The model is about as good as o1. o3 is still better. That's not the exciting part.
The exciting part is how they did it. It's all in the papers they released.
3% Training Costs: They optimized RL and SFT (reinforcement learning and supervised fine-tuning) with unprecedented efficiency.
Unified Gradient Techniques: A groundbreaking approach that merges multiple RL methods (DPO, GRPO, PPO) into one unified formula. (They discovered new math)
Synthetic Data Mastery: Leveraged multimodal synthetic data to achieve results that rival traditional scaling laws.
Comment has been collapsed.
The stock exchanges are heavily manipulated by the big boys who make more money when a stock rapidly goes up or down. Our financial media is very manipulated.
Self regulation is done behind closed doors and the SEC is not included on most of the regulation. The SEC fine pennies on the dollar for settlements of no wrong doing crime repeatedly. That is a cost of doing business to most major firms (if the SEC even collects).
Since November 2024, more than 50% of all trades are being done in dark pools. Price discovery is extremely hampered when the majority of stocks are not traded on the exchanges.
Semi related, Jon Stewart has a nice 16 video about the stock market.
https://www.youtube.com/watch?v=bP74RBTE8kI
TLDR: Nvidia dumped big so someone could make a lot of money. Controlled Media labelled Deepseek as the "cause"
Comment has been collapsed.
It seems you have answered your own question... Mine is back bit different.
Very often the first mover in a market (OpenAI in this case) is not the winner. The company that figures out the way to make it cheaper, faster, in larger volume (here Deepseek) usually wins. But the AI race is far from being finished. It is just one round, but this is the first round won by a non-US company.
As greedy American VCs have been over invested in tech they don't understand, it is time for a correction. Deepseek is just an opportunity to blame their mistakes on China competition. There is no business rational in the market correction. On that we agree.
Comment has been collapsed.
Well chapgpt is no bread winner for sure, I didn't think about Deepseek being open sourced. It still seems outlandish for this to have such an impact on Nasdaq. Maybe nobody else will say it but I will, China has a proven track record with cheap inferior products. Things that are cheap are usually cheap for a reason.
Comment has been collapsed.
You were correct 5-8 years ago. China was about cheap, poor quality and volume. Now, China is about good quality and competitive.
With an open source model, there is nothing to hide or pretend. The entire AI community can test, assess, and review.
Comment has been collapsed.
Yeah, that's just it. It's cheaper ChatGPT. The point is that it's MUCH cheaper while still being pretty good.
Why is it bad for stocks? Much of the stock market is tech companies (Microsoft, Meta, Nvidia ...). And a big part of what these companies do is AI. Nvidia is mostly earning money by selling advanced chips for AI. Meta is building huge data centers just for training AI. And maybe you heard about Trump's plan to invest 500 billion into AI in the USA. There is a lot of money in AI! Well, now it looks like much of this investment was wasted because relatively good AI can be trained on relatively small amount of relatively crappy chips. That's why stocks of American tech companies are down. Because now it looks like they invested money poorly.
Comment has been collapsed.
Not to mention that in previous years you just uttered the word AI and your stock went up. It was a hype bubble with no comparable progress or success behind the soaring value of stocks like Nvidia's. Now there meaningful progress in cost/performance ratio, but not from the actors who promised everything.
Comment has been collapsed.
relatively good AI can be trained on relatively small amount of relatively crappy chips
Is it really the case? Genuine question as I don't know much in the topic.
When Chinese EV companies started pumping out cheap EVs for European market it only hurt the local manufacturers because the government heavily subsidized local Chinese manufacturers. Cost was still the same and in the end price should be on the same level as it is in EU, but in case of China, government poured money into manufacturing abundance of EVs flooding the market and making it seem that EVs should be cheaper, while it is just communist government doing communist things.
I wouldn't be surprised or shocked if most of the money invested has been done by Chinese government to use the tool for spying. Sure, EU and USA also invest in technology, but its not like China doesn't have history of using every technology they have as means for spying.
