You can also learn for your own amusement, and solely for the fun of comprehension; a lot of mathematician were driven by this. It's a shame current social value places so much utilitariaism on learning.
I like baking. There are machines that can bake bread at an industrial scale that I cannot compete with. There are home kneaders that do much of the work very well. I use one of those more often than not.
I still think it worth my time kneading dough by hand. It teaches me the various properties of flour, how external factors like humidity or temperature impact the overall process, and I believe that it makes me a better baker, even when I use a machine, because I am better at controlling what the machine does. When I get a new brand of flour I will make sure to bake everything by hand first to "get a feel".
Kneading by hand is also very relaxing to me. This is probably the main reason I bake in the first place.
Programming, and other activities are not very different. We now have machines that can do it faster, at a fraction of the quality many people deem good enough. If you hate coding, that is probably all you need to use, and learning a new language might just be a frustrating experience not worth subjecting yourself to. But if you enjoy coding then it should make you better at it, even when you use the machines.
As always, it is good to be aware that many people are too anxious about more existential issues to comfortably an consistently expend effort on intellectually-taxing tasks that aren't perceived as directly related to their survival. If it's an issue with social values, it's less of one related to learning as it is to perpetuating artificial scarcity.
>to comfortably an consistently expend effort on intellectually-taxing tasks that aren't perceived as directly related to their survival.
the average person spends 6(!) hours a day on their smartphone, the average TikTok user spends 100 minutes on the app alone. This isn't about artificial scarcity, it's about the average person looking like the Wall-E people
Escapism and making a concentrated effort on something are two different things. Yes it would probably be good if you could flip a switch and use the small windows of time we look at our phone per day to study OCaml, but it's not really realistic..
People who are asking these questions are saying "will me spending my time learning OCaml help me land that job that pays six figures and has health insurance so I can not rot away living on the margin". They aren't saying "I only do things that will make me money".
>use the small windows of time we look at our phone
again, it's not a small window. It's six hours. That's almost half your waking day. People spend virtually their entire leisure time rotting away on low quality entertainment.
>will me spending my time learning OCaml help me land that job
that's a pointless question for one you never know if something will land you a job, new opportunities don't open up before you do something, secondly the relevant question is, should I stop doing X and start learning Ocaml, or Chinese, or take a welding class because all of that even if it doesn't work out beats scrolling through Instagram.
I don't even take offense with the idea that you engage in activity that makes you money, because pure selfishness on that front would be an improvement to what most people are doing now.
“Low quality entertainment” is a very elitist thing to say. Most don’t have time or luxury of even being introduced ( via an east coast liberal arts college ) to read Tolstoy or appreciate the finer motions of Tchaikovsky. Instead the Druski memes will do just fine, mixed in with the AI slop. Or maybe they’re physically or physiologically incapable of enjoying the outdoors or sport. Point is, my gen played strategy games and first person shooters and listened to Eminem, this one marls TikToks until their thumbs have RSI.
As for OCAML vs not, I think the vast majority of even intellectual and studious people would be better served trying to AI max and build some kind of agent serving businesses than trying to get a job at, uhh, Jane Street. 1% of the best engineers in the world get to work in that language, so yeah the parent makes a valid point
Even if those statistics are correct, I do think you're ignoring that smartphone apps are specifically engineered to be low-friction and highly-addictive, and pushed specifically to the people who are anxious about spending money on other pursuits (free time often being necessary but insufficient), as they contend with the artificial scarcity (which this is absolutely about) of affordable housing, affordable food, affordable transportation, affordable guided education, etc.
My experience with OCaml has transformed how I think about programming and complex system design. This may also be true if you learn any other functional programming language, but OCaml is easy and flexible which makes it good imo as a door towards the more formal part of comp sci. Even for steering an LLM I think this might help.
My experience with lisp was exactly that: it completely changed how I think about programming, but much more, how I think about systems and engineering. (I learned it through SICP)
Michael Clarkson teaches OCaml at Cornell. I highly recommend his free course materials [1]. He’s an excellent educator. Learning functional programming paradigms had a major influence on how I design programs. Clarkson also taught snippets from the Pragmatic Programmer, which was equally influential (as it has been for many many others) [2].
You could have asked this question 10 years ago, long before llms. It's not like you'd realistically would get an ocaml job back then when there's so few of them, so why bother?
The answer is still the same as well, people learn ocaml either because they enjoy it, or because learning a functional language makes them a better programmer overall and teaches your brain to approach a problem in a different way.
This mindset (i.e. LLMs are available so why do i need to learn anything?) is highly insidious and will destroy your brain/mind/future if you let it.
