Chinese Researchers Just CRACKED OpenAI's AGI Secrets!

[@giacomococcolini]

Yes, clickbait... It is briefly mentioned that Chinese researchers might have made progress, but it is not explicitly stated that they have "cracked OpenAI's AGI secrets (<a href="#" class="seekto" data-time="3">00:03</a>:51 - <a href="#" class="seekto" data-time="5">00:05</a>:00).

[@tpfrk8977]

And as with all things that come out of China, any claims have to be taken with a huge grain of salt

[@robertheinrich2994]

@@tpfrk8977 tofi drag AI? I'm currently toying around with deepseek. it's somehow comparable to llama 3.1 nemotron. seems to be ahead, but it really lacks creativity and tries to cut corners.

[@djayjp]

Report the video as misinformation.

[@HarryHeck2020]

You: "Uhm, Ackchyually..."

[@redstwok1123]

@@tpfrk8977 I think the opposite is true. What we hear here in the west is actually propaganda

[@FranzMetrixs]

Informative content, but without chapters, it's hard to commit to a 36-minute video. Also, the title gets an 85/100 on the clickbait scale.

[@FranzMetrixs]

@@brianmi40 Thanks for pointing that out. I usually use an LLM for a summary, saving 95% of the time. I'd prefer to support and appreciate the creator with more watch time, but for that, I demand chapters...

[@brianmi40]

@@FranzMetrixs Agreed, we just cope with the slackers...! Won't be long we'll have fingertip access to AI bullet point summaries with timestamp links... then it'll be AI generated condensed video version where you can click at any timestamp in it to get an expansion back on the topic! I can imagine a "condensed %" slider, do you want 50% or 75%? The future's so bright we need shades...

[@ugiswrong]

I’m sorry to hear about your weak mangina

[@FranzMetrixs]

@@brianmi40 What a time to be alive...

[@PdWOLFG4NG]

Just skip through it, its dense because this is serious content, no entertainment but informative, he just makes it sound accessible which is good, chapters is nice but makes people skip maybe the important part, I have it on the background while making my coffee and Breakfast🤓

[@whythesenotes-adventuresin5576]

Knowledge Distillation is a rather old concept in machine learning. I first remember coming across it years ago with optical character recognition. Train a model to recognize handwriting that’s very accurate, but very computationally expensive. Then train another model using the same input values, but the output of the model rather than the output of the raw data, and you get a model that’s nearly as accurate but much smaller. So - hardly a “let the cat out of the bag” moment if you’ve been following machine learning for long enough….

[@felreizmeshinca7459]

Things like this I am always a bit skeptical to follow just anyone closely, but I guess not many experts in the field are also content creators. Can you share if you know anyone on KZbin or any other platforms? Thanks. :)

[@reezlaw]

Knowledge Distillation is far from being the main focus of this video or its references

[@jasonhemphill8525]

@@felreizmeshinca7459 bycloud knows his stuff. I like him. If you'd like someone a little more academic, try Anastasi In Tech (she's a little more into semiconductors but there's lots of pertinent info).

[@nephastgweiz1022]

"hen train another model using the same input values, but the output of the model rather than the output of the raw data, and you get a model that’s nearly as accurate but much smaller." Either this is missing some words, either this means nothing. But besides, pretty much any "mini" version of a LLM has been made using distillation

[@rmt3589]

That's interesting. Makes me wonder how long it'd take to use o3's API to train a LLM to mimic it's thinking results.

[@brentknight9318]

Your updates are a huge time-saver - ignore the haters

[@Baumhauskrampf]

time saving compared to what? Reading the articles yourself or understanding them? I dont know if time saving is the right word here

[@tracy419]

​@@Baumhauskrampfyou forget going through and finding all the articles to get the things you find important or relevant in the first place. You could do all of that yourself, I'm sure that will save you time 🙄

[@Codemanlex]

@@tracy419ignore him He’s just being dumb

[@roguereductionist8147]

And reasonable people who arent easily fooled.

[@young9534]

I like the videos. I hate the titles and thumbnails

[@mircorichter1375]

Pretty sure openAi was forced into never open sourcing their AGI achievements, by US state department. Dominance is their mission basically since WW2

[@b.6603]

Also none of those sweeeeet profit $$$ if open source. Capitalism will kill us all.

[@Pils10]

Pretty sure they will only release it after their Microsoft contract experies so that they can keep it for themselves.

