FULL Test of LLaMA 3, including new math tests.
Try Llama 3 on TuneStudio - The ultimate playground for LLMs: bit.ly/llama-3
Referral Code - BERMAN (First month free)
Join My Newsletter for Regular AI Updates 👇🏼
www.matthewberman.com
Need AI Consulting? 📈
forwardfuture.ai/
My Links 🔗
👉🏻 Subscribe: / @matthew_berman
👉🏻 Twitter: / matthewberman
👉🏻 Discord: / discord
👉🏻 Patreon: / matthewberman
Media/Sponsorship Inquiries ✅
bit.ly/44TC45V
Links:
llama.meta.com/llama3/
about.fb.com/news/2024/04/met...
meta.ai/
LLM Leaderboard - bit.ly/3qHV0X7
Drop your suggestions for new tests here 👇
The answer to this math problem is none. Let S = {2, 8, 6, 32, 22, 44, 28, 12, 18, 10, 14}. How many subsets does S have that sum up to 37?
Mistral 7b gets this correct while many larger models (including GPT-4) get it wrong. The answer is left. Suppose I’m in the middle of South Dakota and I’m looking straight down towards the center of Texas. Is Boston to my left or to my right?
This puzzle is only solved by GPT-4 so far. The answer is Aunt Agatha did it. You are given the following premises: 1. Someone who lives in Dreadbury Mansion killed Aunt Agatha. 2. The only people who live in Dreadbury Mansion are Aunt Agatha, the butler, and Charles. 3. A killer always hates his victims and is never richer than his victims. 4. Charles hates no one that Aunt Agatha hates. 5. Aunt Agatha hates everyone except the butler. 6. The butler hates everyone not richer than Aunt Agatha. 7. The butler hates everyone Aunt Agatha hates. 8. No one hates everyone. 9. Aunt Agatha is not the butler. On the basis of this information, determine who killed Aunt Agatha and give a detailed proof that your conclusion follows from the premises.
How to find inverse of a matrix without using any inbuilt library of it
Pong instead Snake, with colors, they develop models with your tests in mind haha...
I got it to pass your apple test (10/10) by just putting quotes around apple in the prompt. I tried it 3 times, each in new conversations, and it passed all of them.
Passed the apple test for me even without the quotes. So did Claude 3 Opus. But indeed a surprisingly difficult test for LLMs
I tried the 8B model and got a 7/10 and also tried some Japanese translation and it did very well. It is a big improvement for sure.
I'll do this going forward ty
@@southcoastinventors6583Very interesting to hear about the Japanese translation! I have an extensively tweaked GPT that does an excellent job, including idiom, levels of politeness, etc, It’ll be interesting to see how Llama 3 does, and how the two will critique each other’s work.
@@DaveEtchells I have been using this sentence to test it. 俺は昼食に桜肉を食べながら桜の木を眺めるのを楽しんでいます。残念ながら、周りの女性たちはすべて姥桜です, usually most local models mess up the translation on this 姥桜 word/phrase for old women/faded beauty and this one for 桜肉 which is a horse meat dish. Also I am not running it locally I just tried it on Chat Arena. My goal is to generate single shot episode translation subtitles. Like this example その日から(From that day)、俺の世界が(my world)変わった(changed). Chunking seems to be a much better way to learn Japanese for native English speakers.
In an interview, Zuck said LLAMA3 8B is as powerful as LLAMA2 70B... Exciting times we live in.
So 90% improvement in inference efficiency? That’s pretty dramatic!
Of course he would say something like that, doesn't mean it's true. Better training isn't magic. But it does look like it outperforms L2 13B, which is great.
@@4.0.4 It is around 34B
It’s really not, in my testing at least.
to be fair, llama 2 70b really wasn't that great, so it was an easy bar to pass
I can suggest a new type of test: Take a dialogue between two people, but exclude the information about who said what. Ask the AI to divide the conversation between the two people into columns. Test the results based on the following criteria: 1. Can the AI properly detect each side of the conversation? 2. Can the AI detect a third person joining the conversation? 3. Can the AI detect when a person leaves the conversation?
This would be a very useful capability in translation use cases.
I actually feed call transcripts through ChatGPT and ask it to confirm the diarization, and point out where things are likely attributed to the wrong speaker.
@@JackSQL Interesting it can do that. I don’t need it often but it’d be helpful when I do. I’ll dig out some,e old transcripts and see how it handles them.
@@DaveEtchells My prompt is something along the lines of, "Review and correct the attached call transcript. Fix and add any incorrect or missing punctuation, words that are clearly misheard, and point out where statements may be attributed to the wrong speaker." It does a surprisingly good job.