So is it really they are better or could they simply overtake ChatGPT due to heavy government subsidizing eventually meaning it is owned by the communist party. While I support the advancement of technology and in no way I favor ChatGPT, OpenAI or practices they do, getting panties wet over something coming out of china is a bit premature and dangerous (also negatively impacting stock market)
Comment has been collapsed.
Well, I can't 100% guarantee that they're not lying about their backing, but from what I've read the model only cost 6 million USD to train (that's relatively cheap) and they don't have access to the best chips (because of American export restrictions).
The AI model has been tested on "standard" benchmarks, so like a test for AI models to see how smart they are. And it performed quite well.
That's what we know.
Comment has been collapsed.
the government heavily subsidized local Chinese manufacturers.
while it is just communist government doing communist things
Wait, the american oil industry, car manufacturing, and agriculture are also communism then? I KNEW IT, damn lizardmen
Comment has been collapsed.
There's a small difference between capitalism and the joke that passes for communism but is actually just some fascist version of capitalism wearing a fake cloak of republic. If you don't believe that try living in the US for a while where you won't have many if any issues and then try living in China and/or Russia for a while where you will likely have no shortage of problems. Don't be foolish enough to try to make it in North Korea though, they will probably make you disappear.
I mean at least in the US I know I won't have to deal with Trump for the next ?? decades until he dies, he's gone in 4 years. Communism looks good on paper but bad actors always take power and turn it into a nightmare.
Comment has been collapsed.
Of course but it seems far too early to jump ship. If it actually outperforms chatgpt then I would full understand and get behind Deepseek but I'm not going to hold my breath. I'm sure there will be even better options in the future but this just seems like a desperate power grab/shot at the US economy by China.
Comment has been collapsed.
For the stock market it's never too early. If you're the first to buy / sell, you make the most profit.
Comment has been collapsed.
You are missing the geopolitics of this. The Chinese just proved that American sanctions don't work to keep competition away and therefore they are unable to hold a monopoly on the market.
Comment has been collapsed.
Yes, although the concept of sanctions is applied incorrectly here. Restriction is more suitable. These restrictions hurt manufacturers and developers of hardware and software. In addition to the loss of profits, this will also lead to a loss of market shares, which will worsen the prospects.
Comment has been collapsed.
Oh that is true but those sanctions are there for a reason. People may not agree with them but c'est la vie. Personally I wish the US would go back to it's ultra isolationist ways from 100+ years ago. Let the rest of the world handle it's problems, it's certainly not doing US citizens any favors.
Comment has been collapsed.
Restrictions, sanctions, and other prohibitions are always just for ordinary people. But there are no such things in politics. So it will be unpleasant only for ordinary citizens, otherwise nothing will change. And the story of the Monroe doctrine is one of the confirmations.
Comment has been collapsed.
Actually they limit governments as well, if they didn't they wouldn't be utilized. If a sanction has no impact it wouldn't be implemented. If people don't buy Russian fuels or tax/tariff them to be unprofitable then you don't think this has an impact on what Putin can do with military buildup etc? Trust me, people smarter than me or you plot this all out. They wouldn't waste time and resources on something that doesn't impact the intended target.
It's not going to stop something but it will make it more difficult to profit. They have embargoes as well which would stop things but that would for certain lead to war if used against Russia or China. Little people will always suffer more than governments, this is nothing new and not limited to US actions. I would hazard to say Russian and Chinese citizens suffer far more from their respective governments than anything caused by US actions. One big example that comes to mind is China's "great firewall" that severely hampers what Chinese people can do in regards to the WWW. The US has nothing to do with that.
Comment has been collapsed.
You should not believe propaganda indiscriminately. Politicians and people in general are more similar to each other than it seems. Well, the final effectiveness of restrictions can be traced back to many examples from the history of the USSR, Germany, Iran, and China. Only ordinary people suffer, but that's why all these restrictions are imposed. The deterioration of life in one way or another leads to instability in the country.