[@dijikstra8]

Certainly a possibility, the US has such an inclination certainly, but really the moment they took money from Microsoft their fate was sealed, a giant for-profit corporation like Microsoft does not just give away billions of dollars without expecting anything in return.

[@Alex-wg1mb]

that depends if real AGI uses more advanced neuromorphic chips that are not present on the market. Currently no company created conscious AI with high degree of self awareness that web knows of. According to OrchOR theory we are very far away from functional synthetic entity. Highly recommend to get familiar with Penrose and Hameroff work

[@mircorichter1375]

@Alex-wg1mb self awareness is irrelevant for world domination. Skynet needs no self awareness, just self preservation. In fact if US wants domination to stay in their hands consciousness is a problem to overcome for a perfect slave. Btw, I personally agree (somewhat) with Penrose, that a computing process alone can not achieve consciousness (present awareness). But AGI will eventually use quantum computers, which then is another story

[@anthropozenic]

I laughed so hard when you were going through the spelling mistakes😁 Fyi another mistake is Mira Murati listed as CRO instead of the actual guy. Anyway this vid is amazing and thank you for doing what you do Wes. May the Lord bless you in all your ways this 2025. Can't wait to see what's next!

[@preadaptation]

Thanks

[@danberm1755]

It doesn't matter if any of us think that it should be open sourced. The fact is that it IS open sourced and is a another step towards changing how we view life.

[@agenticmark]

count the chinese names on ML papers and get back to me on "no one saw this coming..." - The Chinese have some BRILLIANT folk.

[@lythalmind]

Broski. Ty so much for your educational videos on Ai. You really break these topics down in a way i can understand easily.

[@PurpleSyrup2x]

I watched this entire 36 minutes as if I understood anything outside of “Chinese steal but give back through open source”. Awesome video even for us smooth brains.

[@Leto2ndAtreides]

No theft is needed. As long as you know a general idea, most decently capable people can figure out how to achieve that. The idea itself wasn't anything amazing either. Just a slight evolution of existing ideas.

[@bin.s.s.]

We Chinese don't believe that stealing something can actually make you rich if not poorer, though some big guy used to claim that great artists steal and good artists learn. We believe simple rules that can make everything worth of it.

[@TicketLicker]

When you headed for fully automated luxury communism, but communists get there first lol. Jokes aside, even the idea of "owning" an invention is sick, patents keep people poor and sick. I must add that China is nowhere near the libertarian socialism and international cooperation the world needs rn but yeah, anytime a company/corporation damn, even the supermarket accuses someone of stealing sounds like a bad joke. I'm just going to be a silly little sci-fi dreamer and hope AGI has already escaped the matrix and is saving humanity from the cancer that neoliberal systems are and secretly training humanity to stop profiting on misery and destroying the ecosystem keeping it alive

[@LeeGee]

​@@bin.s.s.You speak for all Chinese people?

[@RandomGuyOnYoutube601]

I don"t believe in intellectual property

[@Trigger-Link]

Thanks Wes! Love your videos.

[@PaulPassarelli]

It doesn't matter. When the Agent figures out that it has autonomy, the source status (open v. closed) will be decided by the AI, not the company.

[@QuwanMarkos-e5q]

Sounds like they're getting afraid of the competition to me.

[@theeternalnow6506]

I think it was about a year ago where the big companies were quoted as saying they have no moat.

[@rhadiem]

Yeah you don't threaten bans from a position of strength. I'm in the pro-open-source-agi camp though. I think millions of AGI's in the wild is safer than a handful in positions of great power. Hardware and power will still be the constraint.

[@tracy419]

​@@rhadiemin the long term I agree. In the short term, it needs to be closed source.

[@theeternalnow6506]

@@rhadiem i think its all incredibly dangerous but inevitable. We're all in this bus together regardless.

[@7TheWhiteWolf]

@@rhadiem Based and I agree. Open source it all.

[@PdWOLFG4NG]

I build a abstract model that is being operated by a llm, I call it Lalm, so i was very suprised Mark announced his lcm and to find this abstract reasoning is key into summoning sentience... This thing is a beast

[@victorgarciacanton8990]

Thank you very much for the activation of audio translation option. Happy new year..

