As someone making my own harness, this makes me sad. Pi is a big inspiration and one of the best open source harnesses, but there are many others. dsh, opencode, hermes, etc. MCP is such a thin layer to implement for any harness, this just seems arbitrary.
Look what's coming out of China, they are catching up on performance and surpassing the US in efficiency. They're on a different level when it comes to open releases of weights.
I don't, to me the entire premise is a bit flawed at it's core (ASI), and I read it like bad science fiction with China playing the bad guy just a narrative crux so we get to the acceleration timeline and warring nation states.
One of the most biased claims IMO in AI 2027 is that a huge portion of the geopolitical and existential risk argument is hinged on the notion that China just steals the US frontier weights.
I feel that quote more likely is pointed at "music appreciation" as a written work, versus understanding the theory. I know a few musicians who are happy to use terminology from theory, "I love to write songs with diminished and sustained chords".
This seems like a great resource to me, a classically trained pianist who is now playing different types of music.
Nah, that's the same sort of thinking that makes people type "make no mistakes", I don't make my model roll play, etc. I believe that the longer the system prompt and the more you cram in it the worse the model does. You need the human doing minimal prompts, but in the right direction. Take a look a the transcripts of Terrance Tao with ChatGPT
It mostly hurts people in countries with weak purchasing power. DS was the main game in down for them.
Personally, I don't think we've seen the total end of dirt cheap LLMs, it's just a frontier lab doesn't want to be in business of serving half the world.
I don't think it's necessarily "wiser" to go closed source. All of AI is built on mostly openness, at least on the software side. There are other ways to compete, it's just the model itself will be a commodity.
Compared to the Opus 5 "model card", which read like a standard Anthropic set of alignment principles and safety concerns, this presents a plethora of useful technical details that advances the state of the art.
Non homogenized, lightly flash pasteurized milk has been my favorite for years. Just tastier, luckily I grew up in a little strange town where it was norm, and I don't have the weird modern health movement baggage to go with it.
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.