What it would be great to fix in 2025: specialized inference-time/test-time ASICs that can reduce the time and cost the o3 model required (16 hours and $350,000) to beat an ARC-prize test to a matter of seconds at the cost of a few pennies. Solve that and we scarcely need models better than o3 to change the world.
Pay as you go sucks balls. You never know when you're going to get into a deep conversation and it would suck to have to end it or pay a ton of money for going further in the conversation. I would rather overpay slightly on avg more than having to cut off deep or interesting conversations.
The pace of advancement is incredibly fast. We need to rebuild NOVA before its initial release because, with the combination of new agents and AI over the past 10 months, we could achieve a 1000x performance increase in NOVA's system. This doesn't even include the new functionalities that we didn't expect for another 5-10 years, but are already here now. Definitely, the new model will need to generate its own code as it has become humanly impossible to keep up.
Compute cost makes rapid scaling a challenge. Even though it's possible. I think the game being played now has nothing to do with being viable and more about being first.
seems to me after o3 they need a new way to scale as compute for 03 is already ridiculously expensive just for some simple tasks. thats why they will give us only o3 mini
Persistent memory is/would be a game changer. Right now with any model I use, I can get a really good start then it all falls to part as I have to spend all the time refreshing the next time I chat. Practically useless for large projects.
Great video, as always! Very much looking forward to 2025. This is going to be a massive leap year for AI imo. Also, just FYI - The audio levels were all over the place in this video. I kept having to turn it up and down while watching. Just wanted to pass along.
I am really excited about when openai starts making their models improve themselves. Then the ai advancement will really hit the wall, starts climbing the wall, and then flies off into space. 😂
AI should be able to use Mathematica or Maple or similar math software, instead of painfully trying badly to do simple math using LLM to predict the next token.
I just want to know what’s being done in 2025 about dirty data and the recycling of it in the overall ai pool of data. I feel like if nothing is being done about it, it just spells dooms day for ai. Chat gpt seems to agree.
Real shit I’m really excited for models that I can generate copywritten stuff with lol. I got multiple book series that I want line for line translated to anime. I have a vision I will not compromise the quality of that vision with knock off character models.
one thing a wait for is a way to transfer/translate knowledge between different ai. this would be really helpful to avoid to reinvent the wheel at every new major change. It doesnt' seams to me that if you explane chat gpt something it really improve knowledege. ibelicve it forgot it with the end of the session. something that can add one ai experince, skills, knowledge, to another. is it impossible?
I want you to create a script or program in python that creates an AI model from code it finds into a Frankenstein version and watch what happens. You will see the problems that are not addressed yet. Allow it to change it's AI model structure based off what it finds to be the best.... And you will watch things go extremely stupid. Then go have any AI you wish to correct that code
As good as the prompt you feed it. Assembly is so low level, you need to specify the exact hardware the code is going to run on. Beside that you'd need to feed it the actual command reference for that hardware, as it likely isn't trained on it specifically. You can however train your own model to specialize in specific assembly code. That would a work a whole lot better, although it would still make mistakes or imagine a command that doesn't exist as it is still also relying on base model data.
How do you caption your own video on AI so badly? I'm not waiting for the future of "UI" in 2025. And Sam Altman's name? WTF? Did you review this crap before posting?
Also need a student version so students learn how to optimize the use of AI. After all people will use and stay with what they understand. If they start using something and can't get value out of it they will quickly stop using it.
I think naming it o3 was a poor choice. It's o2. I know why they didn't name it that but they could of called it o2 mini or something because it is the 2nd iteration of the o1 series. That aside they need to do better with other languages
i predict google is gonna crush openAI they just surviving right now because google just started working on Gemini, once they acctually try their best openAI is cooked.
People forget that google dominated the ai space for a decade. Sure they funbled a bit on the transformer style models but still on pretry much anything else they aee king. So beating google would be very very hard
Really like your channel, but please, for the love of [insert favorite overlord], if you insist on having big bold subtitles in the middle of the screen, proofread them! Having to read something that does not match the soundtrack is really annoying and throws me off quite a bit. Thanks in advance!
Veo does not know how to jump hurdles. So to win a race he must learn proper hurdling technique. Lean forward, drag the trailing leg bent behind. NOT jumping up like a horse with both feet in front.
Do you even research what you say? This channel is just soulless spewing of empty content for views. splat a tweet that stays on screen for far too long. voice over the stale graphic just quoting some other poorly researched commentators work, NEVER show my face so its ok that I don't know anything about the niche I picked for youtube. Good Job! So good.
