- My bet is that other things being equal, where we will be in three years is threefold:
Advances will be more incremental than before and quickly matched. GPT-5 or a similarly impressive model will come eventually, perhaps led by OpenAI, a Chinese company, or maybe a competitor like Google will get there first. Whichever way it falls out, the advantage will be short-lived. Models will continue to get more efficient and less expensive, but hallucinations and reliability problems will persist. Contra Silicon Valley scuttlebutt, neither country will achieve AGI by the end of 2027. Racing endlessly around LLMs will sap resources that might go into developing more original ideas.can I ask the n00b question of the day? Since it's open source, theoretically people have verified the codebase... has anyone ensure it's not like... scanning my system for passwords, crypto private keys, etc and shipping that off somewhere? (US govt, China, someone else?) Sorry to sound paranoid or whatever - just a question from a dude who doesn't know better
Absolutely! Because asking n00b questions illustrates where your blind spots are. Fortunately for you, Stephen Wolfram wrote up a really nice explainer for what the fuck all these "AIs" are actually doing a couple years ago: But before just shoving a massive bolus of words and going "STFU n00b" a few quick answers: So the thing about AIs in general and this one in particular is "what is it trained on." "What is it trained on" is "some form of data". The AI itself doesn't scan shit - could the app that hosts it scan your phone for passwords? you damn betcha but that's not AI related, that's scam related. The company that put it together is basically trying to recreate Renaissance Technologies with modern technology. Renaissance Technologies, you may remember, was what made Robert Mercer rich, applying Markov chains to high frequency training. Robert Mercer then skipped out on his taxes, got sued by the IRS, had his daughter hire Steve Bannon and Cambridge Mathematica to put Trump in the White House and here we are so... stealing your passwords is chump change for them, they're out to steal the economy.scanning my system for passwords
So far my favorite response to the news, as will oft be the case in this era, is the white house statement: Yep. China stole this incredibly disruptive/lucrative LLM technique that we all just found out about today from us. However, I will note that the timing is uncannily optimal to disrupt the U.S. and catch the Trump admin with their pants down to the tune of $500 billion on day 3 or so: I am going to seem like a wizard at work. For about four months, I have been telling all of my bosses that the bubble can't inflate indefinitely, and anyone interested (which is me) in gobbling up GPUs for a computer cluster on the cheap should be ready to make a move. Personally, I don't think this is the bubble completely popping, but it's certainly a big loss of hot air (harhar), at the least. Shoot, I thought I had another few months to negotiate some type of purchasing contract. I think I still might. We'll see, tomorrow should be another bloodbath for U.S. AI as this sets in. Yes the fact that energy consumption was eating into GPT's profit margins in the U.S. at this moment was always the most glaring issue, and my concern isn't for OpenAI. It's for the climate. Regardless of how you feel about AI, it's simply disqualifying. The only pro offsetting carbon footprint is... billionaires make more money? This is all more than a little pathetic for perhaps more than anyone else, Sam Altman.“By stifling innovation at home and failing to cut off China’s access to American technology, President Biden created an opportunity for our foreign adversaries to make gains in AI development,” a White House Office of Science and Technology spokesperson said in a statement.
It's gonna bounce back tomorrow as people realize that they weren't expecting AI to perform, they were expecting AI to pay them. Having a model that really sucks at not getting tricked is a net good for the world; it will Shrimp Jesus everyone into recognizing that you're usually best off assuming AI is lying crap but the Ai techbros aren't about to stop going 'bu bu bu batman' any time soon. AI got nuthin' on Bitcoin and never will. Google's AI summaries suck down about 10x as much energy as a search and if I recall, every dumb image you get out of midjourney pulls down about a thousand searches worth. That puts your dumb Midjourney cartoon at about 0.03 kWh. Bitcoin? read and weep.Yes the fact that energy consumption was eating into GPT's profit margins in the U.S. at this moment was always the most glaring issue, and my concern isn't for OpenAI.
R1 performing so well has me itching to finally try and run LLMs locally with Ollama (once I am over this flu I’ve got). I do think we will see some remarkable nonincremental jump in performance in the next year or two, considering there are more and more vectors that can be used to increase performance. Model size and training data size were thought of as the only important ones, but just the last few months have shown larger context windows (Gemini 2), reasoning time (o1) and learning method (r1) to all amp up results. I’d be very surprised if we have hit a wall in all these dimensions and were unable to come up with new ones. I guess it depends on what one calls incremental progress. Sidenote: unless I’m mistaken, the whole “they created a model on the cheap for 90% less training costs!” aspect of this is a bit misleading. What they did was to continue developing open source LLMs, which were trained at great expense by others, right? It’s still very impressive but it’s not like they started from scratch if I read the paper correctly.
