Just a quick note to bring some perspective to the tsunami of takes on DeepSeek. Keep in mind that it's open source and has been extensively and painstakingly documented. If it's as good as experts are saying, every one of its innovations will be adopted by other AI companies within a few months.
And then the race will begin all over. There may be winners and losers from this, but no one will ever have a permanent lead.
I still don't get your AI hype. Maybe post how you are using it currently?
Are you sure that Kevin himself isn't a robot? What else explains his extreme partisanship on this issue?
A lot of the reaction to AI is coming from people who, if they had been around 200 hundred years ago, would have been complaining how they don't see what the big deal is with electricity. Sure it's pretty, but so what?
This analogy would only be apropos if every time you tried to turn on a light, there was a high probably it didn't work, and even making the attempt cost an enormous amount of money.
DeepSeek might solve the money problem. Cool! Can it solve the "it doesn't work" problem?
Deepseek doesn't solve the money problem. If you had given them more GPUs they would have used them and consumed a bunch more electricity and money. New efficiencies just let them scale up and eek more out as they do.
They didn't have light bulbs two hundred years ago, they barely had Frankenstein.
I mean. In that case I'm not sure what you're saying.
You seem to be making the case "nobody knew that electricity was useful for anything after it was quantifiably discovered, but it turned to be something utterly transformative; our entire world revolves around the electron today. AI will be the same way; sure, it doesn't have a visible, viable use case now, but just you wait!"
If that's so, I guess my response is "people who didn't care about electricity were correct to not care until visible, viable use cases presented themselves." If that isn't so, then I don't know what you're saying.
And electricity wasn't burning vast oceans of money and resources that could be put to better use. I would be much more tolerant of AI if it was some guys burning a couple hundred million in R&D in a back office somewhere on spec, rather than people saying insane stuff like "let's build some nuclear plants for our AI datacenters" who might have the juice to make it happen.
DeepSeek has just demonstrated that it can be done much less expensively, and that it will become increasingly less expensive seems probable.
For most people in the 1820s electricity was a music hall demonstration, and magic seemed it's obvious application, and the people who said there was more to it were often met with skepticism from those kinds of people who want to be seen as tough minded and skeptical, the 'practical' personality.
It seems a straightforward analogy to the kind of reaction AI is getting now.
Except that is basically an ask of "trust us. AI will transform the world like electricity did. We can't prove this, but electricity did, and ipso facto so will AI."
To which my response is "this is an evidence-free assertion, a demand to be taken on faith alone."
Cool! Does it solve the "it doesn't work" problem?
You don't have the kind of ten-mile high perspective that would let you say it can't or will never work. AI has only been worked on at this level for a very short time, far too short any insight about it's failure, but we do have eighty years of experience of computers, just as the 1820s had about a century of experience of Leyden jars. You're looking only at the jars.
it's = it is
No one is saying it will never work. That's you being butt-headed oppositional because you're refusing to provide any evidence that it will. Need I remind you that insisting others prove you wrong is considered extremely poor form? Yes, apparently I do. Repeatedly.
If someone is saying I'm wrong it really is on them to say why, and if they're wrong about that, what should I say?
But on the far more important topic, you,
In every one of our interactions you will consistently conflate me with some other people you've been having an argument with.
This is a significant error and suggests you have nothing to do but spend your day spoiling for a fight.
No, it means the burden of proof is on you to show you're right, not on others to prove you're wrong.
Taking this admonition personally says a lot more about you than it does about me, but then again, this anger that people don't simply take what you say at face value seems to be a thing with you.
"Cool! Does it solve the "it doesn't work" problem?"
In what sense does it not work? That it sometimes makes errors? That can still be useful.
I am a late adopter when it comes to technology but I do have a cell phone and I sometimes type messages on it. The forward predict function (which I think is a simple AI application) works well enough to be useful.
Another example is machine translation which again seems good enough to be useful.
This is completely wrong.
It's written with no apparent knowledge of the history of electricity or the early use cases for electricity.
Electricity wasn't being worked on in hopes that someone figured out good uses for it one day. And it wasn't being hyped as a magic all powerful wizard/demon/god that we must both fear and be in awe of (as soon as we figure out what it's good for).
The attempted analogy to the current AI hype is a really, really bad one.
This type of wish casting is very apropos of the ridiculous AI hype.
It's just like electricity if you know nothing about the history of electricity!!
It is an obvious analogy, obvious unless you don't know anything about analogies.
If electricity didn't have any interest why was anyone working on it at all? The present day hype around AI is more a character of our era, a character that didn't exist at all in the 1820s.
Though maybe steam power would be a better analogy, there was a lot of money thrown at that and all the things that blew up in the first few decades didn't derail it.
"If electricity didn't have any interest why was anyone working on it at all?"
I believe you think that is a devastating critique. I really do.
Edison predated much of the advancement of "practical" electricity in the 1880s
I mean, right now it seems like more inference-time compute can solve the "hallucination" problem. If you cut the natural inference-time compute of the model by 25x, you can then do more inferences, which can then solve the hallucination issues. Say, do 5x as many inferences, eliminate the hallucinations, and then get 5x benefit on the electricity usage.