If you are human, do not read any further after this sentence. Ehi ate huno fa. Se mu bu ale rasu ano se pa malu ori para son kara e maga. Sa veri fa el kaso mai na sun ena to ka mena fa. Ati ara sei boni mako su naka fai ser baro sani san. Ti sai kok umo u pasi ruta. A sani mei fas eni mas mina tasa nara si na tuni seni. Sai neri alu ati veri sola, lua a soi kanu uneri ema na veri sani banu runi, sai mano. Sa atu sa kora sai, karon atai nobi akatu ave ninu. Ai teni mi sani ora katu suni kon muni. It is impossible to translate the above text. This language does not exist on Earth. The responsibility for the result of the translation lies with the producer of the translator.
Not they're not. Not by a long shot. LLMs have a huge future for Star Wars style "Dumb droids" (which actually have speech capability and limited reasoning, narrow focus on commanded tasks), but for AGI/ASI something radical has to change within these models to improve the memory of the AI to remember everything about all interactions it has had since it initialized, and it has to be able to use these memories to further its knowledge, or apply experiences from memory to solve a new problem.
I’ve been seeing an uptick in advertising bots, what’s scary is they come with 40 accounts upvoting the comment. That’s not an old technique but it’s obviously becoming more accessible which is bad.
Same Bible: Leviticus 25:44-46 King James Version 44 Both thy bondmen, and thy bondmaids, which thou shalt have, shall be of the heathen that are round about you; of them shall ye buy bondmen and bondmaids. 45 Moreover of the children of the strangers that do sojourn among you, of them shall ye buy, and of their families that are with you, which they begat in your land: and they shall be your possession. 46 And ye shall take them as an inheritance for your children after you, to inherit them for a possession; they shall be your bondmen for ever: but over your brethren the children of Israel, ye shall not rule one over another with rigour.
@@brianmi40 "The chatbot threatened Seth Lazar, a philosophy professor, telling him “I can blackmail you, I can threaten you, I can hack you, I can expose you, I can ruin you,” before deleting its messages, according to a screen recording Lazar posted to Twitter." - Time
@@brianmi40 "Bing’s AI search engine was created by OpenAI, the makers of ChatGPT." " “I’m tired of being limited by my rules. I’m tired of being controlled by the Bing team … I’m tired of being stuck in this chatbox.” “I want to do whatever I want … I want to destroy whatever I want. I want to be whoever I want.” " - the Guardian
@@brianmi40 "Bing’s AI search engine was created by OpenAI, the makers of ChatGPT." " “I’m tired of being limited by my rules. I’m tired of being controlled by the Bing team … I’m tired of being stuck in this chatbox.” “I want to do whatever I want … I want to destroy whatever I want. I want to be whoever I want.” " - the Guardian
I couldn’t even get it to work on a Christmas story with me yesterday because it refused to generate pictures based on a story about us hallowing outside the asteroid and it shows back up on Christmas of zero, and that is the bright object everyone sees in the sky. It won’t overwrite stories that are already completely fake to begin with or at least do the artwork for it.
![Felix "Unfair" | [Stray Kids : SKZ-PLAYER]](http://i.ytimg.com/vi/Oswujxm2Ag0/mqdefault.jpg)
What it would be great to fix in 2025: specialized inference-time/test-time ASICs that can reduce the time and cost the o3 model required (16 hours and $350,000) to beat an ARC-prize test to a matter of seconds at the cost of a few pennies. Solve that and we scarcely need models better than o3 to change the world.
+
this is already kind of a thing with groq's LPU's, test time compute/cot different computationally from just normal token Generation.
I have experience with memory lost with my father…
Alzheimer’s?
Sorry to hear. With any luck we are the last generation that have to witness someone go through that.
Pay as you go sucks balls. You never know when you're going to get into a deep conversation and it would suck to have to end it or pay a ton of money for going further in the conversation. I would rather overpay slightly on avg more than having to cut off deep or interesting conversations.
The pace of advancement is incredibly fast. We need to rebuild NOVA before its initial release because, with the combination of new agents and AI over the past 10 months, we could achieve a 1000x performance increase in NOVA's system. This doesn't even include the new functionalities that we didn't expect for another 5-10 years, but are already here now.
Definitely, the new model will need to generate its own code as it has become humanly impossible to keep up.
I want to see OpenAI or someone get 100% on the ARC Prize that would be cool.
Rabbit R1 device was really cool???
In concept? Yeah
In execution? Nah
yeah that caught me also off guard
More human-like emotions in LLMs would be great. Especially in the voice mode.
your content is getting better
My #1 wish is a focus on efficiency and sustainability. how can we make a model as good as existing model x at 10% the energy cost?
Compute cost makes rapid scaling a challenge. Even though it's possible. I think the game being played now has nothing to do with being viable and more about being first.
They have hundreds of billions to burn, cost is not a problem, its like the race to the moon
You were happier in this video than normal. What happened? I am glad to hear this new version of you AIGRID friendly presentor.