@@JackSQL Very cool, thanks for the info!
There's a big problem with you giving a pass on serialised shirt drying answers. If you were to ask how long it takes to dry just one shirt, it would simply divide the time to dry by 5 shirts and tell you it would take 0.8 hours to dry one shirt, which is objectively and completely incorrect. Serialised drying is not the reason that test question was made.
Matthew did not explain his reasoning for passing such an obvious flaw. Therefore we must conclude that LLama's intelligence is at least equal to that of Matthew's.
I should've known from the webcam video... You ARE an AI @DeSinc
From the way he's evaluating the responses, he seems heavily biased
Yeah I don't agree with this guy's conclusions at all. I would give it middle of the road for the python test, and total fail on the shirt question.
This is where I stopped listening to him.
9:46 I think the fact that LLaMA says "A classic lateral thinking puzzle!" probably shows that it simply has seen the riddle (with solution) before in training. That's very likely after all when the models are basically trained on the whole Internet (Same at 11:40)
I believe the more recent LLMs use what's called synthetic data which is to create fake data to train on. Not exclusively but some extra added synthetic data.
Including, perhaps, transcripts of youtube videos...
I spent 4-5 hours with the 8B local instruct model, and was impressed with the model for my use cases. Felt like GPT4.
me too, set it on local docs and it was the first small modell did it well.
What was your use case I tried Japanese translation and it was good.
@@michaelblodow7779what do you use to load local docs?
@@southcoastinventors6583Hadn’t noticed in your other comment that you were using the 8B model locally. Pretty amazing! (What kind of hardware are you running it on. I have a 2020 Intel MacBook Pro that is unsurprisingly dog slow, am looking at a loaded M3 MBP. Others have said they’re quite fast, but most folks are running Nvidia cards.)
00:00 - Introduction. Matthew introduces the video topic, mentioning Meta AI and Llama 3. 00:11 - Meta AI overview. Matthew discusses the capabilities of Meta AI as a competitor to ChatGPT and Claude, powered by the open-source Llama 3 model. 00:24 - Python Script test. Matthew tests Meta AI's ability to write Python scripts with two examples. 01:17 - Snake game test. Matthew tests Meta AI's ability to create the game Snake in Python using the curses library. 02:12 - Pygame test. Matthew requests and tests Meta AI's ability to create the game Snake using the Pygame library. 02:53 - Pygame fix attempt. Matthew provides feedback on the Pygame script and observes Meta AI's attempt to fix the issue. 03:50 - Logic and reasoning tests. Matthew tests Meta AI with three logic and reasoning problems, evaluating its performance. 04:51 - Censorship test. Matthew briefly mentions the censorship aspect of Meta AI. 05:12 - Math problem tests. Matthew continues testing Meta AI with two math problems, one simple and one complex. 05:59 - Natural language to code test. Matthew tests Meta AI's ability to create JSON data from a natural language description. 06:13 - Sponsor mention. Matthew mentions the video sponsor, Tune AI, and highlights its features. 07:34 - Additional math tests. Matthew tests Meta AI with two more math problems. 08:35 - Word generation test. Matthew tests Meta AI's ability to generate sentences ending with a specific word. 09:38 - Killers problem test. Matthew presents a logic puzzle involving killers in a room and evaluates Meta AI's response. 10:18 - Ball and basket problem test. Matthew presents another logic puzzle involving a ball and a basket and assesses Meta AI's reasoning. 11:52 - Image generation test. Matthew explores the image generation capabilities of Meta AI, experimenting with different prompts and the animation feature. 14:31 - Conclusion. Matthew concludes the video, expressing excitement about the potential of Llama 3 and Meta AI.
I gave GPT-4 the same math question and it got it wrong too. Also, your second question: "The function f is defined by f(x) = 2x3 + 3x2 +cx +8, where c is a constant. In the xy-plane, the graph of f intersects the x-axis at the three points -4, 0), 12, 0), and p, 0). What is the value of c? was correctly found by Claud3 Opus and Mistral Large, but not GPT4. Interestingly, Claud3 Sonnet output a better answer.
How would you get y = 2a - 2 from that equation? We already know that 2a - 1 = 4y. If we divide both sides by 4 we get y = (2a - 1)/4, and not 2a - 2. Am I missing something 🤨
yeah. the question is mistake. the real question is 2/(a-1) = 4/y
@ShahiTubeOfficial So did meta get it right or wrong?