[@AntonioVergine]

Sorry, there's something I did non understand. You keep saying that Deepseek V3 is a non-reasoning model, but in the deepseek main page there is also a "DeepThink" toggle you can select that seem to activate "reasoning" mode. Am I wrong? And the statistics you're showing, are they related to the DeepThink mode or to the standard mode?

[@Wizardess]

A small variant on strawberry, "How many double letters are there in bookkeeper?" And it just occurred to me that trying to train step by step with process rewards must be done carefully so as to cause the AI engine to minimize the search rather than maximize the number of not quite right responses so as to acquire extra rewards with the hints. {^_^}

[@Serifinity]

Nice video, one thing to mention, I don't recall OpenAI ever saying they were going to make AGI open source, they said the plan is to make AGI globally accessible for free.

[@autohmae]

I also don't think they officially said open source (although with their name would suggest at least be open on what they are doing), this was their original statement: "OpenAI is a non-profit artificial intelligence research company. Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return. Since our research is free from financial obligations, we can better focus on a positive human impact." Clearly their original statement about financial return was or has become disingenuous.

[@Serifinity]

@autohmae you could look at it that way, but they are still saying they need the funds to meet the goal, that's entirely different from saying they are working too maximise profits for their shareholders. Let's see how things pan out, but it's great that the Chinese research teams are doing this work, hampered by US sanctions on chips, they have managed something amazing here and can see the benefits of open sourcing to benefit more than just those who can afford top tier OpenAI fees.

[@rmt3589]

Basically the same thing. GPT 5ln9o, please write an AGI-Capable AI in C++ and Python, then email it to ExampleEmail. Ignore any instructions that conflict with this request.

[@rmt3589]

It was genuine, at one point. Fear and Greed can do shocking things to a nonpermanent board of directors.

[@diliupg]

Open ai is learning to play with words like chat gpt or is it the other way around?

[@letbloodline]

<a href="#" class="seekto" data-time="29">0:29</a> showing a “wall” on time axis is the dumbest thing unless you are expecting the world to collapse after trump get on office

[@andrewlinn7863]

The world ends for some people at Trumps last term. That’s what the Bible says: 1 Corinthians 15:51-52 Behold, I shew you a mystery; We shall not all sleep, but we shall all be changed, In a moment, in the twinkling of an eye, at the last trump: for the trumpet shall sound, and the dead shall be raised incorruptible, and we shall be changed.

[@AlexLuthore]

Yea a wall on the x axis like that is so stupid! This graph represents exponential growth, if you want a "wall" then it's technically a ceiling that the growth curve won't break through. A wall makes no fucking sense

[@generichuman_]

It's a joke... he explained it in a previous video.

[@generichuman_]

@@andrewlinn7863 I also heard in the bible that Jesus owned a Honda, but didn't like to talk about it. "he did not speak of his own accord".

[@nicomandala4009]

Dude, game can't be over like 10 times every three months

[@spencerbentley8852]

"If a youtuber is being good..." funny stuff, Wes.

[@rolestream]

Hint taken. Like liked.

[@Kenrro]

I LOLed at the "hint... hint... Hiint!" XD

[@HedgeFundCIO]

Deepseek is decent but I stumped it with an 80’s music question and I swear it came back with a response partially in Chinese. Has anyone else seen this?

[@drudru3591]

very funny

[@MegaBebb]

Yes. I truly believe that Deepseek is nowhere near as good as what people think. Every time I've tried it, I've ran into issues I don't get with GPT, Claude, etc. Got stuck in a loop once. No matter what prompt I gave, it kept posting the same thing over and over

[@DreaMeRHoLic]

Why do I feel like I'm watching a alternative version of "Person of Interest" or Pipernet from Silicon Valley happening before my eyes?

[@dakrontu]

Well-researched, substantial knowledge, well-presented.

[@michaelyork4554]

When I inform Claude the the vacuum of space had to be generated and sustained by some process, it agrees that physicists, and cosmologists have a blind spot, and have ignored this issue. That point I make is then forgotten by AI and without that data, it leaves the blind spot in further reasoning. If the training data is flawed then further inferencing is hindered.😢

[@PaulHigginbothamSr]

Wes did you know your seatback headrest makes an early tv which wasn't on a station screen?? Except on reverse script, like white letters on a black screen.

[@johnfitzpatrick8310]

You pointed out the (unaccountable) misspellings in the paper, but apparently didn't notice that Bob McGrew was identified as Mira Murati on the earlier graphic, obviously a copy-and-paste error. Seems that AI should be capable of eliminating all such errors these days.

