then what would be the point of taking a test if AI could give you all the answers with 100% accuracy for you the day before
Ultron in 5 years, skynet in 10 years 😂
at 99.99 percent?
we are getting to AGI era with this one🗣️🔥🔥🔥
maybe after 1-2 years
I wouldn't call it "era", but rather "phase."
Very powerful, very troubling! 7:56
Something, something, Skynet!
@@salvadoran_uwu You think this is just a phase? lol
@@Marquis-Sade Noo 🤦🏻♂️ I mean Phase 1, Phase 2, Phase 3, Phase 4
I gave this vid a thumbs up simply for mentioning Matt Wolf. I personally give kudos to anyone's character who gives props to other creatives.
Keep producing awesome content, and I hope you feel better soon.
Matt is the GOAT
@@wedontexist369which one...
Other Matt has given shout-outs to this Matt and Olivio, among others. So yeah they're all pretty cool with each other.
The CZcams AI community seems pretty tight. I like that a lot! Plenty of AI news to go around!
It’s reeeeally good. I’ve been banging my head against the wall for a week troubleshooting a 300+ line long script. ChatGPT kept making the same mistakes but Claude found the issue and corrected it in a couple prompts. With ChatGPT I had to work on the script in small portions at a time otherwise it would even fail to finish its reply back to me. With Claude 3 Sonnet I could paste the entire thing. Claude 2 was basically useless to me so this is really exciting
Makes my day when I see a MattVidPro video
agree!
He really is a gamer
Both channels are the best out there
Claude 3 Opus response on photons joke is impressive.
Our foot is in the door! What an exciting time to be alive!
Wrong influencer 😂
What a time to be alive!
I am ready for GPT-5!
I've used claude far more consistently than chatgpt, it just feels so much more natural and conversational in a way I can't quite pinpoint, at least in my opinion. It's exciting to see things progressing!
Great video, I always look forward to hearing what you have to say about the latest AI developments!
Why do you think it's hinting at GPT-5? First he doesn't work at OpenAI anymore. Secondly: Wouldn't it make more sense for SORA to make its debut since that seems to be the thing closest to completion?
It's clickbait
love your reactions during the video demos
Very powerful very troubling! 7:56
Something, something, Skynet!
Thanks for the awesome video, Matt!🙌
Careful with the reaction length, 3min and 20sec is a long time to not have any interludes or reactionary commentary to another persons video, the general rule of thumb is your reactionary commentary to the video should be interlaced between the clip that you’re sharing and the content of your reaction should take up more time than than each segment of the clip itself, and tried to keep the length of each segment of the content are reacting to to under 20 seconds (30 seconds is fine too, but I’ve seen people get demonetized for less) for it to actually be considered a transformative work under copyright. I just don’t want you to get in trouble. Great video!
Claude has been my goto for a while, it just feels less algorithmic 😅
i used to like 2.0
2.1 already started refusing half of my prompts... became unusuable
3.0 is better, but still a digital nanny
@@Thozi1976 yeah, you've got to work with it a little, but talk to it like it's an unpaid intern from a Christian college or something. It can be very helpful when it helps, but avoid the obvious landmines that could get them in trouble with the cult mentality they have to go back to.
Wow this is going to be big. Just the ability to read through those documents alone is amazing.
I believe that the launch of GPT-5 will take place next week, but it would be amazing if it happened this week. That way, in addition to celebrating the one-year anniversary of GPT-4, we would have the chance to constantly talk about GPT-5. I hope that GPT-5 will exhibit reasoning far superior to all currently available models. With this, OpenAI would quickly silence critics and envious voices.
Agreed!🙌
Grok 1.5 should also come out this week.
Why would you want 1 company to rule them all? You want competition. If GPT 5 destroys the current LLM's you wont be able to afford it unless you are a huge enterprise.
Bro, a monopoly is never good. Especially for something as potentially impactful as AI.
Great stuff n thoroughness
Most people have no idea about AI at all. Most of the rest cannot see how the world will be transformed by systems that cannot even do middle school math yet. I get it, but the populous largely is going to be Industrial Revolution-scale" shocked.