And Putin's military might is even funny. The army of modern Russia is inferior in technology to the army of the USSR from the early 80s. The Russian Federation is even fighting on the reserves of the USSR, which it has not been able to replenish even in decades. Russia is not even a political entity, just like any European country today. This whole war is a bloody circus that should be the prelude to a big war. As it was before the World Wars. On which many people have earned a lot of money and influence. And it's really sad.
Comment has been collapsed.
I don't believe propaganda at all, you clearly don't know me or what you are even talking about.
A government is not going to waste resources on something that doesn't work. Period. And governments don't give a damn about other governments citizenry beyond using them as way of limiting rivals. The idea would be make the people mad enough at their leaders to do something about it. Sanctions are not aimed at civilians but limiting a governments way of attaining wealth/resources/power. It's not even possible to do anything to an opposing government without affecting it's civilian population. I'm really sorry but I honestly don't think I have the time or patience available to teach you the basics and you don't seem to be listening anyway. A big problem is propaganda because a lot of people do just blindly support their leaders as is evident in Russia. As I said these governments are doing far more to limit their populations than the US ever has. Maybe you should do some research in that. I mean there are thousands of examples that are quite easily accessible via the WWW. I mean I haven't even broached the subject of what China is doing in Tibet or Mongolia or trying to do in Taiwan but let's keep pretending the US is the only bad actor around.
Comment has been collapsed.
Did I call the USA a bad actor? I have already said that all politicians and people are the same, and if you switch their places, you will be surprised how similar they will act in similar situations. And this is not to mention how many things look in the "place".
Well, it's stupid to deny that propaganda affects our worldview. At least because otherwise there would not have been this discussion)
Comment has been collapsed.
No, you brought all this into it. You. I freely admit at all times I have no love for my government but they come in shades and the US, hate it or not, is well above where China or Russia is. China has made great strides in recent years, not seeing any tanks rolling over people, but they still make people disappear and still have no reason to be in Mongolia, Tibet or interest in Taiwan. Why Taiwan? The old regime fled there 80 years ago. Let that sink in, 80 years ago. Those people are all dead, their children are quite aged. Why still pushing this crusade? Things like this are why I am very much against Russia and China and certain other nations with really bad governments in place. Calling Xi or Putin a president is a joke. At least that's one thing North Korea is honest about. And no, we don't do similar things in similar situations. If we did then we would all be ruled by lifelong dictators who rob us blind and leave our nations in shambles.
Take one small comparison Japan vs Russia. Japan has no natural resources while Russia has plentiful natural resources. Japan was devastated after WW2 and yet look at their economy today and the quality of life of it's citizens versus Russia. Russia is doing poorly because it has a highly corrupt government and because sanctions actually work to some degree. And Russia has more than enough GDP to modernize it's entire military but all that money is going into private bank accounts for Putin and the rest of his regime.
So while all governments put political self interest before that of it's citizens there is definitely a scale from good to bad to worse. There are reasons why much of Europe/NA/AU/JP/IN are doing better than much of Asia, Africa and South America. That's not propaganda my friend, it's verifiable fact. And it's not solely because they all hang out at the good old boys democratic country club.
I've exhaustively said all that I care to about the subject. I really don't care to discuss it further. All the same thank you for the discussion. I apologize for any slight.
Comment has been collapsed.
I work in an area related to this - so I think I can explain better:
CONTEXT
To understand this, it's important to understand the context.
Since the introduction of LLMs (Large Language Models) mere 2 years ago, with the release of the first ChatGPT, the technology has advanced leaps and bounds.
In general, the LLMs are not something completely new - as ML (Machine Learning) models have been with us for 10+ years.
The innovation with LLMs was - it showed a product that's bigger than the sum of it's parts:
You take an ML model that predicts your next word.
Think of writing a message on a cell phone, and it predicts your next word.
But it turns out that if you train it on a data set that is large enough, instead of being a system that is really really good at predicting your next word... Really really good at it, but doesn't actually knows how to do anything beyond predicting your next word.
Instead you get a system that can actually write a book like it was Isaac Asimov.
Or discuss any topic you want with you.
Or write a computer program.
Or do the countless different things we use ChatGPT for today.