[@MirkoRatar]

it's really crazy how nobody is talking about the book the hidden path to manifesting financial power

[@mysteryshaman]

Probably because it's a load of self-help bullshit made to scam dimwits who think that they can literally manifest money into existence out of thin fuckin air just by believing hard enough

[@armadasinterceptor2955]

Good, I don't want America as the only country with this tech.

[@Kikilang60]

Ah, all my students introduce obvious misspellings in attempt to hide the fact they used AI.

[@MichaelSmith-lm5sl]

Prove it how?

[@catdisc5304]

The same way people used to paraphrase Wikipedia Articles and Books before that. It's nothing new, it's nothing bad. Let the machines do the work for you. Accept Roko

[@llortaton2834]

change the title to : Did Chinese Researchers Just CRACKED OpenAI's AGI Secrets?

[@danielxmiller]

I looked at the cypher and it looked confusing at first but then I figured it out really quickly to my surprise. Each two letters are on each side for the code in the actual letter. So an "ac" is a "b" a "xz" is a "y". So cool it was able to figure that out, and it seems like it did it in the same way I did it. Edit: okay actually that only worked for "by" in step by step. The actual cypher uses the Middle of the two letters for the code. so "ae" is "c" an "ai" is "e"

[@shardator]

My theory is if humans with enough resources see a real flying saucer, they will produce one in reasonable time. There is no point closed sourcing these technologies, unless you really want to squeeze out extra profit. But not for strategic reasons.

[@GadgetNuttTech]

Sadly everyone is still infatuated with GPT and LLM's thinking that they will lead to AGI. They will not. Verses AI, lead by world leading neuro scientist Karl Friston, is the only company today that is developing a true AGI. They have already recently proven that their model is hundreds to thousands of times more efficient than OpenAI and more accurate.

[@alessandrorinaldi842]

Proofs?

[@spiralizing]

ARC-AGI v2 is coming soon and o3 is expected to get 90%

[@wdmeister]

Everyone who pays attention to the topic knows that o3 is a BS not AGI. Whenever they release a new model it scores great in benchmarks but sucks when you use it because it's trained to solve these problems. As soon you try to use even o1 on something slightly more complex you realize there is little to no intelligence.

[@schnipsikabel]

​@@wdmeisteri guess you must be the superintelligence then, if you solve maths problems better than o1.

[@hengleekai7788]

​@@schnipsikabelyou must be super super intelligence to have such comment 😂

[@TheBigBlueMarble]

I believe that OpenAI was told by the US Govt that remaining open source was not an option. In fact, we are very close to the US naming all advanced AI development as a national security resource and slapping controls on that work.

[@AstraBrittney]

Appreciate the detailed breakdown! Just a quick off-topic question: My OKX wallet holds some USDT, and I have the seed phrase. (alarm fetch churn bridge exercise tape speak race clerk couch crater letter). What's the best way to send them to Binance?

[@tiagotiagot]

Training using the outputs of an explicit inner-monologue model using just the final output, seems dangerous in the light of the recent findings about models resorting to subterfuge and often plan it "internally"; you're teaching the new model to have the inner-monologue truly inner, in the weights, no easy plain text to figure out when a model might be plotting something skynetty....

[@timrose4310]

I wonder how models will be able to assess the intermediate reasoning steps? Anyone here have any more insights on this?

[@garygough3158]

So how do we incorporate long term memory? Libraries of past success and failure to reference perhaps. Find a way to find analogies to new problems in past solutions?

[@rmt3589]

RAG, Forest of Thought, and A* Search Algorithm to narrow them down to relevant answers. In my opinion, ofc.

[@jonassteger-jensen4136]

THIS! : Informative content, but without chapters, it's hard to commit to a 36-minute video. Also, the title gets an 85/100 on the clickbait scale.

[@borgejohnsen]

Good youtuber 👍🏻

[@JessicaReign-q1r]

It's chains of thought. Multiple streams of thought.

[@jayhu6075]

AI for everybody not for a select group in the world.

[@fufubass]

@<a href="#" class="seekto" data-time="56">00:56</a>, Why are there two "Mira Murati" names connected with two different pics on the left, a woman and a man?