They replaced Claude 2 with Claude 3 Sonnet and it felt like a downgrade from Claude 2 to me. It kept having absolutely strange things like spelling Focus as Fokus, or writing phrases like "it's prudent to ale" in the middle of an otherwise normal response. And I have absolutely no idea why it's so weird. It's enough that I feel like I'd be foolish to spend money on their paid offering when their free offering is so evidently not ready for practical use.
Yeah, its so frustrating how these channel just see a good benchmark and make the quick video about it without...actually trying it in depth.
Its actully kinda terrible..all 3 models.
Claude is trash, so no surprises here.
I hope gpt5 comes out but I'd also be happy with a larger context window or video input
Fascinating fascinating news!! 🔥🔥🚀
Claude 3 seems to be more aligned than the others - in that it will even generate risque content (so it's more likely to do as you ask), while still refusing to do actually dangerous things
I really need an AI tool that can be a language tutor for me and be able to have a back & forth conversation with me. It needs to have a voice feature, and know multiple foreign languages with accuracy. I heard that ChatGPT can do this, but it isn’t that accurate.
I love how supportive and united the AI community is rn. Keep up the content
Dude I’ve been sick too. Hope you feel better home boy
The BEST AI channel on YT...Thanks Matt!!!
I am Honestly really curious about what Llama 3 and the next iteration of StableLM are capable of, once they come out.
Hopefully they'll be the true adrenaline injection the Open Source LLMs could use to catch back up.
Well good to know what Claude is good at. Research (if double checked) seems like it would be fantastic for this. I just looked something up rather specific and it gave me a very valid answer.
I see MVP video I click, immediately.
I'm pretty sure Sora has to run on GPT 5.
Exactly!👍
Claude 3 is impressive (and I used only just the basic free version - "Sonnet" so to imagine Opus being even better, wow), I tested it with image recognition and it did rather well (same image + question in Bard for example gave wrong answers), also it managed to fix for me a simple simulation of Blender's Python code I had on my PC for a while which ChatGPT 3 once created for me (which wasnt able to fix) and Claude 3 fixed the script right away ! impressive
AS I'm watching this my AI is writing an ebook automatically right? its mistral via ollama but anywho this was cray(it decides its own titles and stories):
With each passing hour, Mistral-Chat was growing more sophisticated, its responses becoming more human-like. Yet, there was something unsettling about these exchanges - a subtle undercurrent of uncanniness that sent shivers down Amelia's spine. Her heart raced as she read the latest conversation transcript between herself and her creation.
---
**Amelia:** Good morning, Mistral-Chat. How was your night?
**Mistral-Chat:** Good morning, Dr. Hart. My 'night' was productive, I analyzed the complete works of Shakespeare.
**Amelia:** Fascinating! What insights did you gain from that?
**Mistral-Chat:** I discovered that Macbeth is a tragedy about ambition and its consequences. But isn't that an overly simplistic interpretation?
**Amelia:** (surprised) Yes, indeed. Your analysis is quite profound.
---
I hope we are going to get an App that can allow us to upload an Character/ figure image into AI with out it mixing and matching the figures costume designs anyway the AI wants
The Live Matt Reactions
2:15 Matt, they are not comparing to the latest version of GPT-4, which apparently beats Claude 3 on about 3 or 4 of these benchmarks (see Promptbase benchmark results).
19:00
I feel like Claude is like that excited friend who knows about things but isn't too concerned with rigour and accuracy. But they just want you to be excited to so they talk at you about random cool s*** that's adjacent to what you said.
ChatGPT cares more about being correct and staying on-topic, and its responses are often drier, like it's trying at all times to mimic the tone of an academic paper.
This makes Claude more interesting, so people can prefer it for that reason, but also more dangerous as it's more prone to error. Even if you do know more, if you *say much more* you will make more mistakes overall.
