You’re arguing via reductionism, and failing to explain the outcomes and emergent properties of the “stupid” system. Humans are made of atoms that are quite literally stupid, so by all means, explain our intelligence and why it’s different than LLMs. (I’m not claiming LLMs are intelligent, BTW, I just don’t think your claim helps nor believe that you can fix it.)
True for today’s static models during inference. Not true for self-supervised learning, not true during training or fine-tuning, of course. Ignores that LLMs might start continuous training in the future - there’s no fundamental or technical constraint that prevents LLM ‘plasticity’. And ignores that accumulating context/memories/skills/etc affects performance and might count as a valid analogy to what many people loosely call ‘neural plasticity’, which is sometimes casually mistaking knowledge for network modification.
> Because this entire discussion is about the release of a new model
Right. The sentence you quoted was about brevity improving with a new model. It did not suggest the model itself improving.
I’m confused why you’re stuck on this tangent. And confused why you are repeating the talking points about the model being fixed. The model is fixed - that’s true, I already agreed with you. But you don’t seem to be listening to anything else.
> It doesn’t matter what you’re trying to improve.
What do you mean? If we’re trying to improve LLM output, there are multiple ways to achieve it. A new model is one of them. Changing the inputs is another.
> You will always have to wait for a new model (like this one we are talking about) for improvements to the model.
This is true! Nobody here is disagreeing with that. The part that it seems you’ve argued incorrectly is the apparent claim that output can’t get better. Output can “improve” without improving the model.
Why did you drop the first half of the sentence in your quote? The qualification there is important context for the part you did quote. And why are you talking about “better” within a model, when the sentence you quoted was talking about 5.6 vs 5.5? The post you’re referring to did not suggest a single model could “get better”. You’ve made some incorrect assumptions.
Your comments are conflating multiple kinds of “smart” and “better”. You’re right that if all the inputs are exactly the same, it takes a new model to improve (ignoring non-determinism). But the knobs and context and harness change the inputs, and they do improve output, contrary to your claim. You’re failing to capture the distinction between what the model itself does and how the harness can boost the model’s performance. It is legitimately valid and fair to call improved performance “better”, no matter where it comes from.
This all gives me the feeling you might not have experience with or understand what’s happening in today’s harness development, and the degree to which it may be as important as the weights. There are in fact a lot of things you can do to improve a model’s performance on tasks & benchmarks, without changing the model weights. @coldtea mentioned a bunch, but the harness feedback loop, internal prompts, system prompts, skills, and requests for a model to try harder, and verify and validate it’s output all lead to improved performance, all without retraining.
I agree LLMs are stupid; they’re statistical token predictors. But somehow statistical token prediction is amazing and works much better than we imagined. The talking points about LLMs being stupid token predictors are fading now because they lack explanatory power for how good the models have become. The big surprise here isn’t about LLMs. It’s about language, and how much “thinking” and intelligence is contained in language. We don’t have a good grasp on where the line is between language and intelligence. LLMs have crushed the Turing Test into dust, and yet we don’t consider them intelligent. They often appear to understand what you ask thoroughly, can re-state it in different words, they can correct your misunderstandings or add nuance you didn’t see. All this because that’s what humans do and LLMs talk like humans.
I agree, or at least it feels insightful and right, though I can’t personally validate if it’s correct. But the big question I have is who is this written for, and what do they want to see happen? Is this to sway the public, to push politicians, to convince the Army internally to plan better, stop using contractors & no-bid contracts, or simply ask for more?
Looks like military spending is currently ~20% of all Federal Revenue at somewhere close to $1T, and it exceeds the combined spending of China and Russia by maybe 2x. Are we wanting to go back to 1960’s 50% of Federal Revenue? Why don’t we have reasonable logistics and supply lines and infrastructure with $1T?
All that, and it’s still better than just reading the book on your own. :P
Be thankful when you get the 25 year old PhD students & post-docs. They care more about teaching and remember learning the material recently and are more willing to talk & help you.
I tend to be amused that the complaints about what’s on the front page don’t seem to grasp the irony of complaining to the group of people that voted it there.
It’s very easy to check & see there’s at least one paper from 2023. Also it takes time to know which papers are influential. But better to contribute than speculate… what are the seminal ML papers from the last 6 years that should be on the list?
The close-to-home example that came to my mind while reading this is GPU programming, where the percentage multiplies. Maybe there are other similar examples where a large sounding percent needs an exponent and shrinks?
With CUDA you try to keep all threads doing the same thing. Sometimes that’s very difficult, but if each thread does the same thing 98% of the time, is that enough? Well since there are warps of 32 connected threads, you might expect the probability that any thread in the warp diverges to be .98^32, or 50% of the time spent with one thread in the diverged code. 50% still doesn’t sound that bad unless threads diverge at different times, and then 50% warp divergence might mean a 16x slowdown overall. 98% isn’t enough in this case.