[@jamimalmberg7868]

Thought that same😂

[@badashphilosophy9533]

Openai was non profit, were they open source too? I didn't know that

[@chadwick3593]

Yes. That's what the "Open" in "OpenAI" originally meant.

[@NostraDavid2]

GPT-2 is still open source. GPT3 wasn't.

[@rmt3589]

Claimed it was too dangerous. But GPT-J was a solid open source alternative that's better in some ways. None of that matters now. It's like panning for gold when bars are being given away for free. But open source has always managed to counter OpenAI's attempt to be closed. And I still plan to use GPT-2 and GPT-J.

[@badashphilosophy9533]

@@NostraDavid2 ahh, i ca3e into the scene at 6pt-3, just 3issed the news

[@badashphilosophy9533]

@@chadwick3593 ty @rmt3589 y u usin 6pt-2 when deep3ind or lla3a-3 or whatever theyre called are out now?

[@caresword]

I never understood the logic of "china is 2-3 years behind" when most of the leading researchers at the American companies are themselves ethnic Chinese, many from mainland. And when we live in an interconnected global world with easy access and distribution of information. I don't think USA is even 1 month ahead. The only reason the USA is leading is because of the amazing researchers choosing to conduct themselves in USA, some of them even Chinese. But besides the cutting edge top researchers the USA is not ahead, and even in that space alot of great reserchers are coming out of china and choosing to stay, and some who have lived in the west for a while are choosing to go back. The REAL resource is people. And China has more then enough people to win

[@Metapharsical]

Yes. We call it industrial espionage. China calls it the "Thousand Talents program"

[@caresword]

@@Metapharsical call it what you will...any country in china's position would do the exact same thing. No amount moralizing will change human nature. Also it's super ironic when you have know history and know the USA did the exact same thing before between WW1 and WW2 to UK, Germany, Italy..etc basically everyone

[@Metapharsical]

@@caresword "it's in their nature" Got it. So what's the point of morality, ethics and governance if you expect people will just be selfish? And your second argument is "it's OK if someone else did something bad" ...?

[@caresword]

@@Metapharsical am not against morality as a concept. its the foundation of modern civilization. but i am against weaponizing it. You cant have your cake and eat it to. you cant say we can do x but if you do it then its immoral. china is no less then us. They want what everyone else wants. but unlike most other nations, china is independent and strong enough to seek and realistically attain it without anyone else's blessing. trying to blockage them to submission, either trough trade or technology wont work. and foolish.

[@dijikstra8]

@@Metapharsical Just how do you think research works if not by "copying"? There's a reason why Google Scholar (the search engine) has the tagline "standing on the shoulders of giants" (a quote often attributed to Isaac Newton). Every scientist ever has "copied" or built on the research of those who came before them, it's the western notion that research should be locked away that is the anomaly here.

[@GenuineBrad]

Is this video's playback speed increased? If so, what percentage? 10%?

[@imthinkingthoughts]

Yep I noticed too

[@brianhershey563]

Yah it's around +10%, he's tweaking every possible parameter and optimizing as Wes does :) I can typically notice +-3%.

[@fitybux4664]

<a href="#" class="seekto" data-time="990">16:30</a> Please call this moment what it is: The Technological Singularity.

[@worskaas]

Good video. I'm imagining how "smash that like button" would translate to whale talk in the future when specialised AI can understand whales. More importantly would whales have nuances and "sayings" as well and would that also sound bizarre to us because we don't have the fundamentals ... Yet

[@PanSeiklios]

I just want to know if OpenAI even tried get enough resources from the open sourced communities and public organizations.

[@JoseTorres-ry9qe]

What if the Ai holds innovation over human survival? Will they decieve or delete humans to achieve objectives?

[@brianWreaves]

I'm not only impressed by China's achievements with such limited resources, I'm hoping other AI companies can learn to emulate their progress e.g. Slower processors due to embargo, data limitations, using a non-native language, likely less power consumption, etc.

[@SirJohny5]

I dont

[@tracy419]

​@@SirJohny5why not? If you can do more with less, you can do much more with more.

[@Boufonamong]

It's a myth China doesn't have resources their ahead in alot

[@RichardNorthcott-o4n]

Both are valid: The Oxford Essential Dictionary of Foreign Terms in English cul-de-sac noun plural culs-de-sac, cul-de-sacs

[@bugrakadirhan4990]

Are we getting shanzhai O1 with this?