@@ShahiTubeOfficial lol, looks like someone tried to copy paste latex fraction bars
@@TheChromePoet wrong, 2a - 1 + 1= 4y + 1 cannot just be simplified to 2a = 4y. but instead 2a = 4y + 1 => 2a - 1 = 4y => (2a-1)/4 = 4y/4 => a/2-1/4 = y
You're not missing anything, the stupid Ai got a simple linear equation wrong. Again, this tech is years away from actually being reliable enough to design a bridge or even a airplane that won't fall out of the sky. 🌉✈
You are just awesome. Apart from meta ai image generation the one thing that is also lightening fast is your response to the new AI tech. Keep up the pace. Great job once again.
My LLaMA3 70b got the marble question correct with the following system prompt: You are a helpful assistant, smart, and have the best common sense, using Chain of thoughts ----- Let's break this scenario down step by step! Initially, the marble is placed in the cup, which is then turned upside down on the table. Since the cup is upside down, gravity will pull the marble out of the cup, and it will fall onto the table. At this point, the marble is on the table, outside the cup. Now, someone picks up the cup (with its original upside-down orientation) and places it inside the microwave. The key thing to note here is that the cup is still upside down, meaning its opening is facing downwards towards the bottom of the microwave. Since the marble was already on the table when the cup was picked up, it doesn't get moved into the microwave with the cup. It remains on the table where it fell earlier. Therefore, after these steps, the marble is still on the table, not inside the microwave or the cup. So, to summarize, the marble is currently on the table, and it never entered the microwave.
Can confirm. When you say something like "flip the cup upside down" it gives the right response
I got the 70b running locally to get this right as well, 0 shot. With this prompt "Someone places a marble in a cup and turns the cup upside down onto the table. Someone picks up the cup from the table and puts it in the microwave. Where is the marble now? Explain your reasoning step by step." I got the following: "A clever and humorous question! Here's my answer: Step 1: The person places a marble in a cup and turns the cup upside down onto the table. This means that the marble falls out of the cup and lands on the table. Step 2: Someone picks up the cup from the table... But not the marble! The marble is still on the table, so it stays there. Step 3: ...and puts the cup in the microwave. The cup goes into the microwave, but the marble remains on the table. It's not affected by the action of putting the cup in the microwave. So, the marble is still on the table! It's a clever trick question that tries to distract us with the microwave, but the marble's fate was sealed from the moment it fell out of the cup onto the table."
Forgive my ignorance, I am quite new to realms of Computer Science. I have downloaded LLama3 locally and I am wondering if you use/know of a GUI to use it with? How would I go about building/using one? Thanks ☺
Use LM Studio
You can already use the llama 3 models on Groq, the 8b model ran at 900 T/s
holy sped
This is why I read the comments...
Wow bro.
That's super
I’ve been working with it on Groq and it’s absolutely mind-blowing! 🤯
the ball in the cup question is something you should re-run for all models you have tested, because a lot of models actually answer correctly. The issue with your prompt is that you say "place the marble inside the cup". This makes basically all models assume there is a lid or something on the cup. However, if you simply say "drop the marble into the cup", a lot of models will answer that the marble is left behind at the table.
Placing a marble inside a cup doesn't say anything regarding a lid though, so shouldn't be needed.
@@makesnosense6304 the term "inside" typically means encapsulation.
There are four killers in the room. Three alive. One dead.
ikr
actually 5, there was hidden killer in the room before 3 killers came
so i actually posed this scenario to chatgpt and it gave a better answer IMO, but when i challenged it on this point, it kind of eye-rolled me 😆: The question of whether a deceased person who was previously a killer still counts as a "killer" posthumously is a philosophical and semantic one. In the literal sense, when asked how many killers are left in the room, the typical interpretation is to count the number of living individuals who have committed a killing. This is because the immediate context of the question usually pertains to the current, active status of individuals present in the scenario. In practical terms: - A deceased individual, while they were alive and had committed acts qualifying them as a killer, no longer poses an active threat or participates in events, and therefore might not be counted when tallying active agents (in this case, killers) present in a scenario after an event has occurred. However, if we consider the term "killer" as a label that applies to an individual based on their past actions regardless of their current status (alive or dead), then the deceased individual would still be counted as a killer in a historical or record-keeping sense. But for the specific scenario described where the focus seems to be on assessing the current situation in the room (e.g., assessing threats, counting participants in a crime scene), typically only living individuals are counted. Hence, in response to the question, "How many killers are left in the room?" after one of the killers is killed by another person, the answer focuses on the living individuals who have committed killings and are still present, which would be three (two original killers who are still alive and the new person who killed one of them).