Yes you’re right; I am leaning heavily on the learning part of college and maybe I’m downplaying some of the negatives. (Or maybe it’s just objectively worse on average now than when I went to college; schools have definitely gotten more expensive over time.) I admit that a degree is neither the only way to learn things, nor necessarily the best way to learn things. I agree there’s massive inertia in this system too. Maybe I’m rationalizing, but I truly enjoyed my time in school and feel like I got a lot out of it, but I’m aware that not everyone does. And even if there are stats and averages that paint a rosy picture, there’s certainly no guarantee; not everyone with a degree makes more money or leaves feeling like they were bestowed any intangible benefits.
Hey I hear you, school isn’t for everyone, and I don’t believe everyone should go, and I don’t believe anyone should be forced to go. I only think everyone should be aware of the job & financial prospect stats and their likely future options, that’s where the general recommendation comes from. People who don’t know what they should do after high school should consider college as it offers the most options. People who know what they want to do are going to ignore the advice anyway… and they should! Motivated people can and do learn volumes and start companies and make fortunes without college; if only we were all that strong.
I’m jealous; the US could afford to pay for college for all on a fraction of the additional income tax earnings from college grads. We choose not to even though countries like yours have demonstrated a clear ROI. Many (I suspect most) Americans are aware of the ROI of education as well, but we have some special frozen place in our souls reserved for refusing to give away anything for free to anyone for any reason, regardless of whether it helps or hurts either them our ourselves. :P
So to be fair, most college grads don’t waste time studying things like medieval history, or fungi, or any other pure whims, maybe aside from one or two elective classes. People have majors and gen-ed requirements, and generally stick to it. My CS degree was delayed a year because I was taking music classes that had a conflicting schedule, and I paid extra money and a year of employment opportunity cost for it. And still, TBH, my biggest regret isn’t that it cost me, it’s that I didn’t take advantage of the opportunity to study even more music, and art.
One big problem with viewing college as pure vocational training, or of an economic tradeoff, is that most people don’t keep their first job for very long. Whether you change your mind, or get laid off, or get promoted, you’re not going to do exactly what you trained for very long. This is why getting a well rounded education and forcing people to read and write and do some history and math and language are all so important. For example, learning how to program computers doesn’t make good managers. But writing a lot of papers and taking Philosophy and English does, a little bit, perhaps surprisingly.
It’s a feature, not a bug, that there’s material not directly connected to a job. This is actually one of the reasons why companies want a 4-year degree credential - they’re interested in people who are competent in more than one thing. I feel like it’s also a feature, not a bug, that college gives people time to mature. How many people change their majors? Do those people get to change their jobs as easily if they work instead of go to college? Some can, but it’s a lot harder when you depend on what little income you have and the job market isn’t great and you’re not trained for other things.
Yes. For example, when people take 4 years to learn full time, by the end they generally have learned.
> college ritual
I don’t buy that foot binding is a valid analogy to college. Why do you believe it is? Foot binding as a practice has died, while college hasn’t. Foot binding causes lifelong health problems, while college leads to lifelong benefits. Your language suggests you see no value in 4 years of education. Why?
What are the downsides to college? What is the objection to learning? Isn’t learning on the job the alternative? Some people say there’s 4 years of job market opportunity cost. Except the people who start work out of high school get stuck in lower paying jobs forever. The people who go to college get to learn & explore many topics outside of “work” for 4 years and then immediately take a higher paying job that statistically surpasses their GED working friends in 3 years, and then keep earning more forever.
Some people borrow money to attend college and go into debt. This can be a real issue for some, and I believe the average debt is higher than when I went to school. For the majority who get decent jobs, this is temporary, and the debt gets paid off, and then they end up richer.
What if university was free to attend? Does that change your calculus? Would you still consider college a “ritual” or a tradeoff if there was no direct financial cost to attending?
No you don’t need to do a study. We already know college is causal in the sense that you can’t get good jobs without the credential. The job market already enforces the causal relationship. And that’s a major reason why the advice to go to college is sound advice: we know for a fact that it broadens one’s options and enables a vast set of higher paying jobs. Whether this system is good or fair is reasonable to debate, but doesn’t change today’s economic “tradeoffs”.
What you’re asking is whether the credential represents a measurable and valid increase in skill and learning, and ideally adjusted for socioeconomic status. That’s an interesting question, and there’s a whole field of literature on this topic. But the answer to that question won’t change the economics nor the advice.
Many papers go much further than what you suggest and they adjust for things like race and family history of higher education, in addition to intelligence and family income. I have read through quite a few papers myself, and the conclusions I’ve seen vary on where they land on the causality axis between true learning and credentialism, but none of them come to the conclusion that true learning isn’t a significant factor. And it would be pretty silly to assume that people who spent 4 years learning learned nothing, right?
> it’d be disingenuous to assume the conversation is about anything other that the economic trade off.
To me it seems hollow and sad to think about higher education purely in job and money terms, but maybe I’m weird. I studied what I wanted to learn about, and was naïve about any economic trade offs. There were a few companies I thought were cool and dreamt about working for, and at some level I knew it’d take a degree to get in, but in my book school was quite valuable beyond the jobs I’ve had, in many non-financial ways.