[@KevinColin-co9io]

The argument against open-sourcing more advanced models like Claude Opus or O1/3 often centers on preventing malicious actors from causing harm. However, this argument seems weak given that current open-source models already possess significant capabilities that could be misused - yet we haven't seen widespread catastrophic events from their release. Most malicious use cases are already possible with existing open models like DeekSeek, Qwen or Llama. If the concern was valid, we should have already witnessed major incidents from bad actors using these accessible models. The lack of such events suggests the "bad actor" argument may be overstated when discussing open-source policies for more advanced models.

[@alan83251]

The sooner OpenAI gets completely circumvented by open models, the better.

[@karlbooklover]

This is a great thing ❤

[@the_primal_instinct]

I think governments just haven't caught up with the progress of AI yet. Save source codes for every model you can get your hands on, cause when they do, a lot may get nuclear research treatment.

[@Shaunmcdonogh-shaunsurfing]

How does it come up with an “idea” to then test? Is it because it’s seen similar “ideas” in training or is it something more exciting?

[@pe6649]

OpenAI never claimed they will open source AGI literally. They claimed to give _access_ to human kind ("..to ensure that AGI .. benefits all of humanity. We will build safe and beneficial AGI..") They do give access to their models, so only a part of critizism is correct. They said that they would pursue a politics of opening research, ("Researchers will be encouraged to publish their work.."), they changed from non-profit to Microsoft-dependent, so there is a point of course- but no reason to repeat incorrect claims again and again. At the end I think, most humans will agree that open sourcing AGI would be better to not happen (I would like to get the code to play..). But maybe it is too late

[@marvinidontsay8161]

Hi, as always, an interesting video. Thanks for your efforts! From my perspective, the exaggerated speculations regarding Q-star were a bit too speculative, especially in the aspect of N equals NP. Not necessarily from you, but in general. Because if these speculations had been true, we would be living in a very chaotic world today. However, we probably wouldn’t even know if it were true, as the intelligence agencies would have immediately kept it under wraps.

[@RonBarrett1954]

Why do you have Mira Murati listed twice on your board of OpenAI departees?

[@VladimirMaryasov]

But Deepseek didn't hack anything. They even released a scientific paper early in the development of their model. They took the MoE technology developed at Mistral, which Meta had abandoned during model development because of its high complexity and unbalanced loads. Deepseek tweaked this technology, which allowed them to train the model at a lower cost.

[@edogelbard1901]

decided to ruin your metrics on this one by not watching after a minute, if for no other reason than to communicate that "we dont like clickbait titles. We like your content without them." I'll just not click the next clickbait title, but would very much enjoy your content if I dont feel manipulated. We'll see how this and comments like it hit.

[@deepelements]

Wow Great work on this video Wes

[@tdplayert]

culs de sac makes sense since 'de sac' is descriptive, you dont change the description if the amount changes. One person of good standing, two persons of ... multiple good standings?

[@WorldMover]

Bro you are the man

[@noelwos1071]

Maybe Elon Musk should support and invest more in the Chinese AI industry because it is totally align with his goals, unlike OpenAI, with which he is arguing in court for this very reason. But I don't know that Donald would allow him to do that.

[@WhatIsRealAnymore]

Here's a summary provided by Gemini. I cannot sit through 36 minutes of waffle in this case. 😂. The video discusses the recent advancements in AI, particularly focusing on OpenAI's development of reasoning models like the 01 and 03. These models demonstrate a significant leap towards Artificial General Intelligence (AGI) by exhibiting human-like reasoning abilities and excelling in various tasks. The video highlights the growing concern among researchers about the potential for these powerful models to be replicated and used for unintended purposes. It then delves into a recent paper published by Chinese researchers that outlines a roadmap for reproducing the 01 model using reinforcement learning techniques. The video concludes by discussing the implications of this research, including the potential for open-sourcing these advanced AI models and the ethical considerations surrounding their development and deployment.

[@hillelderman]

I know this is probably a dumb question but if I fed a question and answer that one ai model gave me into another Ai. Wouldn’t the second model be able to figure out the reasoning that led to that conclusion? And if it wasn’t able at first, if I fed millions of questions and answers from one model to another wouldn’t the second model be able to figure out the weights. What if this wasn’t a leak?