Why has all this preamble been important?
Because the main difference of LLMs from regular ML models, is the fact that they are large. Which is an understatement. They are HUGE.
And each time they create a new model, they make it more and more HUGE. Like 10x 100x 1000x bigger every iteration.
Which caused several changes in the technological market:
Small companies can no longer compete and innovate in regards to AI.
You cannot be the next Google or the next Meta in AI space.
Because you need the resources of a behemoth like Google, Meta or Microsoft to create and train a modern LLM.
Even Apple doesn't have it's own LLM yet. That's how difficult it has become to create/train one.
To train a single LLM model, you now need HUGE (HUMONGOUS) computing power.
Literally thousands or tens of thousands powerful processors.
And as it's more efficient to do on GPUs (Graphics Cards).
(Has to do with the difference between generalized processors, and specialized processors)
Which is also the reason for the rise in the Nvidia stock in the last 2 years: Huge amounts of GPUs are needed to train the LLMs, and the bigger and more complex the LLMs, the bigger the number of GPUs needed for their training.
And in addition to that, it takes HUGE amounts of electricity.
To the point that the companies working on new LLM models, are now building their own nuclear power plants, to support their electricity needs, for training AI models.
https://www.atlanticcouncil.org/blogs/new-atlanticist/tech-companies-are-showing-a-new-strong-interest-in-nuclear-power-heres-why
Which is why it costs Billions (with a B) of $$$ to develop new LLM models, and only huge & rich corporations can afford to do it.
DEEPSEEK
Now that we understand the context in which we operate, we can better understand what is Deepseek, and why it's important.
Deepseek is a small start-up which (supposedly) was able to develop their own LLM models.
From everything I just discussed, one would presume that this would be impossible, as it would cost them Billions of $$$, which small start-ups don't have.
But they were able to train their model on a "minuscule" budget of only $6,000,000.
They claim they were able to do it, but using a lot of optimization, cutting corners (their model is slightly less accurate than the best ChatGPT model, but is infinitely cheaper) and by allowing the model itself to participate in training the model.
EFFECTS
This innovation is very important due to a large variety of different reasons:
Comment has been collapsed.
Ok this was well explained, thank you. I just hope they stop freaking out. Maybe people haven't noticed but when bad things happen to the American stock market it spread to the rest of the "free world", that is to say places without lifelong dictators.
Comment has been collapsed.
In the long run - it worn't matter.
Either Deepseek will turn out to be fake (or not as good as advertised).
Or it will turn out real, and other companies will copy it and build upon it.
So companies like OpenAI, with their infrastructure and hardware, and using Deepseek's optimizations, would be able to create LLMs x1000 times bigger than current ChatGPT or Deepseek.
In the long run - not much will change in the stock market.
Even in the most extreme scenario, that Deepseek will become the next OpenAI - it will simply be traded in the US stock exchange.
GPUs will still be needed.
So will Google's product.
Or Meta's.
Comment has been collapsed.
Actually I would be more worried that China would take full ownership at some point. I wouldn't doubt it's just stolen code of chatgpt to begin with. There's a lot of shady political stuff going on in the background that nobody seems to be paying attention to.
Comment has been collapsed.
I actually checked the numbers - and the whole "stock market crash" is greatly overstated.
The biggest players in the AI arena, were not affected at all:
Microsoft stock remained the same.
Google stock remained the same.
Meta stock ever rose somewhat.
Nvidia was the only major player that seems affected.
It's stock dropped $150 -> $120.
On the other hand, it's stock was $100 only 6 month ago.
It was $60 one year ago.
And it was $15 two years ago.
So it's barely a "correction".
Not really a crash.
Comment has been collapsed.
3% in a single day is a crash by any stretch, it bounced back though. It's almost back to where it was. A 3% swing is pretty major when you are dealing in hundreds of billions my man. You could try to argue that it had nothing to do with the AI news but that would be rather suspect that it happens on the DAY of the announcement and specifically to NASDAQ while NYSE remained on the uptrend. It's all kind of moot now with the hype over with, at least for now.