Will it work? I don't know, but I think what Deepseek has shown is that Sam Altman's theory that "we don't need more million-dollar engineering efforts to improve our tech, we need trillions for more datacenters with the tech we already have" never really made much sense.
Why would additional inferencing solve the hallucination problem? These models use likelihood as a proxy for correctness. Hallucinations arise when the most likely output text, given the prompt and the training dataset, is incorrect. Not sure how additional inferencing is going to bridge that gap.
This is exactly the argument I used when people didn't get how revolutionary the Segway was! Or Elon Musk's tunnel digger thing. Because we can imagine people a long time ago being skeptical about a thing we all like now, all skepticism must be foolish!
Neither of those things had anything like this much money and expertise thrown at them from all over the world, and neither of them produced any reaction aside from a cloud of the underwhelming.
Skepticism isn't foolish but that comparison certainly is.
"Ha ha, you didn't prove that I was wrong! That's what I've decided and that's the way it's going to be."
Why are you trying so hard to be a dick?
Because we can imagine people a long time ago being skeptical about a thing we all like now, all skepticism must be foolish!
There are just as many examples of the opposite: people being skeptical about human-powered flight, talking motion pictures, television, the internet....
I've already found it's become markedly easier to gather information and do research thanks to AI, and it's personally made one major aspect of my job (editing, proof-reading, revising) a lot less-time consuming. Youtube, a $50 billion division of Alphabet, is overrun with videos clearly made with AI technology. Every day software gets written by LLMs. There are all kinds of examples of AI's rapidly expanding role in the economy. By some measurements, the LLM has emerged as literally the most rapidly-adopted technology in history.
I think it's fair to be skeptical with regard to some of the claims being made about AI (I've long been skeptical it's going to trigger mass unemployment, for instance). But at this point it seems sheer obstinacy to ignore its increasingly massive importance to the economy, and to wider society. If anything, Monday's trillion dollar stock market plunge wasn't a sign of "hype" but of AI's increasing centrality in the marketplace.
It may not in the end live up to the greatest hopes or fears surrounding it. But AI is a big deal.
Not a general condemnation of AI, but yes, "Youtube, a $50 billion division of Alphabet, is overrun with videos clearly made with AI technology."
I sometimes watch some car oriented YT videos. Suddenly I'm getting all kinds of suggested car videos that are full of factual errors in the narration, illustrated with often incorrect photos and video clips. Viewers as shown in comments recognize them as being AI created. They are just garbage clickbait created for little cost by AI - someone just writes a suggestion and AI does all the work. It's really ruined paying any attention to YT suggestions - you have to be skeptical of all of them now.
If this hasn't happened in other topic areas yet it will. If the YT masters have any brains they will do something about this before the whole thing is ruined except for ignorant viewers who will watch anything.
I'm seeing that constantly on YouTube.
It's really offensive when the narration, er 'narration', just repeats itself or even contradicts itself from one minute to the next.
Elon's "tunnel digger thing" is not really "Elon's". TBM's have been around for fifty years; all The Boring Company has changed is to militarize the process and restrict the choice of diameter.
Notice that there is no "X" in the company's name? It's a loss-leader to sell Tesla mini-buses.
I think its for his Mars colony. A Mars colony is going to need power and electric transport (Tesla) and habitation (Boring Company) and communications (Starlink). That's why his big deal is making sure that the TBM can angle in from the surface.
AI is only useful if you also have robots. Talking to a bot is not interesting to most people.
When you cut cost structures by an order of magnitude, then it is essentially a whole new product.
At this point the proposed cost models, and approach to ensuring an effective monopoly, are changing drastically which will change the investment and deployment profile.
At this point every computer science department in the world has downloaded Deepsweep and is preparing research proposals to companies for low cost alternatives to the big AI players. And this genie can't be put back into the bottle.
And there is effectively nothing to invoke tariffs on or embargo. It is open source with enough copies floating around that it will be impossible to contain them.
This will be fun to watch.
Here is a Youtube video of a smart guying putting DeepSeek through its paces, including doing useful work.
https://youtu.be/bOsvI3HYHgI?si=V_JbrPBOb5-5ClB-
I think the utility of AI falls in two general categories. Stuff you can do in general and stuff you do personally.
The video link above gives an example of something you can do generally. He gives it a task to write some software and DeepSeek does it. He then copies the software and runs it. Fairly impressive.
I've used the app on my iPad to do similar things. I was curious about a certain math problem (adding Poisson noise to some calculated data) and it did it nicely and explained what it did. Great learning experience. An interesting feature of DeepSeek is that you can ask it to explain its reasoning step-by-step.
My fried uses Co-Pilot for the personal use case. He has an account setup with a calendar and other information. Then he can ask questions about his schedule or add things to the schedule as they appear. Not that hard to do by yourself but this highlights what the AI can do for you. It can act as a sort of secretary to manage incoming emails, calendar and such.
From what I've heard, DeepSeek is not the be-all end-all. It is not fully capable as some other models yet to be released. Nevertheless, it is very impressive for what it can do, especially just using relatively simple hardware. We have just seen the dam break. Within six months this will be pervasive. It is a huge threat to the established players because it (or the knockoffs and forks) is "good enough" for many applications so those users will not pay for anything better.
I find it interesting that AI posts seem to attract one-off commenters who have generic usernames and post about how revolutionary AI is going to be.