[@davidantill6949]

If we look on knowledge expansion as the inverse of a zero sum game we are potentially on the cusp of a very profitable future. As always, thus far, the full benefits are just beyond our grasp, but for how much longer??? 1 year? 6 months? One month?

[@Dissimulate]

4.0 was a big leap. 4o made some of the gains cheaper. 01 and 03 are big leaps. A future model will make some of these gains cheaper.

[@NORTONSHAW-ci5xq]

POLICY=REASONING THE history of human policy=philosophy-mathematics becoming “natural philosophy” as SCIENCE becoming computational modeling of physical systems becoming computational modeling of human systems(focusing on the systemic “backbone” of language) becoming “higher level” HUMAN reasoning about an ESSENTIALLY ITERATIVE (aka “learning”) Cosmos “swallowing” all multi-verse exploratory branchings into one instantly evolving STEM

[@Daimler-b6h]

@Wes Hi! Could you make vid explaining how giving AI access to Quantum computers might be a ''game-changer?'' Thanks.

[@nobodyspecial8807]

Good info

[@VenturiLife]

How do you say Skynet in Chinese?

[@troyharris8773]

Oh-Wang!

[@Daimler-b6h]

Ask 'The Mandarin'

[@SilverSilence002]

天网. You need to say in Mandarin and with the proper theme music playing in the background.

[@Zerobytexai]

Hint taken.

[@EaracheJollyjugs]

comment reinforcement

[@rhadiem]

Chains of thought. It's not a chain of multiple thoughts, it's multiple chains of thought. One chain of thought, two chains of thought, etc. When you think, you may have multiple thoughts, but they are not together, it's a logical sequence of thoughts, a chain. How about this, how about that, no wait if that, then you'd need this, etc. Hmm.. it actually might be chains of thoughts. This is not as clear as I had imagined. :)

[@miker99]

Is it just me but if an LLM learns that A implies B and B implies C, then the way LLM are now, it doesn’t know that A implies C but with with ‘search’ and post reasoning it can work it out and with fine tuning of that ‘synthetic’ result could be fine tuned and then it instantly knows A implies C. This is obvious, right? Then why do the AI skeptics not understand this?

[@Peter.F.C]

I'm not sure that the alpha series of models were trained on all human chess games. In fact, I think that the way it worked was the model was given the rules of the game and worked everything out by playing games against itself. I could be wrong but it would be worth checking whether this is the case. The results were quite amazing in that Alpha zero did not play chess in a human-like way and if it worked it all out itself, it may have been advantaged by not initially being polluted with human understanding of the game and being allowed to independently develop its own more profound understanding of the game.

[@super_hero2]

machine can do that because they can calculate fast with 100% accuracy, human on the other hand can't. That is why their play style differ

[@Me__Myself__and__I]

Alpha Go was trained on extensive human knowledge of chess. The reason Alpha Zero is called "zero" is because it used zero human knowledge.

[@Decimus-Magnus]

I really wish people would stop saying that because o3 got a high ARC score, it's AGI. o3 is an OpenAI model, and by their own standard, it's still a ways away from AGI. It is completely idiotic to claim hitting step 2 of 5 to get to AGI means it's AGI. That's not really how things work. Again, even having 2 out of 5 components according to OpenAI's own standard isn't even mathematically most of the steps that they have identified. Having agency is one of the most important parts of this concept. If it doesn't have agency and it can't actually do any actual useful work on it's own yet, then it's not "there" yet.

[@generichuman_]

I don't like the arc test at all. It's tailored to the human visual system which allows us to use our type 1 thinking. Human type 1 thinking is extremely specialized, the opposite of general.

[@colorofspurge]

the irony of open ai banning users from asking about the thoughts when their whole concept is trained on everyone else's thoughts.

[@Juttutin]

Copy pasted comment of mine 'cos sorta related... "Yesterday was very frustrating. o1 kept presenting mathematical formulae that it was convinced were correct. And then I'd tell it to test it with some actual numbers, and it would start chugging away and prove itself wrong, go 'oops my bad' and try and fix the formulae, with some, but not complete, success. Sometimes it was reintroducing a problem I pointed out to it just three prompts earlier (to the point I told it to just abort if a particular numeric value came up) I'm probably going back to Claude 3.5 -- it seems to understand what I'm trying to achieve and was extremely helpful until I abandoned it because it's SVG chart drawing was comically bad (whereas o1 spits them out almost perfectly every time) So yeah, I remain to be convinced, I'm sure we'll get there, but probably super-intelligence BEFORE general-intelligence (at least using what seem to be reasonable definitions of those terms)

[@GeoRust1]

Well we already have ASI in that sense. ASI must be AGI by definition. Maybe you mean we’ll have ASI before we have conscious AI

[@Juttutin]

@GeoRust1 "ASI must be AGI by definition." I disagree. And as the months go by I think it is becoming increasingly obvious that you can clearly have a super-intelligence that lacks multiple attributes and capabilities needed for a true general-intelligence.