Comment has been collapsed.
Also, the code is open source; if that were the case, openAI would have claimed that the code was stolen(I'm sure they are studying their code right now)
That's not how open source works in the world of machine learning.
The LLM itself is a "black box", nobody knows what goes on inside.
Not even people who created it.
"open source" in LLM is akin to what was once called "shareware".
You can freely share the product with others, or modify it.
It doesn't mean you can see what's "under the hood".
Comment has been collapsed.
We have to consider the possibility that the whole revolutionary aspect of DeepSeek is fake, aimed to achieve temporary political gains by China:
https://www.youtube.com/shorts/rTbFWN3VFXM
Comment has been collapsed.
Let me add the markets' point of view. The meteoric rise of some companies (Nvidia, OpenAI, etc.) was based on the assumption that, in order to create an LLM (in essence, "do" AI), you needed a lot of computing power. Computing power, which, because of software specialization, was mostly coming from Nvidia. Now, DeepSeek has proven that this axiom is not true.
Therefore, it seems that the market was operating on false premises and is now readjusting the value of a plethora of stocks based on new information. This happens all the time, except it rarely happens on such a large scale.
Comment has been collapsed.
It doesn't matter where this LLM was devised. It only matters what it can produce and if it can hold up to the claims that the parent company is making. As of today, it appears to be. It is certainly not as good as the top of the line LLMs we have had until now. But if it can produce 85-90% of the quality of, say GPT-4o, while using 2-3% of the resources, it's going to swarm the market. And rightly so, in my opinion.
Comment has been collapsed.
My opinion is based on the available facts and data. Of course, if the data changes, my opinion will change along with it. I hope yours will too.
Right now, I can copy the whole model creation toolkit for DeepSeek from GitHub and tinker with it. Granted, I don't have the hardware to train it, but I could rent it and have my own copy of a fully functional LLM. Even if the code license changes at a later date, I can still legally use the copy I have. I cannot do that with any of the proprietary models. In my mind, that's a clear and definite advantage.
DeepSeek R1 (which was also open source) was tested extensively, and it wasn't so good out of the box. Now that people have started to play with DeepSeek V3, we'll hopefully get more data in the upcoming weeks and months.
Comment has been collapsed.
Trust me, I hope it furthers technology by being opensource. That's not really my concern. I just thought it was quite foolish to have such an impact on mostly unproven algorithms or programming or whatever you want to call it. Just like all the nutters who think the world is going to end because some orange misogynist is in power for 4 years. It's good to ponder what if's, not so much to take them to as fact.
Comment has been collapsed.
It's much more cost effective than ChatGPT and therefore the barrier to entry is much lower. All of a sudden, you don't need 500 billion USD in hardware investment to run AI. That's the long story short.
Comment has been collapsed.
Sadly hype plays a big part in things. I could sit here and say I developed a decent AI on my potato PC for $6 and cause the stock market to fall apparently. It will undoubtedly bound back if people see deepseek is not all that but shots have been fired.
Comment has been collapsed.
Open source in LLMs is not the same as open source in software.
In software you get to see the source code, and how the program is built.
In LLMs you get a black box, which you can use freely.
But you have no idea what's inside the black box or how it was created.
You would need to reverse engineer their model to be able to understand what's inside.
Comment has been collapsed.
After reading through this thread I reached two conclusion. One is that the stock market panics at anything unexpected, which to be fair I was already aware of, and two that any company that manages to optimize LLMs to run using less resources can crash the stock market overnight apparently. Did they not expect technology to improve when literally everyone who can is trying to do so? This whole industry is just a few years old and they were treating it as if how it works was set in stone for the next few decades, this is a hilarious lack of foresight coming from people that claim to be good at predicting trends.
Comment has been collapsed.
I just skimmed through the comments and didnt see some details, so i will add my 2 cents
IM NO SPECIALIST btw- just curious follow news here and there. By coincidence i caught one excited news about deepseek that led me into a bunch of other videos- and the general idea i got have a little bit more meat to it then just being cheaper
Its a package of reasons for the hype- and it does include disapointment with current models.