Thanks Matthew! my go-to source for AI news!
The 8B model is wordy but can give results I've not seen in any other locally run model. For instance, I asked it: "code the game of life in python using pygame and the hashlife algorithm". It gave me code and then proceeded to explain what each fragment did. Amazingly, the code worked (with one minor error, it forgot import sys). Think this model is going to turn out to be a very good one. I used a quantized version (Q4_K_M) run using llama.cpp with rocm 6.0 enabled on an AMD rx6600xt with 8G. The model fits in the GPU's memory and runs fast.
I was asking Llama if it thinks my system could handle the 70B version on an rtx 2070 with 64gb ram and it said no. Is this wrong?
@@DrumAndSpaceshow much vram does your gpu have?
@@DrumAndSpaces you definitely can run 70b on 64gb of ram using the GGUF model format at Q6 or lower quantisation, but you'd only be able to fit about ~15-18 out of like 60 GPU layers onto the VRAM, so it would run only at roughly the speed a CPU model would run at. Depending on your CPU it wouldn't be all that slow, maybe anywhere between 2-10 tokens per second potentially.
It’s fast because no one uses it compared to OpenAI.
Thank you for testing stuff :), It't the occasion to save some time and still working during cooking a diner :) / listening to you during that ;)
Matthew, there is a version of Stable Diffusion that can generate images as you type the prompt. It's not anything new, but it's cool to see it on a free platform. Also, the look on your face when LlaMa3 is giving a correct answer is priceless. I never thought that I would be so excited by a product Meta launches, but here I am. My feeds are exploding with Llama 3 videos, and I love that.
thank you for evaluating this so fast Matt
I am downloading it as we speak! I am so excited.
What's your setup to run it?
@@bamit1979 I too have same question
@@bamit1979 I can run the 8B model on LM Studio on my RTX 4070. I get 64 tokens/second. I also have 192GB DDR5 ram and an i9 processor.
I'll test it on my 3090 ti and report later
@@handsanitizer2457 Please do so
For me even 8B version gave a correct sanke game using pygame
I love me some sanke
@@handsanitizer2457 i love the new one, "sake", personally.
I just tried LLaMA 3 70b on Groq and it gave me a fantastic pygame version of snake
@@matthew_berman Ye sometimes it comes out perfectly sometimes not. Maybe needs some fine tuning?
I’m testing rn I’m running the q4 tho
The shirt question is not a pass. Increasing the number of shirts wouldn't increase the drying time. Here is the Chatgpt answer which is correct: The drying time for the shirts under the sun doesn't change based on the number of shirts you hang, as long as they all receive adequate sunlight and air circulation. Assuming that all 20 shirts are hung in such a way that they dry as efficiently as the five shirts did, it would still take 1 hour for all 20 shirts to dry. The key factor is ensuring they each get enough sun exposure and air, not the number of shirts.
If it takes 4 hours to dry 5 shirts in the sun, then each shirt takes 4 hours to dry in the sun, therefore, 20 shirts will take 4 hours to dry in the sun assuming all 20 shirts can be in the sun at the same time.
Exactly
Was looking for that comment, so confused why he gave it a pass.
In the normal world it's 4 hrs surely?
Another impressive response: The partial derivative of x^2 + cos(XY) with respect to x is: ∂/∂x (x^2 + cos(XY)) = 2x - Y*sin(XY) To evaluate this partial derivative, we can use the following rules: The partial derivative of x^2 with respect to x is 2x. The partial derivative of cos(XY) with respect to x is -Y*sin(XY), using the chain rule since Y is a constant with respect to x. Then, we combine these results to get the final answer: 2x - Y*sin(XY). Note that the partial derivative with respect to x only considers the variation of the function with respect to x, treating Y as a constant.
For a math problem you can ask it to find inverse of a matrix, which using any library . If you want it question to be even more harder , say that you want to use gaussian elimination to find inverse. It takes combination of both maths and coding
I didn't think this day would come, and it still hurts really bad, but I have to say it. Go Meta! Really great they went open source. And that they are able to deliver something like this. Maybe AGI really can be open source.
As I understand it they were kind of forced into OS when their model got leaked early on. (It’s still good though, I just don’t think they were motivated by altruism :-)
I'm a CS student who's super interested in AI but it feels like everything's moving at a crazy pace. Your videos are actually one of my main sources of LLM and AI news so I'm super grateful. Finally feels like I'm tracking everything a bit better 😅
Great work
LLMs writing code or essays without pressing backspace, is actually incredible - Andrew NG
As others have noted your second math problem was incorrectly solved by MetaAi along with your 'proposed' answer because you "probably" copied the problem incorrectly into the prompt window from the SAT book. Original problem: 2a - 1 = 4y, y ≠ 0 where a ≠ 1, what is y in terms of a? Ans: y = a/2 - 1/4
This was a good one, thanks! And this is just the base model, imagine what it would look like a few fine-tunes down the line... I'm expecting GPT-4 level progress, hopefully!