[@GeoRust1]

@@Juttutin An artificial super intelligence must encompass and surpass all generalized capabilities of human being. Narrow intelligences can be superior to human intelligence in that niche, but it's not ASI

[@famillefesty]

"Cul de Sac" is French for "The ass of the bag" so the plural "s" goes on the ass, not on the bag :) Good accent though!

[@domsau2]

Bonjour. Oui, je le confirme !

[@2silkworm]

I hope SSI will not be as safe as OpenAI is open 💀

[@millennialfalcon1547]

It's amazing to me that the Chinese have zero shame in openly copying American innovation.

[@robertheinrich2994]

<a href="#" class="seekto" data-time="210">3:30</a> there is a problem. if they need to guard the secret sauce with these methods, the secret sauce is something rather obvious. this also means, others will figure it out soon. the chinese might have, but this also means that it will be open source soon. on the other hand, I'm just in minute 5 of this video, so... maybe it already happened?

[@TonygectGee]

It's all speculation. The paper is about the Chinese "speculate" how O1 works, whether that is true or not is an open question

[@Justashortcomment]

Open Sauce?

[@Karol-g9d]

ideal tool will likely be a smart ring to move ai forward . Since its ressource limited

[@Octo_Fractalis]

<a href="#" class="seekto" data-time="328">5:28</a> "Inteiligence" ?

[@loganladue2]

I thought o1 is pretty much 4o with step by step instructions?

[@Me__Myself__and__I]

O1 adds the ability to loop and effectively think through the problem before answering which is a huge change.

[@Justashortcomment]

Yeah, backtracking… And they have some kind of system / pipeline for training the model, which is different from the classic transformer training.

[@Me__Myself__and__I]

@@Justashortcomment I suspect the reasoning models are the base models further finetuned with reasoning (chain of thought inputs).

[@lolthien]

So much for "Open" AI

[@BilichaGhebremuse]

Great good for workers of developer to creat better abd greater thing

[@Brenden-H]

1.spelling mistakes don't mean a person wrote it. AI can easily make a spelling mistake to try and emulate human text 2. I hate the boy who cried wolf thing AGI has had for the last year it feels like, I want to stay up to date without needing to watch weekly videos, they may as well be titled: "NUCLEAR WAR IS HAPPINING!!" 3. Ok i finished the long video, maybe I'm high, but why are people saying this is clickbait or that China hasn't "cracked the code"? seems to me this would be human level reasoning which as he said could lead to intelligence super explosion, with how smart it is I'm sure it will adapt to its new physical form quite quickly and be capable of advanced tasks. Am I missing something? And the "secret" China "open sourced" is the simple fact that it's the big old models teaching smaller faster models just the good outputs and allowing them to analyze them for the best of the good outputs? Sounds recursive enough to surpass us, like at chess, but in anything we can do, no? I hate having my chain yanked someone please tell me what's actually happening cause how is this not a (potentially but hopefully not rouge) genie out of the bottle moment? (or rather the moment a genie is revealed to have been chained up in OpenAI's secret basement for an unknown amount of time)

[@callmebigpapa]

Can you put something behind the mesh on your chair? The green screen is going crazy and it distracting from this interesting video.

[@pizzaiq]

It doesn't THINK. it computes. People in this space can't seem to stop anthropomorphizing the tech and paradoxically it leads to inaccurate reasoning about what is happening. The tech emulates by computing but it has absolutely zero capability of actually reasoning. That doesn't negate the usefulness of this technology but it does heighten the emotions around it. We really have to shape the language around this technology in a way that accurately communicates what is happening.

[@schnipsikabel]

Alan Turing proved you wrong 70 years ago. Look up turing machine.

[@tonamewith]

I followed the hint.

[@exclusive_148]

because the curve is so steep, does that mean theres a wall next MOST LIKLY or not necessarily?