Good Morning. Hello. How are you? #1422
DeepSeek analysis, cuz that's what you want from an aging ad guy who hides in the woods.
Good morning, hello. How are you? All well? I do hope so. Sorry I am late. Just back from the Orthopedist. Good to see her again after my luckless run of tennis elbow therapy. Tennis elbow: year 3, still going strong. Anyway, seems I ruptured my not-the-bad-one bicep muscle, which means I don’t need surgery, I just have to wait out the pain. Which, let’s be honest, that’s what I was going to do anyway so, you know. Nice to get official backing of my half-assed remedy.
Attention Facebook readers: GMHHAY is leaving Facebook, migrating to Slack. This will be over a few months, and we’ll be completely off by March. I love you all. Please come join us in Slack. The invite link is in the comments below. If you’re an email reader, reply to this email and I’ll get you the link.
We are listening to someone called TheBabeGabe. I do not know why or how I found this artist. She has a lot of singles. Each single has a collaborator? I guess? Including “The Human?” I don’t even know, man.
Annnnd… once again. I forgot to include a photo of the Band Aid organization. Reader revolt. Which is quite reasonable since it is probably the single greatest thing I have done in my life. Except for maybe getting Shakira on Twitter. Who knows.
It comprises of sunglasses racks, wall-mounted with double-stick tape, except that fell off of my slippery gloss closet walls, so now they are screwed in. The cases are mostly from the craft section — flash card cases, pencil cases and the like. Simple as pie. Bob’s your uncle. Anyway, here you go:
We are doing to talk about DeepSeek today, and I am sorry. But I do think it would behoove us all to have a lay of the land, so to speak. Plus, you know, I work in adtech, god help me, and this whole “AI” thing is a big deal, for dumb reasons. But here is the situation as I see it.
(If you are caught up on all this you can skip the first batch of bullets and move on to the analysis in the second batch.)
Silicon Valley is obsessed with “AI” but they really mean this new technology they have developed where they hoover up the internet to train “models” to talk back to us or make pictures or, maybe, help us with the easiest tasks that we do.
By “hoover up the internet” I mean that all of these models stole everything humantity has made and posted online to make their for-profit models. They stole your words, they stole your drawings, they stole your music. Just a giant theft. Everyone knows this. Eventually, assuming we still have a democracy, there should be a reckoning about this but we’ll see.
They have been raising just insane gobs of money for this, under the conceit that they need more “compute” and the more “compute” they have, the better the models will get.
This part’s important, as you’ll see later: Even though I do not believe this, it’s important to note that they do believe this: there is no wall on these models. They believe “this one simple hack” of mashing up all of the words on the internet will lead to, eventually, AGI, or human-level intelligence, and then it will go on to achieve superhuman levels of intelligence. Or, that it probably will. This is, of course, a completely insane notion. But they believe it.
They have, thusly, gone on on for a year or two now about needing more chips, faster chips, and more power. Each of these is relevant to our story.
More chips: the “best” chips for all of this are Nvidia chips, so Nvidia stock has gone through the roof, it’s become like the most valuable company in the world. They’re the classic “pick axes in a gold rush.”
Faster chips: Nvidia wins here too, but also there’s a geopolitical angle, because people are all worried China will make an AI. I do not share this worry, but I do share worries about what else China can do with fast chips. Anyway, this lead the Biden administration to severely curtail the speed and awesomeness of chips that China can buy legally, since “the west” (i.e. mostly The Netherlands and Taiwan and a little bit of America but not much who are we kidding) makes all the good chips.
So the AI companies — and Facebook and Google who want to be AI companies — have been buying more and more of these chips, just hundreds of thousands. Just stupid amounts of them, and making new models, “training the models” on tons of chips, and making incrementally better models. They’ve been launching a new one every month or so across a handful of AI companies: most notably for our purposes OpenAI, but also Perplexity, Anthropic, yadda yadda.
Other companies — Apple, Microsoft, Google, your bank, every god damn company under the sun — have been cramming these models into their products and forcing us to use them even though, so far (he says, generously) they are all dumb and useless and annoying. But hey, maybe someday they will be able to do useful things? Maybe? Unclear.
So this has been going on for months, the models have been getting “better” and having “giant leaps forward” but they are still mostly useless aside from a few very specific things that conveniently lead to unemployment: coding, writing (sort of) and illustrating (sort of). Actually they really only seem to be good at coding. Very Smart People insist they will imminently get better at more-or-less everything. Maybe they are right. Maybe they are wrong.