Humans are the ones who Understand, while LLMs only Know.
So inter-disciplinary/cross-disciplinary insights, new modes of thinking/reasoning, flashes of insight etc. are still in the purview of Humans only. AI/LLMs can help focus and short-circuit the study of various subjects but their understanding can only happen within a human "Mind". If you do not even have basic domain knowledge (i.e. unknown unknowns) how can you even prompt/query an LLM for answers?
A few illustrative examples; a) Newton came up with limits/calculus out of a need to measure continuous motion with varying speeds b) Kekule came up with the benzene ring from a dream where he saw a snake grab its own tail c) Descartes came up with the cartesian coordinates in an attempt to solve geometry via algebra etc. Each of these was a novel leap of insight bringing together various concepts to create entirely new knowledge domains.
So one should learn/study the core concepts/ideas in various domains and then push the tedious mechanical labour onto the machines. In this regard see also the concept of "Active Learning" - https://en.wikipedia.org/wiki/Active_learning
Cultural inertia: until recently, you couldn't just press play and get even a wide selection of music: for most of history, if you wanted music, you had to make it or hire someone to do it for you.
More recently, you had to go to the store and buy it, which meant you didn't have much variety.
Today, learning an instrument is for social status, inheriting the shine of the past, where music was rare and costly. The reason to learn an instrument today is because the former situation was romanticized.
It'll probably take a generation before people ease into guilt-free enjoying infinite, fully generated music.
Sure, I guess. But today, something like 30% of people play music or sing regularly enough to say they do it (ie, not very much at all). Even a couple of generations ago, it was much higher. It's not going to die out, but I think a lot of people are asking themselves if they want to bother.
Seeing that you already know programming, I'd say it'd be less risky for you. But the only reason you're able to pilot an LLM to do programming for you, is because you understand programming and architecture.
But what about the future generations skipping the step of learning the OCaml's, the C's, the Python's...? It's quite concerning.
> So (asking genuinely), should we learn these things?
I wanted to learn a functional programming language with powerful type capabilities and I chose the Lean Language for that and not OCamel or Haskell. Reason being: Better type system (dependent types!), applicable in formal domains and can use it to learn math too.
For your bread and butter programming, there is already JS/Go anyways.
you are eventually going to have a very bad time if you do not have a solid mental model of the code the LLM is writing, and indeed if you cannot steer the LLM so that its code conforms to your mental models. learning ocaml is a great way to add some valuable tools to your toolkit when it comes to thinking about code and how it fits together.
I think it's helpful to have a deep understanding of one c-type language, one lisp, and one ML-type language. There are so many things influenced by these three language families that being comfortable with them makes it so much easier to understand a wide variety of languages and libraries.
I would not make that dependent on LLMs. If OCaml covers a use case you have, why not.
Personally I try to stick within my own niche though - ruby, java and also python (ruby is unfortunately losing grounds really hard now, the writing was on the well in the last some years though, and people such as DHH are now indeed a liability rather than an asset to be had, but that's a side topic).
I think what LLMs will force in the long run is to make programming languages used by real humans in a traditional way, more effective. That is, writing code by humans will have to become a lot more efficient, both time-wise and speed-wise. And for that there is always a use case IMO since LLMs are, despite the promo, incredibly stupid.
Didn't know the "Highest scoring EU model" has a low bar, lower than Qwen3.8-27B, but still congratulations on the milestone, hopefully next iterations will get better from here
It is surprising given how many parameters it has that it scores so low. But, hopefully this will build up domestic talent and understanding and let Europe compete on the world stage with this.
3.8-flash-next quantized in a "large" Q4 that just fits in 128GB RAM even more so, in how close it can get to state of the art in a number of benchmarks. Or a large Q8 version of it that fits in under 190GB. Competing against things that are closed weights/opaque information about the model and might very well be 600B+ in size.
I don't think proposed plan with tariffs will be effective, because once scaled up and generating enough revenue, MBA will come and outsource everything, to make more money
what if it was because of quantization and they haven't released the new benchmarks for it?
Anything which changes the model needs new benchmarks I guess to compare with other models, otherwise you can benchmark Fable, and distill it to student model and keep claiming this is the Fable model
ARC Prize has retested Luna after the discount and validated identical performance.
(Also, quantization isn't inherently bad or damaging when done properly, e.g. QAT).
These APIs are used heavily by enterprises at scale; with lots of performance telemetry, live evals, etc. You can't really silently nerf API models at scale without people noticing.