What they did was demonstrate that making the gadget train itself algorithmically was pretty much as effective as running the gadget through a gauntlet of fine-tuning and tweaking, with certain important caveats. https://x.com/koltregaskes/status/1881446103180062872 Much of wypepo AI kerfuffle has been about "how badly does it fuck up when you give it stuff you know it's going to fuck up" with all the AI boosters constantly asserting "I'm sure it's just a glitch." All the buzz around Deepseek is about the fact that if you aren't even vaguely testing for fuckups, China got there hella faster and cheaper than anybody else. "Created a model" is a great button to put on it - "all models are flawed, some models are useful." I will again remind everyone that AI came for my job first. Dugan has been at it since 1974. Sabine Feedback Destroyers started showing up in the '90s. Izotope introduced "total mix" in 2010. And hey - a lot of DAWs will fuckin' transcribe now. But you don't know how to use a feedback destroyer, and if you try using it without understanding it, your room will sound like shit. I watched a $1500/hr sound mixer lose his job for trusting a Dugan card over his own ears. Totalmix was such a catastrophe that Izotope burned it off the internet and let's talk about those transcriptions, shall we? I've been watching a lot of Hoarders lately which is probably bad for me but I noticed that a lot of the transcription team was guys I know I've worked with before. And I twigged to the fact that now that I can do it in the box, all those guys are going to be typing a lot less. NOT A FUCKING ONE OF THEM IS GOING TO LOSE THEIR JOB because if you need transcription, you need accurate transcription. Lemme pull that out because i'm going to refer back to it: If you need transcription, you need ACCURATE transcription. See, I can now transcribe in the box, which means I can get transcriptions where I couldn't before. They're pretty close but they sure aren't ready to send; i need to tweak them. I can tweak, clearly. That extends my reach. Likewise, the transcription agencies are no doubt jumping all over AI transcription because it allows them to do more with less, lower their prices, increase their customer base and generally provide more for many - it's a job of terrible scutwork and experience-derived skillsets and they are CONSTANTLY looking for workers. And sure - there are outfits that are going to just use the AI without a transcription service but they weren't using the transcription service before. it's an added bonus for them. Because if you NEED transcription, you need ACCURATE transcription. You and I had an adventure whereby you recorded your fiancee's musical performance in a church. It didn't occur to me to say "by the way if you hear any annoying squeaks you want to eliminate them right away or they will absolutely dominate the performance" because I assumed you'd give that sucker a listen and flatten out anything obviously horrible. Thing is? You aren't a sound mixer - you aren't an expert - so you don't know what's easily fixable and what isn't. Each and every one of us has had a discussion about the fact that nobody can understand the TV anymore. Much ink and pixels have been spilled as to why - nobody wants to say "sound mixers aren't being hired anymore because the media companies are too cheap." It'd take me an hour to fix most bad television, trust me I watch it. But the shredditors sitting in the hot seat have no fucking idea how to do sound, they didn't train for that. They know everyone at home is just going to turn on the subtitles. Which are transcribed. By humans. Because if you need transcription, you need ACCURATE transcription and you know what? Transcribers make hella less than me. People love to make fun of closed captioning. What they don't realize is that's usually a volunteer position. It's some nice old lady down at the station, typing in short-hand real quick. She's likely to lose her job; she's mostly there for local content anyway and we all know that shit's gone. The transcription that doesn't need to be accurate is gonna be an AI extravaganza in a couple years because the local TV station doesn't care about being memed. Warner Discovery? NVidia was hyped to shit because all the techbros refuse to acknowledge that if you need transcription, you need ACCURATE transcription and since they (in general) understand exactly fuckall, including the big words in the prospectii they don't read, they're absolutely convinced they're nine months from having a robot girlfriend. It's not entirely their fault - for twenty years, advancement in professional tools has occurred because of the massive financial investment in consumer electronics. If you need a better CMOS chip for seventy million phones it's more likely to have a bigger R&D budget than a better CMOS chip for seven thousand ENG cameras, QED. A well-trained AI would argue that better consumer goods can be predicted to filter down to better professional goods. But since none of the techbros understand the professional goods, how they work, how they're used or who uses them, they're utterly unprepared to evaluate a situation where prior trendlines don't extend. Much like a well-trained AI. So the AI techbros have a choice - they can point out that DeepSeek sucks at reasoning and deal with the blowback over the fact that all their shit sucks at reasoning only slightly less or they can pivot to "this isn't ackshully an improvement" which, true, but what it lays bare is that the whole "training" thing is a fucking sham. And there goes the market. No shade - if a decent recording of your fiancee's musical performance was important you would have hired out to get it done right. As it was, it was fun and if it didn't work the biggest blowback would be disappointment. You came at it like a dilettante with a toy which was entirely appropriate. A professional with a tool would have solved those problems immediately and - if you were to try again, you would, too. There's a human pipeline between "dilettante with a toy" and "professional with a tool." We all know it, we all recognize it, and the AI Techbros have been big on leaning on "training" to convince us all that there's an AI pipeline, too. it's bullshit. It's STRAIGHT bullshit. There's "mistakes it's going to make over and over" and "mistakes that have been spackled over a piece at a time so that Tay doesn't start spouting Nazi slogans within twelve hours of meeting Twitter." What the DeepSeek paper says is that everyone else's "training" is, in fact, spackling and if you're willing to utterly disregard Tay and the Nazis you can have a model in minutes. We knew that in 2016 but Sam Altman figured he could WeWork it. And here we are. Right now? AI is, for all intents and purposes, at the "toys for dilettantes" phase. DeepSeek demonstrates that as toys for dilettantes go, Chinese crap is always going to be cheaper than American techbro nonsense. But more than that, it demonstrates that any aspect of AI that isn't "toys for dilettantes" is hand-applied spackle. And if your spackler isn't as good at your professional's job as the professional is, the toy will NEVER be a tool.Sidenote: unless I’m mistaken, the whole “they created a model on the cheap for 90% less training costs!” aspect of this is a bit misleading.
It doesn't seem like it. These chain-of-thought models kinda broke the mold. https://youtubetranscriptoptimizer.com/blog/05_the_short_case_for_nvdaWhat they did was to continue developing open source LLMs, which were trained at great expense by others, right?
Skipping labeled data? Seems like a bold move for RL in the world of LLMs. I've learned that pure-RL is slower upfront (trial and error takes time) — but it eliminates the costly, time-intensive labeling bottleneck. In the long run, it’ll be faster, scalable, and way more efficient for building reasoning models. Mostly, because they learn on their own. https://www.vellum.ai/blog/the-training-of-deepseek-r1-and-ways-to-use-itThe team at DeepSeek wanted to prove whether it’s possible to train a powerful reasoning model using pure-reinforcement learning (RL). This form of "pure" reinforcement learning works without labeled data.