It cannot be stressed enough just how much money these people are raising for this bullshit. They are talking insanity like “building 10,000 nuclear reactors” using more chips than humanity has ever made. Just absolutely insane talk. Because they believe — and this is important — there is no wall in sight. Dan Primack said “The quantums of capital are just so much more than anything VC has ever before disbursed,” the Trump white house bragged about ten billion dollars or some shit for data centers.
Data centers and chip fabs that will, let’s not forget, break the planet but they ignore that part. All of the older companies involved in this (Meta, Microsoft) used to talk about how they were going to become green and make their data centers green and not kill the planet, but they have walked all that back. They are out to kill the planet.
So this has been the state for a while, just the worst state an industry can be in, people all ganged up together participating in a mass delusion in order to steal more money from the government (national security!) and pension funds and everyone else.
And! The fun part! This week it all blew up! Not, sadly, in the way I thought it was going to blow up, and not, sadly, in a permanent way. BUT still. Good times.
A Chinese hedge fund announced they made a new model, called DeepSeek, that is better than the best new Silicon Valley models. Which on its own is hilarious but also Silicon Valley would have just pretended this meant they needed more money and more “compute” and needed to kill the planet faster. And they would have gotten it. They were already calling it a “sputnik” moment. Really tryin’ to spin it.
Except then people read the paper and realized a few things (sub-bullets!):
The Chinese hedge fund made their model for way way cheaper than the Silicon Valley models. The exact amount is unclear, the paper is a bit evasive about the true all-in cost, but it’s clear it’s like orders of magnitude less.
The Chinese company (supposedly) made the model on older Nvidia chips, ones that are still allowed in China. This turned out to be maybe not true, or not completely true, but (sub-sub bullets!)
They definitely did a lot of awesome, low-level efficiency coding at the systems level of the Nvidia cards to get every last drop of juice out of them. Because they were being thrifty. It never occurred to OpenAI to be thrifty.
They definitely did a lot of the work on older cards.
Even if they did use new cards/chips as well, they only used, people think, 10-50,000 of them (which the hedge fund supposedly bought back when it was legal) which is, relatively speaking, a drop in the bucket
They open-sourced the model, and gave the world an version you could run on a laptop (people are checking to see if this is a secret back door, but so far it seems to be legit), which would mean, hey, if you have a free model, who needs OpenAI?
The Chinese company totally used “distillation” which means they queried and hoovered up data from the current “best” models, OpenAI, Anthropic, etc. This is hilarious because just this morning I got the joy of watching OpenAI, the perpetuator of the largest intellectual property crime in the history of mankind, whine that someone stole their intellectual property.
They also used “synthetic corpuses” — as in, yes, they hoovered up the internet, but they also hoovered up fake internets. This is cheaper. The research papers think it will work. OpenAI, etc has already been exploring this, this is not really an innovation of theirs. Personally I think this is destined for a global ELE, but that’s just me. In terms of our story here, it is not super disrupting.
So, everyone freaked out and shorted Nvidia, since OpenAI, Anthropic, etc., are private and you can’t short them.
The thinking here was that if you can do all this work on old Nvidia cards, and you can use them more efficiently, then people aren’t going to need more Nvidia cards.
So that was all very fun to watch as a schadenfreude, but I do think it was misguided. But you gotta unpack things a bit to see exactly how.
(If you are caught up on all this you can pick it up here).
Nvidia is not screwed, at least not by DeekSeek, at least if you are an AI true believer.
Because, remember: none of these people think there’s any upper limit to these models. As in, they think the models will keep getting better, with more compute and more power.
Deepseek did not change this conceit. This is important.
DeepSeek showed that you can get more out of each card, but it did not break the foundations of all of this: that there is no upper limit to the models’ potential power.
We can safely assume that OpenAI and Anthropic and everyone else will eventually read this open-source paper, learn how to do the low-level coding required to eek even more out of each of their Nvidia cards, and they will — if you believe all of this AI shit — then make an even better model.
Nothing DeepSeek did actually permanently makes it so that “more chips + faster chips = better model.”
More chips and faster chips are still gonna “matter” (if you believe in LLM-based AI).
SO, if you believe all of this shit, Nvidia should be fine.