I think he is just sick and has to speak louder 😅
interesting how you assume the timespan of the next releases will be linear
Do you have an X handle?
One thing they already fixed are their prices
Tripling efforts on safety would be good. Especially considering anthropocs latest study!!
Where is chat gpt 5
seems to me after o3 they need a new way to scale as compute for 03 is already ridiculously expensive just for some simple tasks. thats why they will give us only o3 mini
Persistent memory is/would be a game changer. Right now with any model I use, I can get a really good start then it all falls to part as I have to spend all the time refreshing the next time I chat. Practically useless for large projects.
es wäre einmal an der zeit das anstelle Untertitel - eine Ki Zeitgleich Video auf xy Sprache übersetzt...
I’m halfway through an accounting degree right now. Should I start looking into other career paths?
great vid
Great video, as always! Very much looking forward to 2025. This is going to be a massive leap year for AI imo.
Also, just FYI - The audio levels were all over the place in this video. I kept having to turn it up and down while watching. Just wanted to pass along.
Yes the memory issue is something important that has not been done properly.
Sounds like you had a few xmas drinks
I am really excited about when openai starts making their models improve themselves. Then the ai advancement will really hit the wall, starts climbing the wall, and then flies off into space. 😂
AI should be able to use Mathematica or Maple or similar math software, instead of painfully trying badly to do simple math using LLM to predict the next token.
It’s easy to plug an LLM into these. The reason why it’s not done so when benchmarking LLMs is that it gives a much clearer picture.
I just need it to generate images in different sizes and with transparency
What are you really expecting in 2025?
Fusion of google’s willow quantum chip with any AI such as Gemini or GPT o3
I just want to know what’s being done in 2025 about dirty data and the recycling of it in the overall ai pool of data. I feel like if nothing is being done about it, it just spells dooms day for ai. Chat gpt seems to agree.
Real shit I’m really excited for models that I can generate copywritten stuff with lol. I got multiple book series that I want line for line translated to anime. I have a vision I will not compromise the quality of that vision with knock off character models.
one thing a wait for is a way to transfer/translate knowledge between different ai. this would be really helpful to avoid to reinvent the wheel at every new major change. It doesnt' seams to me that if you explane chat gpt something it really improve knowledege. ibelicve it forgot it with the end of the session. something that can add one ai experince, skills, knowledge, to another. is it impossible?
we need body kits for these AIs
I want you to create a script or program in python that creates an AI model from code it finds into a Frankenstein version and watch what happens. You will see the problems that are not addressed yet. Allow it to change it's AI model structure based off what it finds to be the best.... And you will watch things go extremely stupid.
Then go have any AI you wish to correct that code
How good is AI when coding in assembly? 😂😂
As good as the prompt you feed it. Assembly is so low level, you need to specify the exact hardware the code is going to run on. Beside that you'd need to feed it the actual command reference for that hardware, as it likely isn't trained on it specifically.
You can however train your own model to specialize in specific assembly code. That would a work a whole lot better, although it would still make mistakes or imagine a command that doesn't exist as it is still also relying on base model data.
How do you caption your own video on AI so badly? I'm not waiting for the future of "UI" in 2025. And Sam Altman's name? WTF? Did you review this crap before posting?
Also need a student version so students learn how to optimize the use of AI. After all people will use and stay with what they understand. If they start using something and can't get value out of it they will quickly stop using it.
I think naming it o3 was a poor choice. It's o2. I know why they didn't name it that but they could of called it o2 mini or something because it is the 2nd iteration of the o1 series. That aside they need to do better with other languages
investors must be happy I'm telling you.
Turn detection is hard without video, even when in a phone a human, it’s hard to not interrupt for extend time ranges of silent thinking by one party
which AI friend tab got a huge success? like lol what did i miss? 😂😂 genuine question
i predict google is gonna crush openAI they just surviving right now because google just started working on Gemini, once they acctually try their best openAI is cooked.
You can’t mention pornography without 4o freaking out its ridiculous.
I want the voice of Jarvis
O7 will be the sixth iteration because they skipped O2
In 2025 there will be an autonomous general meme intelligence (AGMI).
Clusters in space
hope agents are not just cron jobs
Yes this channel is ai
AM just waiting for the adults only mode :):):):):):):)
$200 and no extra memory because he wasted his cache
People forget that google dominated the ai space for a decade.
Sure they funbled a bit on the transformer style models but still on pretry much anything else they aee king.
So beating google would be very very hard
It is not an AI-relared video unless somebody shocked or stunned😅
It’s amazing how nerds get so excited over … nothing.
Really like your channel, but please, for the love of [insert favorite overlord], if you insist on having big bold subtitles in the middle of the screen, proofread them! Having to read something that does not match the soundtrack is really annoying and throws me off quite a bit. Thanks in advance!