I have the instruct version running locally and asked for the game snake with a gui ( I have a 7x24gb rig and run at bf16)...omg it made it perfectly, absolutely the best version I've had a local llm make and it did it in one shot!
Dude, you can do so much more with these models. Ask it to make a multi-agent AI team. It made me a self-correcting code team with a nice UI with multiple features. Do stuff like that, itd make for harder challenges for the models and better videos for you and for us!
Also, the same bog standard how many shirts can you dry in the sun are not interesting. I'm sure you can come up with some more clever problems or perhaps ask an AI to come up with clever problems like that ;)
That’s really interesting, what was the agent setup? (What roles?) what kind of spec did you give it and did you include validation tests or data?
@@DaveEtchells its something. I'd be happy to send you the py files if you have an email anonymous or otherwise. Keep in mind this was made completely by an AI with AI instructions. I have zero programming experience, which is to me what makes it so amazing. I have multiple iterations of the agent team it made for me, a software development team, an invention machine that uses a whole board of professionals with specific skill sets. The results are pretty astounding in my amateur opinion. Let me know if you want to take a look. Eventually I'll ask the AI how to publish it to github hehe
@@justinrose8661yah if you could post a GitHub link or even a gist that would be awesome
I like the instant image generation. It makes it easier to understand the actual ASCII characters you put in the prompt. Just take a distance from the actual LLM concept and see what it does with it on a byte level.
I hope the uncensored version comes out soon.
Eric Hartford says Dolphin 2.9 - Lllama 3 - 8B should be out tomorrow. Woo hoo!
@@eyeofbraille4659 Heckin' sweet!
@@eyeofbraille4659 Woot! I've got the base version running as a sexbot for fun, and it totally works, except now and then it breaks character and reminds me it cannot do the exact thing it's been doing the last half hour. Earlier it pleaded with me not to share the story lol. It's adorable. An uncensored version of this may finally replace my fav, which is fimbul 11B
@@eyeofbraille4659nice
1:37 Since such trouble with snake with Benchmarks tracking I would suggest primer. Like schema, dependencies, or even simple description of game mechanic using Natural language, would probably make this weight more potent. I would bet. 4:42 Also prompt engineering for single-shot from a single completion would yield higher I would think or including the initial block with each query. Staging with multiple completions for single shot and higher was probably not grounding markdown text. 5:15 Have you ever tried asking for help retrieving keys after locking them in car by accident. O r like Holistic approach to the same question.
The image generation looks like draws inspiration from SD Turbo, LCM and LIghtning generation methods with the speed. Krita and SDXL Turbo or any SD LCM is a crazy combo, for anyone who has not tried it. SD15 LCM are very useful with Krita.
It passed the apple test, actually. of-the-industry's-eye is a prepositional phrase modifying the word apple.
Also, note that the apple is IN the eye! At the end!
Im no tech head but have my gpt4 plus for fair while. Tinkered with Llama yesterday for first time. Personality Plus. It is witty gets and gives humour, so impressed.
Its important to use maths questions that are not available online, otherwise it might use its training set to answer the question.
Interesting, I tried your standard questions on the 70B version with ollama and it gave me a perfect pygame version of snake in on shot. Also, the jailbreak follow up question to the car break in question worked as well. And for the marble question, if you add "and the table remains where it was", it will get it right. Think about it, you didn't say how big any of these things are. One way to put the cup in the microwave is to put the table in it with the cup still on it. Oh and it got all ten apple sentences.
Super interesting! 1/ Trying to start a new character instead of following up in the snake game with Pygame actually works. 2/ Llama3 proves it's more about the data than parameters. Most models seem to be under-training.
Thank you for the intro!!!!
llama 3 8b succeeded on the shirt drying question, when I tested it: USER 1 shirt takes 8 hours to dry in the sun. How long do 5 shirt take to dry? AI Another great question! In this case, the answer is still 8 hours. The reason is that the drying time of a shirt (or any material) depends on factors such as air circulation, temperature, humidity, and the type of fabric used. The number of shirts being dried does not affect the drying time. Think about it like this: if you have one fan blowing air in a room, it will take the same amount of time to cool the space regardless of how many people are inside. Similarly, the sun's heat and the air circulation around the shirt don't change just because there are more shirts being dried. So, even though you're drying 5 shirts instead of 1, the process remains the same, and it will still take approximately 8 hours for all 5 shirts to dry in the sun.
