mcbuilder

I consider myself to be in that cohort as well. :)

LRMs are plateauing for sure, not that there won't be gains to be had in the future, but it's not like the era of rapid progress that was the past year any more.

Yeah, it's actually the case. Researchers have shown that the models response doesn't always follow from the reasoning. Whether you consider that an internal language or not really depends on what you're speculating the neural network is doing. I think there was an Antropic paper on it.

I'm a white dude from Iowa, working in top levels of AI/ML. I'm in the minority at work/conferences. I hardly ever even interview homegrown US job candidates. I'm just saying, that the reason I think you see more people from Asia and India is the education levels of most of the candidates. I'm not faulting these other countries, just pointing out how I see an educational gap based on demographics, and one that is rising up the ranks.

I've always found this line of reasoning troubling and uninformed.

Chinese models first of all can be hosted on your own hardware, I'd argue they are way more transparent than US companies, by well releasing stuff.

Second, the "smoking gun" of DeepSeek training off Claude isn't as bad as you may think, and the amount of tokens was deemed trivial. Did you also know that if you asked Claude's it's name in Chinese it would respond as "DeepSeek" until just a few months ago until they patched it?

Third, I find it a little hypocritical to call out Chinese for "industrial-scale" theft when anyone could create Studio Ghibli style image gen photos. How could they do that unless US companies trained on copyrighted works.

Chinese are just innovating faster at this point, DeepSeek V4 is an actual technological advancement (KV Cache compression) more than a cheap clone.

The administration does have it backwards, but IMO it's more them playing into the big tech companies plans (of course they have their favorites) instead of actually investing in education, and research like the Chinese do.

I did the exact same thing except a virtualized opensense router and bare metal kubernetes on one host. The kubernetes broke and I downgraded from 32GB of RAM to 16GB . I actually may revisit the setup since opensense FRR and Cilium BGP to peer your cluster and home LAN is actually a really seamless way to self host things in kubernetes. Maybe there are other ways, maybe there is something simpler, but a homelab is about fun more than pure function.

Free Monads are everywhere. Learning Haskell at this time was such an amazing experience. Haskell has incredible library stability, the kmettoverse feels the same, my is still good enough for most situations, there are new streaming libraries but accomplish the same things as conduit and pipes. LLMs are as decent as you would expect on Haskell, and have helped me debug some situations where I would be fighting GHC usually with some flags turned out. AI has actually has been helpful in learning since in Haskell once you figure something out you solve it for a whole class of problems, the issues is sometimes figuring that one thing out it's so abstract you feel like you are hitting a cliff. Excited to be writing Haskell still in 2026, I hope it continues to avoid success at all cost.

Opencode has been a thing for a while now

Yeah, this is exactly what I was thinking. LLMs don't have precise geometrical reasoning from images. Having an intuition of how the models work is actually.a defining skill in "prompt engineering"

After using it a couple hours playing around, it is a very solid entry, and very competitive compared with the big US relaeses. I'd say it's better than GLM4.6 and I'm Kimi K2. Looking forward to v4

Yeah, financial bubble != useless technology. Maybe coding agents do cost $50 per month in the long term, but I might just pay that for entertainment and personal stuff. Like, I don't even try to vibe code my job, but in the evenings having a cool slop generator is good times.

Right now I use a Chinese vibe code plan, really good value.

This article stands as complete hype. They just seem to offer an idea of "replication training" which is just some vague agentic distributed RL. Multi-agent distributed reinforcement learning algorithms have been in the actual literature for a while. I suggest studying what DeepMind is doing for current state of the art in agentic distributed RL.

Alan Turing had a great test (not definition) of AGI, which we seem to have forgotten. No I don't think an LLM can pass a Turing Test (at least I could break it).

I think it gave up trying to solve Pokemon. :) Seriously, aren't these ARC-AGI problems easy for most people? They usually involve some sort of pattern recognition and visual reasoning.

That's wild, I thought it was referencing a popular 80s action movie by James Cameron, but yeah then I clicked, and realized it was neither.

I feel like 70% of open source projects on GitHub say written in the language that they were written in

Unfortunately the speed of AI/ML is so crazy fast. I don't know a better way to keep track other than paying attention all the time. The field also loves memey names. A few years ago everyone was naming models after Sesame Street characters, there were the YOLO family of models. Conference papers are not immune, in fact they are greatest "offenders".

Did they change the system prompt? Because it was basically "don't say anything bad about Elon or Trump". I'll take AI sycophancy over real (actually I use openrouter.ai, but that's a different story).

It certainly could be, but not all technological advancement is necessarily dystopian. You say, currently everyone now has access to this, while before it was only available to nation states who could hire teams of skilled analyst s. I mean, I agree it's scary that now a stalker could track a victim, but cars and cameras probably help as well. So, I think it's fair to challenge "dystopian", someone will use it for non-nefarious purposes.

It definitely means something, probably an app designed around being interacted by with an LLM, upon first hearing it. Browser interaction is one of those things that is a great killer app for LLMs IMO.

For instance, I just discovered there are a ton of high quality scans of film and slides available at the Library of Congress website, but I don't really enjoy their interface. I could build a scraping tool and get too much info, or suffer and use just clicking through their search UI. Or I could ask my browser tool wielding LLM agent to automate the boring stuff and provide a map of the subjects I would be interested in, and give me a different way to discover things. I've just discovered the entire browser automation thing, and I'm having fun have my LLM go "research" for a few minutes while I go do something else.

On the trip to my local recycling center I immediately did a double take when I saw a pair of Eizo GX540 monochrome medical diagnostic displays sitting there. I've got these babies hooked up and I can see myself using this e-ink mode and the grayscale nature of these monitors. Although these monitors weren't intended for productivity they are very good at editing B&W photography, terminal work, and even old films.

How many students rely on AI, is that number really going up?

My son is 14 and knows about AI (I'm a researcher in the space, so I've been mentioning advances in it for years). He seems to code with some of his peers, and it seems like normal to me (python scripts, HTML, js type stuff written the old fasioned way [written by hand or copy pasting into a notebook.exe equivalent :P]). I try to be super honest with him, and I tell him AI is incredible, but we also joke about it and I explain the incredible drawbacks of vibe coding, especially while learning.

I wonder overall though how LLMs are going to effect CS education. Will students avoid using the tools, or will they be accepted? CS homework projects were always easier to cheat on vs say fine art, since of the ease one can copy and paste code, but AI tools makes trivial work of many homework exercises that would in theory be harder to implicate someone.

Funny I was just thinking about this yesterday! Now it's on the top of HN.

I think this is old news, but this model does better than llama 4 maverick on coding.

I have a stack of T40 and T60 series in my shed. All 32bit processors, but man what beautiful machines. I kinda feel like the guy with the classic Thunderbird in his garage.

Too bad Nikon practically stopped making DSLRs

But you have all the assets of the actual finished game as well as the code used to run it, using your example. You don't get the game dev studio, i.e. datasets, expertise, and compute. Just because someone gives you all the source code and methods they used to make a game, doesn't mean anyone can just go and easily make a sequel, but it helps.

Surely the architecture released as a HF transformers python file counts as "open source". https://huggingface.co/deepseek-ai/DeepSeek-R1/raw/main/mode...

Yes training is left as an exercise to the user, but it's outlined in the paper, and a good ML engineer should be able to get started with it, cluster of GPUs not included

I feel the phrase came into common use during COVID pandemic, so things certainly felt more doom and gloom then. The connotation I think is with the type of negative content being consumed, which exasperates your own feelings.

What a guy, managed to meet him once! Sad all those American Spirits probably caused his untimely death.