Have you played Arc 3? It seems like more of a simple optimization problem (think Sokoban) than anything approaching fluid intelligence. Whether a multi hundred billion dollar company would spend time benchmaxxing a highly publicized benchmark that claims to confer AGI is an exercise left to the reader, but I doubt Claude Plays Pokemon is suddenly going to get past Mt. Doom now.
Just like we trained on human language to create LLMs, we can train on human keystrokes with a similar algorithm and spit out believable (at least statistically) "human" keystrokes generated by machine.
We have probably hit a limit to scaling LLMs through raw parameter count alone, at least we're not seeing the exponential pace. I personally think we'll end up with a nice sigmoid curve plateauing in the sub 10T parameter regime. The amount of tokens processed (in inference) is increasing exponentially though (I've been following open router usage stats for years and it's always been exponential). We will of course make technological advances in hardware efficiency, and model parameter efficiency, but I think a much more plausible future is that VRAM needed for loading and serving individual models will slow down or even stop. We will need more chips, and more power, as demand continues to grow of course, but the operational lifetime of GPUs today will be a lot longer than the SoTA cards from 5 years ago.
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.
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.