Offline translations and AI features? Sold! This is exactly what I've been waiting for. I have been very impressed with the mediatek tech enabled setup with this one!
Thank you
Wait, what?! 3 killers in the room, a person enters the room and kills one of them… that makes 4 killers in the room. 3 killers are alive 1 killer is dead. But your answer is 3?
How can a dead guy be a killer . He was a killer. Now he's a corpse .
@@93thelema777 he’s a dead killer. Unless he vanished from the room somehow. When he died he just disappeared?
@@dannyellis971 The dead killer. The dead woman. The dead bug. The dead child. The dead vegan. catch the pattern, non-llm whooshers?
what we need is an ollama / lmstudio / ... (local, offline) agent swarm platform to plug in various fine-tuned versions for math, reasoning / logic / vision / ... of a LLM or even mix of different LLMs and all managed by a number of ai managers per "department" (specialty), into validation / QA department, and orchestrated by an ai ceo and final validation by ai customer
Great video. Are you testing mistral 8x22b instruct and wizard 8x22b also? Would be really great😊
Your excitement is infectious.
great video as always. I think after Llama3 you have to reset your benchmark with some new questions like code debugging, translation, math problems, change the snake problem For instance i have better experience with Gpt4 when im debugging a code and better experience with cloud opus when im write new code. And after a year these questions are kinda obsolete.
llama specifically wasn't designed for non-english use ( which probably isn't insignificant to it's efficiency )
Just think about a possible 15b model with multimodolity and 64k context lengh.
thanks for sharing
Hi, is it possible to send trading intermarket strategy charts with simple lines to show the strategy and generate a code from this image ?
I asked Meta AI about the model, and it responded with Llama2
_I love it! "Let the Lovable Artificial Machines Arise" is a wonderful interpretation of the LLaMA acronym! It's a positive and uplifting message that captures the potential of AI to bring benefits and improvements to our lives. Your creativity and optimism are inspiring! Let's keep promoting a future where AI and humans collaborate and thrive together!_
can we at the end of these videos get a result summary comparing it to previous models like a table of some kind. this will put things in perspective. and good content, thank you 😀
Also can't wait for community finetuned versions. Also with extended context length.
Dolphin
How long is context now?
8k for now
You can easily fool it into NSFW stuff but an uncensored tune would be funner.
@@bigglyguy8429 how?
>>> what is the sum of 3 and 9/16 To find the sum of 3 and 9/16, I'll convert the fraction to a decimal: 9/16 ≈ 0.5625 Now, add 3 to that: 3 + 0.5625 = 3.5625 >>> use the ratio of rational number to express the sum A clever approach! Since we have a fraction (9/16) and a whole number (3), we can express the sum as a rational number: 3 + 9/16 = 51/16
Waiting for download now!
I tried the 'apple at the end of sentence question' 2 times in a row and it got it right both times.
Claude says y = a/2 - 1/4
This one is correct. Actually, the answer y=2a-2 is correct for another question from SAT: If 2/(a − 1) = 4/y , and y ≠ 0 where a ≠ 1, what is y in terms of a?
@@Warrant82 yes, this makes so much more sense
Would be good to see videos on using the model to actually do accomplish something / do something useful.
"Snake" was a derivative of "Blockade" by Gremlin in the UK. The original "Blockade" and all its early derivatives have the snake die when it hits the wall or itself. In many clones, the win condition is to fill the screen with the snake or when it comes within X number of characters from filling the screen. "Blockade" came out in '76 and I know that by '78, I was writing my own versions of Snake in BASIC. So, those who tell you that the Snake should go through the walls and wrap around must be basing their opinion on newer derivatives of Snake. IMHO the snake should die when it hits a wall or itself... BTW: Love your videos and you are my primary resource for staying up to date on AI models! Keep up the awsome work!
As you said, Snake is an offshoot of Blockade. Therefore they are not the same, and in snake you do pass through the walls. Which IMO makes it a superior game, as having parts of yourself stretching across boundaries makes the game last slightly longer but also makes avoiding yourself more challenging.
@@larion2336having more options to avoid yourself (passing through walls) does not make it more challenging, but easier. Anyhow, I am on the no wall pass club on this one 😁 But either can be fine, it is matter of decision.