Chatgpt woes
Think about it- sillicon valley and all the ai bros made their strategy mount hype upon hype, AGI next corner, cure cancer what have you- only thing missing is calling it the second coming of christ.
That is a lot of BS and bloat, theres potential yes but they need agressive growth and money for that so it became a hype machine...
What happens when you have something built on hype showing signs its flattening development instead of exponentially rising?
People crave news and alternatives.
Then we have chatGPT wich started pronouncing itself as a nonprofit being all but- and heck theyre officially trying to change it to for profit last i heard. The mask is off.
That said their arguments at the start is what many specilists and enthusiasts belief- that its a field that requires oversight and openess.
So OpenAI starts on that moral high ground, shows everyone its not really about high morals and is not open at all.
then comes...
Deepseek
The last bit is relevant. Unsupervised Learning has been disfavored by most companies and models out there, if im not mistaken it can be harder, take longer and starts more prone to mistakes, delirium etc. BUT in theory unsupervised would be the closes to how organic inteligence(human) works- it may take longer but uninpeded/unbiased is what would in theory lead to better lateral thinking and holistically deduce or reflect... many specialists consider unsupervised the one candidate to reach AGI.
I personally dont believe large language models alone will reach agi- maybe, but i have the suspicion the amount of data and processing to reach agi by llm alone isnt materially feasible in our world. I feel current models are 'primitive' by inteligence standards and definetly completely unoptimized- our processors and binary run linearly fgs. I also suspect theres more to human memory then just building a database like they do... anyway that aside i do believe unsupervised is the closest to that goal.
...all im saying is im not buying the kool aid that AGI is around the next corner, and that includes Deepseek.
But if you think or rather bet it is, Deepseek kept training unsupervised and now reached chatgpt, on some benchmarks better. THAT MATTERS, A LOT.
Its like a flip for unsupervised.
Before ditching unsupervised for one of the alternatives improved results and growth steadily, reducing errors too, but we know that increase is starting to flatten now.
So some chinese guys decided to keep going with unsupervised- just feed it more data and for longer... and they kept doing it long enought it reached(and may be surpassing) chatgpt, all by itself basically. Theyre not improving on optimization, new forms of alignment or supervision or anything like that... raw freaking data to the same student that was repeating class, now that student is somehow getting the same scores as the best student, few better, and it literally have more experience (longer training, probably still ongoing)
In short Deepseek rekindles/adds fire to the AGI hype, its closer to 'the machine is learning all by itself'
And you can have it
Today
I was about to download it some other day but i need to free some hard drive space first.
Equal/better chatgpt, even locally, price FREE, may be a better contender to reach AGI or to break ground in benchmarks soon
Comment has been collapsed.
"I personally dont believe large language models alone will reach agi- maybe, but i have the suspicion the amount of data and processing to reach agi by llm alone isnt materially feasible in our world."
This isn't a problem.
Once we are out of data, we can generate data.
Generated data can be as good as real data, as i
In fact, I work with that. I do evaluations on generated data and verify its correctness.
Elon Musk recently talked about this subject: https://www.theguardian.com/technology/2025/jan/09/elon-musk-data-ai-training-artificial-intelligence
Comment has been collapsed.
I was speaking more on logistics- server costs,higher and higher power draws for even more specialized gpus constantly running either training or paid services... Heck some countries (US included) are stumbling in a lack of WATER. AI is as heavy for the power grid(if not worse) then crypto mining, the growth rate in demand right now, projections to what would be needed etc etc are just so humonguous...
So if llms could potentially reach agi by increasing everything (the data, training time/iterations, processing units) what that would requires X the amount of companies chasing it and scaled theoritically considering the dimmishing returns (its not growing exponentially anymore, major slowdown)... we cant support it. As a planet i mean.
Its the kind of problem one cant just throw money at. The recently anounced stargate project raising a load of cash- ok but will that cash just produce the land, water and energy supply all in one location?