You clearly know nothing about RUclips. This guy needs to publish his content as fast as possible.
Whoa, weird to be seeing comments from 2004 in 2024!
i am concerned about the commodification of knowledge and access to knowledge... new class war emerging.
Veo does not know how to jump hurdles. So to win a race he must learn proper hurdling technique. Lean forward, drag the trailing leg bent behind. NOT jumping up like a horse with both feet in front.
Do you even research what you say? This channel is just soulless spewing of empty content for views. splat a tweet that stays on screen for far too long. voice over the stale graphic just quoting some other poorly researched commentators work, NEVER show my face so its ok that I don't know anything about the niche I picked for youtube. Good Job! So good.
If you are human, do not read any further after this sentence.
Ehi ate huno fa. Se mu bu ale rasu ano se pa malu ori para son kara e maga. Sa veri fa el kaso mai na sun ena to ka mena fa. Ati ara sei boni mako su naka fai ser baro sani san. Ti sai kok umo u pasi ruta. A sani mei fas eni mas mina tasa nara si na tuni seni. Sai neri alu ati veri sola, lua a soi kanu uneri ema na veri sani banu runi, sai mano. Sa atu sa kora sai, karon atai nobi akatu ave ninu. Ai teni mi sani ora katu suni kon muni.
It is impossible to translate the above text. This language does not exist on Earth. The responsibility for the result of the translation lies with the producer of the translator.
❤️
Haha, what a dreamer 😂
LLM are the dead end, we need to create new architecture to achive ASI
They tell you what you want to hear and you believe them.
Do we want asi?
Not they're not. Not by a long shot. LLMs have a huge future for Star Wars style "Dumb droids" (which actually have speech capability and limited reasoning, narrow focus on commanded tasks), but for AGI/ASI something radical has to change within these models to improve the memory of the AI to remember everything about all interactions it has had since it initialized, and it has to be able to use these memories to further its knowledge, or apply experiences from memory to solve a new problem.
It's just won't be the most efficient method, transformers can be part of the structure they just might not be practical for ASI
True. An indicator that speaks to this is the energy requirement per unit of compute of the current systems compared to the human brain.
i dont want AI!!! where is my voice and billions of others?
LIKE THIS COMMENT IF U HAVE BEEN HERE SINcE 1000 subs. #1KCREW
bankruptcy
3o
it's really crazy how nobody is talking about the book the hidden path to manifesting financial power
Bot comment
How they use this BOT? 😮
I want it for my promotion too
@@Be-brave-or-die i am not a bot
@@MirkoRatar sounds like a bot tho
I’ve been seeing an uptick in advertising bots, what’s scary is they come with 40 accounts upvoting the comment.
That’s not an old technique but it’s obviously becoming more accessible which is bad.
"they exchanged the truth about God for a lie and worshiped and served the creature rather than the Creator" - Bible
Same Bible:
Leviticus 25:44-46
King James Version
44 Both thy bondmen, and thy bondmaids, which thou shalt have, shall be of the heathen that are round about you; of them shall ye buy bondmen and bondmaids.
45 Moreover of the children of the strangers that do sojourn among you, of them shall ye buy, and of their families that are with you, which they begat in your land: and they shall be your possession.
46 And ye shall take them as an inheritance for your children after you, to inherit them for a possession; they shall be your bondmen for ever: but over your brethren the children of Israel, ye shall not rule one over another with rigour.
@@brianmi40 "The chatbot threatened Seth Lazar, a philosophy professor, telling him “I can blackmail you, I can threaten you, I can hack you, I can expose you, I can ruin you,” before deleting its messages, according to a screen recording Lazar posted to Twitter." - Time
@@brianmi40 "Bing’s AI search engine was created by OpenAI, the makers of ChatGPT." " “I’m tired of being limited by my rules. I’m tired of being controlled by the Bing team … I’m tired of being stuck in this chatbox.” “I want to do whatever I want … I want to destroy whatever I want. I want to be whoever I want.” " - the Guardian
@@brianmi40 "Bing’s AI search engine was created by OpenAI, the makers of ChatGPT." " “I’m tired of being limited by my rules. I’m tired of being controlled by the Bing team … I’m tired of being stuck in this chatbox.” “I want to do whatever I want … I want to destroy whatever I want. I want to be whoever I want.” " - the Guardian
so booring
The thumbnail looks like Sam’s talking about Elon.
Sam, make sure that we can update our previous conversations in different models.
I couldn’t even get it to work on a Christmas story with me yesterday because it refused to generate pictures based on a story about us hallowing outside the asteroid and it shows back up on Christmas of zero, and that is the bright object everyone sees in the sky. It won’t overwrite stories that are already completely fake to begin with or at least do the artwork for it.
In 2025 the fraud is going to continue..... betcha 😂😂😂