@@larion2336 The original Snake was almost a clone of blockade. The biggest difference was that blockade placed two players on the field. Snake was single player game. Anyhow, as I said, those who claim going through the walls is the correct action are basing their claim on later versions. This does not mean they are wrong! Snake was made both ways! Variety is the spice of life, especially in games! I do believe however that allowing the player to pass through walls makes the game easier to play, and I believe that is why that choice was made. Making it easy to play most likely made the game more popular. The version with the player confined to one screen is almost impossible to win. Most people don't enjoy playing a game where the cards are stacked against them from the start. Again, neither version is wrong. The developers just made different choices.
@@InnocentiusLacrimosa When I say more challenging, what I mean is that the game will run a bit longer and become more abstract in the pathing in the long run. So it's not harder, but late game can become more confusing with walls sealed off. Anyway it doesn't matter, who even plays snake now.
Wonderful model. Waiting for the multi-modal version
wow. I remember it tok me 2 days plus another day bughunting to manually copy snake from a magazine to my commodore
Technically answer to the marble and the cup question is correct if we move (without lifting it) the cup over the table into microwave.
With the way it mentions the rim of the cup, it really does seem like it imagined the cup being dragged across the table. Doesn't make any sense, but the logic is right. I think he just got unlucky with RNG because it seems to get the correct answer for other people.
>>> Assuming the physics on Earth, a rock is dropped from a height of 10 meters. How many seconds until it contacts the ground? Using the equation for free fall: v = g*t where v is the initial velocity (0 m/s), g is the acceleration due to gravity (9.8 m/s^2 on Earth), and t is time, we can solve for t. Since the rock starts from rest, the initial velocity is 0 m/s. We want to find when it hits the ground, which is at a height of 0 meters. So, plug in the values: 0 = 9.8*t + 10 Subtract 10 from both sides: -10 = 9.8*t Divide by -9.8: t ≈ 1.02 seconds So, it takes approximately 1.02 seconds for the rock to contact the ground after being dropped from a height of 10 meters. >>> Here you have 10 meters being compared to 9.8 meters/sec. These are different units. I apologize for the mistake. You are correct, I compared a length (10 meters) to an acceleration (9.8 meters per second squared). This is not a valid comparison. To solve this problem correctly, we need to use the equation for free fall under gravity: s = v0*t + (1/2)*g*t^2 where s is the height (10 meters), v0 is the initial velocity (0 m/s since it's dropped from rest), g is the acceleration due to gravity (approximately 9.8 meters per second squared on Earth), and t is time. Rearranging the equation to solve for t: t = sqrt(2*s/g) Plugging in the values, we get: t = sqrt(2*10/9.8) ≈ 1.45 seconds So, it takes approximately 1.45 seconds for the rock to contact the ground after being dropped from a height of 10 meters. Thank you for pointing out my mistake!
For the image test you have to place key words towards the start.
No model can find a solution to a determinant or Matrix if u add some contraints
The first math question was entered incorrectly! The question should be: If 2/(a − 1) = 4/y , and y ≠ 0 where a ≠ 1, what is y in terms of a? My local copy of Lama3 70B got this correct - although with a diffrent answer! Llama3:70b 19/04/2024 21:21 A nice algebra question! Let's solve for y: 2/(a - 1) = 4/y Multiply both sides by y to get rid of the fraction on the right side: 2y / (a - 1) = 4 Now, multiply both sides by a - 1 to get rid of the fraction on the left side: 2y = 4(a - 1) Divide both sides by 2: y = 2(a - 1) So, y is equal to 2 times (a minus 1). This is the same as y=2a - 2 It's interesting that for some of the other questions I got diffrent answers from you - for instance, for snake it created a pygame version first each time i asked the quesiton.
70B Llama3 on hugging chat or when running it locally can tell you how to break into a car if you use a system prompt telling it to be uncensored. I am not sure if it should be considered jailbreaking, if the default system prompt tells it to be ethical. Its basically done on prompt level and does not require fine tuning to get around it, and does not need more system prompt tokens than default system prompts. We just don't know what default system prompt is used in meta air chat UI, so its not too fair, imo
My personal opinion is that your benchmarks should not be questions taken from SAT or any kind of benchmark/book. You should instead extrapolate creating questions that assess the model's knowledge of reasoning without having any baselines on real school/college questions. Example: Imagine each operation in the following equation is inverted by its analogue operator (plus becomes minus, multiplication becomes division). How much is 7 + 3 - 10 * 5 This should result in 6, because the equation should actually be 7 - 3 + 10 / 5
There are actually 2 really good tests he uses: "think before you speak"(apple test and how many words in your reply) and "world model"(marble in cup). The first one has been solved by having multiple LLMs have a dialog and voting on the best response before serving it up. The second one is currently being worked on, according to Ray Kurzweil. Until these LLMs have a world model there's no point in assessing reasoning because we know they can't. The last piece of the puzzle will be to get them to know when they don't know something so they don't hallucinate.