Its similar to the recent woes with lithium- we increased our output(mining) but if alternatives dont come up the world will run out of it, simple like that. Theres already prospects for shortages soon because how many phones we do, cars, house appliances etc... Everyone that can is trying AI now and if anything the arms race for it barely started...
My point wasnt about generated data but frankly all ive seen from it was terrible- any amount of generated data inputed into training was like poisoning the well, decreasing results and benchmarks like crazy. If you look up images and videos its mind boggling how much it affects output- i mention media because because we can clearly see the botched results but the same decay in quality and consistency happens to text as well.
Wich makes sense- imagine every little benigm mistake in output + ai delirium + ai lying +gross mistakes here and there, now add that back as the source and not only it will be learning from mistakes as reference it will ruin the finicky grasp it was getting from good correct data- since it wont tell them apart.
Even if a coherent model could be trained in lots of generated data without just mumbling nonsense(most likely) it would be full of insane imagined whatevers. It could answer a question on history getting right some names and dates (because i imagine those facts repeat much more through the data then ai imagined) but then inject it with a whole bunch of inexistent events and developments because, idk the ai at some point was romantizing history since its also fed by books(including fiction) and it went into the whole generating data to train more...
Dont get me wrong i believe work like yours can be great- thats a filter, validating the data. But when people throw that idea out like musk did it ignores the whole picture. Sure it theoretically could be possible... but the amount of data needed vs how much it NEEDS to be perused/curated vs time and labour required just dont add up. Were reaching a point the amount needed that ai could generate in a month wouldnt be curated by a entire country hired just for that in a year. The time it takes 1 human to check one simple thing quickly the ai already generated what, 20, more?
And thats ONE (not the only one) reasons why i dont see LLMs alone as reaching AGI. Llms are a great start but the baseline is primitive and ineffective. It took the world by storm because of how much(and how impressive) it could all by itself by just dumping data blindly into it but i feel it created this false expectation that alone will take it all the way.
If llms could work in a similar fashion to how evolution did(wich we still cant be sure that alone will, but optmistically bet it does) were talking insane amounts of iterations way beyond what we have currently. Organic brains evolved over millions of years in enviroments with more pressures and selection including elimination... and heck im not sure they can create world models like living beings have done not being in the world, but then simulating a world for then to train is beyond what were capable... the list goes on
Comment has been collapsed.
It doesn't seem to be as unsupervised as advertised as people have found censorship coded in relating to Russia and China.
I'm firmly with you though in that I think AI still has a long way to go. Will be interesting to see where it is in 10 years.
Comment has been collapsed.
As i suspected, anything chinese will at least censor on china. The way things goes the chinese party dont even have to actively get in the weeds for companies to censor- they do it by thenselves because if they dont government will take then down.
Frankly i take everything AI with a grain of salt- heck not even that, some kilos of salt really- so they claim its unsupervised but i dont take it at face value. But id do the same with western companies claiming the same. Theres so much money and interest going around the field so much to gain by mere optics i wonder if any spoke person for any model is ever trully transparent... too much to gain for putting up appearances, embelishing... hence why so so much is weighted on projections and future, that way theyre not 'lying' to the face of investors just painting a optimistic 'but tottally possible and coming' future...
I suspect AI will slowdown a bit. It will probably take speed again soon in how its USED, how to better apply what we currently have etc- the actual benchmarks and leaps in 'inteligence' however i think will be missing for quite awhile. Improving but at a snails pace, dimishing returns.
All could change by major breakthroughs tough- in hardware, methods or heck some breakthrough with quantum computing + AI... but breakthroughs cant be counted on. A single game changer tomorrow, a few impactful ones over time or decades without any- all possible.
More likely AI will be a bubble because the amount of investment and especulation is hedging all in much bigger disruptions and leaps that are much more likely to not happen before the bubble bursts... because the bubble may be bursting very very soon
Comment has been collapsed.
Btw, Qwen is even stronger than Deepseek. We don't have any information cost wise, but their AI is insane.
You can find them here: https://chat.qwenlm.ai/
They were also released today, and they belong to Alibaba.
I tested their video generation, and it's way better than Sora.