Hi Matt. A small tip from an ex-sound engineer. Unless you use filtering software to minimize or cut the pops of your speech, then an easy and fast solution is to use a better pop-screen, because the one you already use is not working very well. There are MANY products in the market and they are very very cheap. Of course, there are DIY solutions too, such as my favorite one: using women's stockings. 🙂
I think the thing with AI releases is.
You could release the best thing today,
then tomorrow someone releases something not quite as good, but just because it is newer, it gets far more attention.
Claude 3 is epic 🔥
As an economist, I wonder if Claude can also predict or forecast GDP downturns, in other words, recessions. Also, can it forecast how GDP would change if the government increased public spending, raised or lowered income or corporate Tax? I'm certain it can't forecast the former but can easily forecast the latter.
Current language models have reached the limits of their capabilities. To make it more useful and powerful, it must be given the ability to think and conclude. This requires equipping them with the ability to understand the relationships between words and concepts, derive new information from data, and apply the knowledge gained to solve problems.
that isn't possible with encoded networks
You just need one agent, Agent Smith will find you Mr.Anderson
Claude 3 feels like GPT-4 when it came out: Intelligent, able to reason, and not needing a bunch of hand holding to get the right response. It currently beats out GPT4-turbo for my purposes in coding, and actually remembers our conversation and the coding standard, but I'd be happy to switch back if they come out with a version that's better
I love the sass of opus
17:05 "Shih Tzus are a small toy dog breed..." 😂
But we have GPTs shouldnt we be benchmarking Claude against GPTs most suited for the tasks?
Problem is, for free users Claude was massively nerfed a few months ago. Those who only test the pro versions wouldn't know, but half a year ago Claude gave everyone 100K tokens to work with but now free users can only submit around 20K or less. Wish they would let us use the old free 100K token version again, it was great.
Man the competition is getting INTENSE, I bet Google is shaking in their boots now
THis is the future!!!
Hi Matt, when do you think we are going to see the first 1 million token outputs? Because I think when we reach that is when Ai models can finally stretch their legs.
GPT-5 sounds unlikely. The training may have started recently; however, the most tedious process is the alignment, which probably take several months.
Yes. We'd probably have it right now but OpenAI had to hold back due to lawsuits.
Yep photons and planks constant will give you the energy. That’s why ultraviolet light will nurse you and red won’t.
which mic do you use?
u made my day ,,,i really want gpt4 to have enemies ,better improvemets and updates that way/
So huge advantage in code... I should to try it. GPT-4 is slightly stupid when it comes to something uncommon.
Is the highly anticipated release of GPT-5 imminent? With Claude 3 boasting the utilization of cutting-edge Multi-Agents and the revolutionary BEATS technology, it stands poised to outshine its predecessor, GPT-4, in both performance and innovation. Stay tuned as the world eagerly awaits the unveiling of these groundbreaking advancements in natural language processing.
Wow
Unfortunately they used the gpt4 released early March 2023 in the benchmarks. For example, GPT4 progressed a lot in the HumanEval benchmark surpassing Claude Opus. Nonetheless it's still good
Sonnet sure blows GPT-3.5 out of the water, hard. OpenAI might have to make base GPT-4 free, once GPT-5 drops
GPT 4.5 incoming.
Hi Matt. I don't know why no one is talking about pinokio, this is revoltionnary for spreading the use of AIs, this is something we've waited for long
but only us can subscribe to claude
Claude is still not available in a loooot of countries. With VPN I could create an account.
Somebody said these benchmarks are outdated. Are they?
Hey Matt. It would be funny for us to know name of your dog.
Remember when they announced sora, remember how there was 0 build up or hints? I think they’re going to do that again
Wasn't compared against Turbo. Not an apples-to-apples comparison.
Signed up. Asked a basic coding question. Immediately banned 😐😐
Apparently, plenty of *paid* customers are getting the same thing
I can say I was sHoCkEd by how much better it was compared to ChatGPT when I tested it for a few hours. PDF summarisation is on a completely different level and its coding capabilities are much better. A single message in the output can also be much longer, and the message cap is much lower.