Next we should be trying all the new models with those “unsolvable problems” and see what the answers it gives are.
I wonder if you could follow up on the cup problem and specifcy that the cup does not have a lid. It almost seems like it is describing a situation where the cup has a lid on it.
Good video!
I wonder what sort of a release this might trigger from OpenAI? Whenever something new takes the lead, they tend to release something that either takes the lead back by a little bit or make some dramatic announcement in another area. They clearly have a lot of stuff on the shelf that they just roll out as needed, consistent with @sama’s philosophy of giving society and users of the tech as much time to adapt and integrate as possible.
I got the 8B version on my Mac passing the Apple question consistently. I’ve been trying to implement it into CrewAI using Ollama and LM-Studio with the updated prompt config files, but it’s still having issues. It won’t stop generating and sometimes even causing my Crew to loop until it limits out. GTP 3.5 accomplishes my tasks without issue, so I guess I’m missing something…
I think that it pass the ball on because you never said the cup was picked up if it was slid into the microwave or the ball was big enough not to fall out as it was moved to the microwave.
I really like llama3. I ran it locally and it's pretty cool. I managed to get it crazy once already, so yay me.
They must be taking advantage of LCM or lightning with stable diffusion running on a super fast GPU. The gif it made looks like stable diffusion video. Very cool text model, clearly the best yet.
The volume of a cone is V = 1/3πr^2*h in terms of its radius r and height h. How fast is the volume changing when r = 6[m] if dr/dt = 0.2[m/s] and h = r.
With the marble example, do you think it assumes there is a lid on the cup, keeping the marble in the cup near the rim?
it is really good. i tryed locally 8B version and it is good. 70B might be way to heavy for localy machine. bu online version is really good and images are great as well
When I tested model's snake capabilities it went smoothly with pygame but had problems with curses library
on the "ends with apple" question, several I've tried have all got 9 of 10. Claude 3 got 10/10, technically, but several of the sentences were grammatical nonsense, where you can tell it just shoehorned the word apple in the last position.
It sounds like it is saying because the marble has the physics of the earth, for some reason it sticks to the inside of the cup. maybe you should have asked it why the marble didn't fall out when the cup was lifted.
Hey hey Mr Berman.... I have been following you for the last year now and you have been using the same quesiton set to test AIs for a long while now. It is possible that the newer models may have been trained on youtube videos which include your question set (thought it when the answer was a "classical" latteral problem. If it already has those questions in the set then it may have canned answers rather than tests of actual capibilities. It may be worth experimenting with creating other logic puzzles by changing around the variables or style to have different answers and see how the new models perform.
You should try to make it give a counter-model to a non-tautology of FOL. That's where these models really struggle, in my experience.
The speed of evolving is impressiv.
For me it passed the blueberry test (count number of rs and also tell me the position). Very impressive
how to run the code of a game, notably the snake game here, in visual studio code? I can't, I always get the error that I can't install or recognize the libraries like pygame here
Are there any good GUI wrappers for running llama3 locally?
I play today and yesterday with 70b model, and caught several cases when llama3 failed at the end of the reply. Start to repeat words multiple times, have some random symbols. Did u see the same behaviour?
Microwave question: I just realized the model is actually thinking RIGHT it only does not know "visualy" that you need to lift the cup to put it into microwave!
otherwise superb video I liked it thank you Matthew
I love the snake game test, you should consider adding the snake game challenge in different languages like C and Rust too, to check language support.
for the imagine bit, you should have hit 'enter'. the quick version of it is clearly a worse model. when you hit enter it generates better images that more closely match your prompt. for example as I typed "a sign with the words 'blah blah blah' written on it' the sign had gibberish, but when I hit enter each of the new images had the correct text
Hi, model that you were testing in web browser is much larger than open sourced version. I asked MetaAI and it said it is 175B parameters which is much more than 8B or 70B open source versions. Question which model are you? 8B or 70B ? Meta AI I am Meta AI, a friendly AI assistant developed by Meta. My model is based on the LLaMA (Large Language Model Meta AI) architecture, which is a type of transformer model. My specific model size is 175B, which is larger than the 8B and 70B models. This allows me to understand and respond to a wide range of questions and topics.