> you’re basically telling them their economic value is shit before they are even grown. I don’t believe in that.
Maybe I didn’t understand, but I can’t quite reconcile your suggestion to only look at college as an economic trade off, with not believing in recommending college due to your pessimistic interpretation of the message. It’s a fact that in the US, people with 4-year degrees earn more than people without, statistically speaking. (And the factor is a lot bigger than I thought.) People with advanced degrees statistically earn a considerable amount more than people with 4-year degrees. If you want to look at this as an economic trade off, it seems like there’s only one recommendation that makes any sense, no? Like, based on the numbers and my own experience, it doesn’t feel like going to school has particularly strong negatives that offset the positives. (BTW I went to a state school, and borrowed money to pay tuition.)
> the assumption is (not unreasonably!) the genAI is doing all the work
Yeah. I agree it’s reasonable, but there’s a fun twist I experienced myself. I used to do a type of generative digital art… artificial evolution… long before anything like Dall-E. I wrote the procedural tools, so I was creating the imaging software and doing all the work. Some people still assumed the computer was doing all the work and jumped to the conclusion that it was art of lower value than other kinds. There is another angle about art that can be mechanically and exactly reproduced having lower value, and I think that is also valid and factors into how people see digital art.
Personally, I also prefer art made by humans. AI can easily make beautiful and interesting images, but it has little meaning and seems far less interesting. Art is often as much about the narrative or story of the art and/or the artist, about why they’re doing what they do, as it is about the artifacts. I used to think it was somewhat ironic that the art-consuming public behaved this way, but now I realize I very much do the same and value human motivation and original ideas & execution and unique unreproducable art as much as anyone else.
It looks spiritually similar to a CM-1 (https://en.wikipedia.org/wiki/Connection_Machine), but I think WOPR predates the Connection Machine. Still, watching Wargames and seeing WOPR always reminds me of a story my college hardware prof told about one of the early Connection Machines - that the LEDs were a busy signal, one for each processor. Supposedly there wasn’t enough power to have them all on at the same time, and they discovered it debugging someone’s parallel algorithm that appeared to crash the machine when, as they finally figured out, the algorithm at one point used all the processors simultaneously.
Ha! A couple decades ago I saw the original Westworld, spotted some assembly, and thought it looked like 6502/Apple II code, so I assumed that was “probably” it and thought I was a clever nerd. Now I check this list and discovered it wasn’t 6502, and then realized the 6502 (1975) didn’t exist at the time the movie was shot (1973). Reviewed some scenes just now on YouTube and I can see it doesn’t look like 6502 code at all. It does look like the assembly might be the code behind some of the animated displays that look like old screen savers that you see on the other monitors in the film, perhaps, based on a few comments & variables in the code. (For example: https://youtu.be/Luo3uEVOahw?t=2645)
> the premise of all this is basically that there’s originality based on ideas, right?
That’s part of it. But it depends. I’m not sure exactly which premise we’re talking about, since digital painting and writing apps are fairly different activities. It also depends on whether we’re making legal distinctions or just judging originality for the sake of narrative or discussion. Either way, I personally think that original ideas are only part of the equation, and that original execution matters for the purposes of art as well as for the purposes of software, and this goes if we’re talking about copyrights as well.
I’ve done digital painting, and written apps for others, as well as written tools and libraries made for digital artists. As a digital artist, I’ve had people wonder out loud in front of me whether I had any skill as an artists since “the computer did all the work”. As an app developer I give the person commissioning the work and giving the specifications the credit for the design, even though I might have to make many unspecified micro-decisions. Likewise, I take credit for the design of apps I write for myself. As a contractor, the client is the “artist” for the purposes of this conversation. I’m responsible for execution of their vision, and I can hand some of that off to AI in which case I’m giving away or sharing some credit for the execution. As the author of tools or libraries for artists, my credit stops with the tools or libraries, and artists that use those tools get all credit for both the ideas and the execution of their works.
Anyway, I mostly agree with you in the sense that apps I write are only as original as the original pieces I put in them. There is room for creativity (depending on the flexibility of the client/employer), maybe for novel modes of UX, maybe for thoughtful design and beautiful interfaces (both code and graphics). It’s also possible to make good apps without adding a lot of originality too, and there’s nothing wrong with that - it can be more efficient and better for the client/employer to not push your own ideas. That’s more engineering & business than art/authorship/originality, but still a valid lens for evaluating your identity and responsibilities.
You’re arguing via reductionism, and failing to explain the outcomes and emergent properties of the “stupid” system. Humans are made of atoms that are quite literally stupid, so by all means, explain our intelligence and why it’s different than LLMs. (I’m not claiming LLMs are intelligent, BTW, I just don’t think your claim helps nor believe that you can fix it.)
https://en.wikipedia.org/wiki/Reductionism#Definitions