AND, theoretically, so should OpenAI and Anthropic and Facebook since, theoretically, they will still have the most money and most compute to make the best models. This did not change!
The actual practical value of these models, what they can do and what you can charge for, has an unknown upper end. It is clearly not zero, even I have to admit that. I mean, I have thoughts about people who pay for Chat GPT, but there is clearly a market there. I will be an optimist and assume it’s because a lot of people who can’t write well wish they could. Fine.
Can bigger, better models do more and charge more for it? Unknown. OpenAI has a new model and product, Operator, that “does things for you” at your computer. It costs more. Some people are paying. Will it last? Is it useful? Probably. I mean, I don’t want it, I think productivity is a trap, but people do.
BUT, DeepSeek did take the bottom out of the market. At this moment, the $200 (or whatever) version of ChatGPT that “does stuff for you” is safe from DeepSeek. For now. But the “writes better for you” and “codes for you’ and “steals the work of all the artists on the planet to make you a comp that you could have gotten with a Google Image Search and Pinterest” will be free. But, then, it already was.
SO where are we then? Assuming you believe all this AI hype?
Nvidia is probably fine. They were the wrong people to sort. yes, I own $10k — or probably $5k now, lol — of Nvidia stock but I bought it cynically six weeks ago, knowing full well we were probably at the top of the market. In my head I did all of this, I am very proud. I am kidding. Sort of. Anyway, they’re probably fine. Chips are still gonna matter.
It’s a bit of a stretch to assume that someone out there will keep buying lots of Nvidia cards and make the lowest-cost products free, but not that big of one: Facebook is doing it, DeepSeek is. We can assume there will probably continue to be parties commoditizing the lowest-hanging fruit offerings of the “Big AI” companies.
This will keep them from ever having a long-tail of low-paying customers. Cuz that shit will basically be free.
But presumably they will be able to keep offering ever-more-useful functionality that the bottom feeders cannot offer. And they will be able to charge a premium for this. The new “do stuff for you” offering from OpenAI apparently already makes as much for them as the “write shit for me” version.
Buuuut this is gonna make it harder for OpenAI to keep raising gobs of money, because half their market is gone, and half their market will always disappear, every time they make a new leap. They are going to be even more of a high-wire act.
It’s very hard to imagine OpenAI having just as easy of a time raising as they have, given that half of their revenue base is going to continually disappear.
Facebook and Google can afford to keep going but… what’s in it for them, really? All they have to do is match whatever OpenAI/Anthropic can raise in this new world of halved-total-addressable-market.
Sam Altman is a modern-day PT Barnum. He is a great showman. But I do not see how he will manage to raise more money out of this mess, and seems safe to say he will raise less.
It’s possible all of this will kill off OpenAI and Anthropic, but I suspect there are enough true believers out there that believe in imminent AGI and superhuman intelligence that they will believe that the upper market, the high-end products only, will be enough of a business.
Or they’re just rich-ass zealots and won’t care either way cuz they are trying to get to the point where the world burns. So I suspect at least one party will live on, working the high-end. A Ferrari of AI, for the rich people. Sound familiar? Great, right? We’ll all get the free AI tools and they’ll get the posh ones.
Of course this assumes there is some useful tool to be had out of contuing on this path. I don’t personally believe this, but enough rich people do. One of them will live. Unless Google or Facebook grudge matches them to kill them off, and keeps up with them and then open-sources everything. Not inconceivable.
It is an absolute indictment of Altman, OpenAI, and every other American AI company that this Chinese company had to come and show them how to actually code for the chips they were buying. Never believe any of these people when they claim they have good engineers who are 10x. They obviously dropped the ball and people should be fired but lol no one will be.
That’s all I got.
The new Mogwai is amazing.
My daughter was on best behavior last night. It was great.
And you should buy some girl scout cookies from her!
Got a moody and quiet playlist for you today. Lotta stuff you’ve been seeing on other playlists lately. Love the Dawn Richard Spencer Hahn album. Still on my Anders Parker shit. And the XTC of 1999 ish. And Circus Trees, god, Annie was so right they are so good. Got Arooj Aftab from Pitchfork they still give good recommendations now and again.
Talk tomorrow about non-computer things. I am sorry.
—
Thanks for reading.
And hey! Maybe buy one of my books!
Good Morning, Hello, How Are You vol 1.