Of course, what I said doesn't apply to non-API consumer sub models; there's many documented and officially confirmed instances of under-the-hood "juice/effort" adjustments. (Juice = a number your effort tier maps to underneath the hood; much like Inkling's effort=0.00 to 0.99).
you are right, but we also can't control other people who are posting there, lots of AI/ML news happen there because where else can you post them? I don't think instagram, snapchat or facebook is good forum for such news / one-off thoughts.
If the populist campaign is to Make Affordable DRAM Again, then it's not a terrible solution.
The current datacenter owners love a compute-bound world anyhow. A moratorium on new datacenters would increase their valuation, encourage efficiency and make computers cheap again. If Chinese labs can ship frontier models under 1T parameters, why not American labs too?
If you think tanking Trillions in investments, warming the earth and increasion ocean water levels, creating water shortages and brown-outs is "good for all of us" - well, the rest of us beg to differ.
these GPUs make computation faster, I understand as of now maybe all the computation is used to generate yet another junk LinkedIn post or unnecessary RFC, but at some point this craze should settle and we will be left with powerful computation machines, which can be used for computing more useful things
The GPUs being paid for w/ billions in investment will be obsolete and e-waste in a few short years same as a Cray-2 was just a decade after its release.
It's fine if you're one of the people selling shovels to gold miners for a while, but sucks to be building houses in the boom town?
I am not selling shovels, I just like to see when there is real competition on the market and I hate regulatory capture what Anthropic is trying to do for models, because they are scared of Chinese open weight models.
In terms of GPUs whole world with 8B people have only couple of viable options: Nvidia, AMD, Intel - and largest part of their inventory is going to enterprises to run those LLMs, and its impacting every consumer / hobby projects, like cheap phones, DIY electronics projects and so on.
The Strix Halo and DGX Spark are pretty danged slow, relatively speaking. I don't recall exact numbers, but with MoE models in this size ballpark (Laguna S 2.1), I seem to recall I was seeing about 20-25 t/s with a big context, which is close to usable. Qwen 3.8 27B crawls on this hardware, though, at 10-16 t/s, definitely not comfortable for interactive use. (Though this makes it seem like you can cook pretty good with a 4-bit ROCmFP4 quantization: https://github.com/julianmb/q38rocm the model does get notably dumber below six bits.)
A model similar in size to Laguna S 2.1, but with only 6B active parameters, should be a notable amount faster, so I would imagine 25-30 t/s would be a reasonable guess for where Qwen 3.8 Flash Next will land.
DFlash2 might improve all these numbers. It wasn't available last I was testing new models on the Strix Halo; I've only used MTP (which doesn't generally improve MoE models, but I believe DFlash2 can).
Given software improvements, I'm hopeful an MoE in this size range will be the sweet spot that pushes past 40 t/s and is also smart enough for real work. Qwen 3.8 27B is finally a self-hostable model that's smart enough, but it thinks so hard it still isn't really useful for agentic interactive use.
Note also prefill with large models is pretty slow on the Strix Halo (300 t/s, maybe). Time to first token is a painful wait, when using it interactively with large models.
No. In Linux, you can ignore the BIOS (well, set it to 512MB) and set GTT to allow almost the entirety of memory for GPU. If you don't run a desktop GUI (and why would you), you don't need memory for much of anything else. Just a few GB for the OS and llama.cpp.
I think Steve somewhat predicted the future well. He might be slightly off, because he is overly optimistic and operates at the edge, but look at Gas Town, when it was released it felt like dystopian, today it looks not too far away, I am sure most of your orgs are already running some kind of agent to triage the tickets and in some cases automatically open the PR in your git repo.
This weekend I stumbled across a codebase unlike anything I’d ever seen: Swamp Club[]. Came out of nowhere, lots of useful features, but with a coherent philosophy and architecture and a large body of working code. Change volume is high - notably so, but it looks nothing like the vibe slop I’d expected.
A “swamp” contains tools that you build - with the help of agents - intended for use not just by apps and scripts, but agents (to add deterministic behavior, like code mode). Tools are ops things, product things, whatever.
*And so it’s a tool plus a factory to update the tool.*
The factory though isn’t only used for building your tool - that’s just a set of skill.md files. It can build your whole app, which itself can _use_ your tools - either for product features or ops tasks.
*And so it’s a factory that can build any app, including a tool which can operate any app. *
Right now, looking at the commit volume of the swamp stuff, it’s not stupid to suppose that the swamp is being used to create the swamp.
*And so it’s a factory which makes, among other things, factories.*
Lately, I keep asking myself, do I need to learn this new thing, should I force myself to learn this thing, LLMs know it anyways and so on.
So (asking genuinely), should we learn these things?
reply