I canceled my ChatGPT subscription after testing their Qwen 2.5 Max Model, as there is no point into paying for GPT.
I mean, GPT is awesome, and it has some amazing features, for example, it accepts images as input, which Qwen model doesn't. It also accepts voice as input, and many other things. However, GPT does have free features, so I plan to use GPT only when Qwen can't handle it.
Comment has been collapsed.
It's not just a cheaper alternative, it's a free alternative. Now someone might ask you to pay for it if you're too lazy to set it up. Not only is it open source, it also runs exceptionally well on shit hardware. It was also made with a 6 mil budget which is hilarious considering facebook wants to drop 50 billion into ai this year, and nvidia dropped a few hundred bill already.
Also, it's not just chatgpt, it can do images, music, music, whatever requires ai. They have models and ai for everything listed, and because it's open source, you could take the code and implement it as you want it. Note that it does have censoring against certain things like I hear you can't make things with the pooh bear and China's tyrant, or mention a certain event where people got massacred but because it is open source, you can figure out how to disable that and go with it.
Comment has been collapsed.
?Is this the Chinese government's policy of currency stability for the Chinese New Year?
It seems like every year around this time, some such topic comes up and disappears.
You may also want to take a look at the People's Bank of China.
DeepSeek's AI avoids answering 85% of prompts on 'sensitive topics' related to China | TechCrunch
https://techcrunch.com/2025/01/29/deepseeks-ai-avoids-answering-85-of-prompts-on-sensitive-topics-related-to-china/
Microsoft investiga si DeepSeek obtuvo datos de OpenAI - Bloomberg
https://www.bloomberg.com/news/articles/2025-01-29/microsoft-investiga-si-deepseek-obtuvo-datos-de-openai
OpenAI says it has evidence China’s DeepSeek used its model to train competitor
https://www.ft.com/content/a0dfedd1-5255-4fa9-8ccc-1fe01de87ea6
Well, I don't think I'm going to try to get my hands on it.
Incidentally, it seems that there was a trial-and-error process the other day in the oriental island nation on how to bypass the whale censors with questions about Tiananmen Square.
Some interesting responses were found in the process.😸
The other question is that if the AI was a copy, there are currently no reports on the hardware that runs the AI. I think the right thing to do is to evaluate it patiently after this period is over without rushing into it.🙄
Comment has been collapsed.
US Navy bans members from using China's DeepSeek AI app out of security fears
https://nypost.com/2025/01/28/us-news/us-navy-bans-members-from-using-chinas-deepseek-ai-app-out-of-security-fears/
🙄Oh...Things seem to be moving more quickly than expected.
Comment has been collapsed.
17 Comments - Last post 53 minutes ago by SketCZ
1,456 Comments - Last post 1 hour ago by eeev
2,549 Comments - Last post 2 hours ago by Marius11
234 Comments - Last post 3 hours ago by Volcanic
111 Comments - Last post 6 hours ago by mourinhos86
7 Comments - Last post 9 hours ago by EvilAaron
81 Comments - Last post 9 hours ago by ahha
5 Comments - Last post 9 seconds ago by Aydaylin
63 Comments - Last post 17 minutes ago by SolubleCrow
504 Comments - Last post 44 minutes ago by CarLNC
1,400 Comments - Last post 45 minutes ago by LycanKai
67 Comments - Last post 55 minutes ago by sunmi
57 Comments - Last post 55 minutes ago by LycanKai
97 Comments - Last post 1 hour ago by Masafor
Anyone care to catch me up on exactly what Deepseek is and why it's crapping all over my investments? From what I understand it is some kind of cheaper alternative to chatgpt. Beyond school kids asking these things stupid questions in an effort to stump them or get bizarre nonsensical answers to plaster all over reddit and their blogs I don't really get it. What am I missing? Why are these so important? I get that AI learning is certainly a big item moving forward but right now it just isn't all that impressive. Just seems rash for Nasdaq to drop over 3% in a single day over what is probably a very limited cheaper AI butler.
Comment has been collapsed.