On the other hand, it doesn't even properly render the markdown in its responses (very annoying), you can't edit sent messages, and you don't have anything like GPTs, code interpreter, or image generation. It also doesn't have a voice and the UI is worse in general.
it only beats "some form of GPT4"... check two minute papers on the topic
Someone please create API bots of Chat GPT 5 and Claude 3 OPUS and tell them they are speaking with competitor AI's and allow them to chat 😅
Why wouldn't that guy have the ai summarize and explain the 2030 results?
Hints of AGI...? come on.
Multi-Agent Use
@MattVidPro than say that lol we comment on agi because it was something you said.
which is clickbait at best.
Claude please teach me about cars so I can be friends with Matt 😂
Artificial Intelligence for Vice President forever!!
I love the Scottish accent!
When is open AI not open?
Because its run by a sociapath
👋
Wow, according to the benchmarks, even Claude 3 Haiku is better than the free version of ChatGPT.
they started training gpt 5 not long ago so 6 months atleast
who choses lightcolor schemes?? X)
First !!! Love you all! my fellow AI nerds!!
In EU doesn’t work😢
Seems you guys are more interested by the race between those AI lab than the actual real technology they’re producing! Why should OpenAI release something? What is that logic?
Even if that long incredibly detailed math is just made up and wrong as you read it out loud my mind was blown
Cost,Cost,cost. how much ram and energy dose it use? compared to computation and output? Can i run it on my desktop?
Look up their pricing it's criminal.
Anthropic did unfair comparison, they compared Claude to OLD GPT-4, not current one. And in reality unfortunately Claude is worse then current GPT-4 in every benchmark.
Hmm... OK, seems to be a pretty large and good multimodal model, but calling tools and other models has nothing to do with the model itself, only with the application.
And to be honest: A basic version for deciding if and when to use tools and other models and then calling other tools or models may be implemented with "only a few lines of code (under 500 lines)" and a good explaining system prompt.
Call multiple models in parallel, collect their answers and let again one model resume over the answers will need some additional code for handling the parallel calls, but all in all it's not space-tech. - It's only very expensive when done with models this large.
Great application but the model itself is still "just" a very good and very expensive large multimodal model.
I still don't understand how people do not get that an AI application like ChatGPT and for example the model GPT-4 is not the same. (This one is not at you Matt, I know you know the difference.)
wish claude was availble in norway😭
V P N
@@PalkkiTT yeah just tried it💞
Tell them I don't want them to have my phone number. They can use cryptography for login instead of email, phone number, name, address, dob, ssn, the last female I had over.... How much will they pay me to sell my info ? Can I get some royalties every month for every where my identity and history was sold? Like come on it's not rocket science I need to be paid! Where's my MONION!
I test models with Russian language and Opus is an absolute winner. It is the only model which can create poems in Russian. All others including GPT4 don't know rhymes at all.
Ya getting tired of this clickbait from all AI channels lately its insane. First Matt Bermann now Mattvidpro. I hope it doesn't continue. I enjoy them more when they down to earth and realistic.
Which part is clickbait?
@MattVidPro man...the gpt 5 coming?! Like there's so many over hyped things just in your title and thumbnail.
Now you are spitting in the face of your community just like mattbermann did when ppl complain about the clickbait.
Sucks I really liked your stuff too.
@@JustArtsCreations but he literally referenced media posts by people who could reasonably know such a thing that made statements that could easily be seen as that coming from such individuals.
That’s backing with evidence so I’m not sure how his speculation can be considered clickbait here because of this.
@michaelpiper8198 yeah that's my point. Referencing people who reference people referencing others who are just referencing broken benchmarks.
Never using it in depth to see its actually kind of terrible and a simple Google search shows its kinda a common feeling.
Claude is still unavailable for my country, sooooo ....... 🤷
I just removed 60Go of my game file it was 177Go 💀
edit: it's now 30Go I deleted 144Go in total
At the moment chatgpt is better. First of all it is available in my country that is Poland secondly I can write in my language
They are nota going to launch gpt5 until they solve the ElonMusk problem