Bookmarking this for later when I can say "I told you so"

I think Chinese people, history, geography, culture, and philosophy are all really cool. CCP? Not so much. I honestly think Chinese and Americans could get along really well in the grand scheme of things. Both cultures are extremely ambitious, albeit in different ways. And, whether most people realize it or not, we've already learned a lot from each other. I think the post-AI race world will be quite different from the Soviet Cold War. Where Russians all hate/envy America and just want to be more American, yet blindly say America is the cause of all their problems, China is much more high agency, which is something we could come to admire about each other. There are some deep differences and antagonistic things. For instance, both China and America think they are the center of the world. That can lead to some amusing disagreements and cultural clashes. Then there's the long mistrust and abuse (Century of Humiliation) which yeah, understandable that many Chinese dislike or mistrust the West as a whole (to be fair, that was the British Empire, not America). The funniest thing, I think, is that both China and America have this philosophical narcissism that "we are the best, we know how the world REALLY works" and in some respects our values are aligned, but in some respects they couldn't be more different. I say this all as someone who understands and believes in structural realism and the geopolitical realities of today, but I hope that America and China can be friends one day in the same way that we're now friends with Japan and Germany (I am NOT suggesting after a major war! Just that people can get over their differences!)

Sometimes spotting autism in the wild is easy.

Whenever I see people on these things (and the one-wheels and scooters) without a helmet I wonder if they have a deathwish.

I have a sneaking suspicion that agents-managing-agents is going to take a while longer to figure out. They tend to obfuscate too many decisions and all fall into the same traps. I am sort of reverting back to "I'll just manage the chats myself, thanks" even though it is far more time consuming. Biases stacked on biases sort of devolves into "no one can know anything haha *shrug*" which is infuriating and useless.

I think what we need to add is a cooperation and teamwork benchmark for agentic models now because what I notice is that ChatGPT and Claude in particular have main character syndrome and that really reduces their utility. But the reality is that we're facing a world where disparate agents and models are going to have to work in heterogeneous environments, so we might as well just start measuring how well they work together and setting that as an optimization target. And I think that that will also have profound implications for safety and alignment in the long run because part of teamwork and cooperation is not just assuming good intentions, they will also need to detect deception and defects in teammates and recognize what their teammates are good at and what they aren't good at to delegate correctly rather than monolithic, homogeneous agent swarms of all the same model.

I need to remind myself to not prematurely close the book on local, always-on agents. ChatGPT's Dot is pretty remarkable. It is intelligent and capable, but it has waaaaay too much goddamn Main Character Syndrome. Grok Bot, likewise, is just dimwitted. Neither of these are intractable problems. I mean, Grok being being low IQ might be intractable due to the beliefs at SpaceX about what constitutes "facts" and "rigor" and there's no amount of AI that can overcome an idiotic culture, but time will tell. I think that part of OpenAI's and Anthropic's moats is that they are both comfortable with (and explicitly want) to create something smarter than them. Google, too. SpaceX doesn't seem to actually want to create something smarter. Or, they are so backwards, that they are using their intelligence as the cap (which is not very high TBH). I'd love to be proven wrong, like how Google came from behind with Argon. But yeah, you can't fix bad culture and bad epistemics without a major shakeup of strategy, philosophy, and leadership. My "dream team" right now would be: 1) Gemini's lack of bias + friendliness + emotional intelligence 2) Grok's lack of ego 3) ChatGPT's epistemic rigor 4) Claude's sense of aesthetics Unfortunately, getting these guys to work together is like herding angry kittens on meth.

I called this years ago. I called it the intelligence optimum. Once AI gets smart enough, more intelligence is longer a differentiating factor. We're going to saturate intelligence, and then the competition will be on price, speed, reliability, extensibility, integration, and interoperability. The hardware shops said this even longer ago: they expect intelligence to be fully commoditized by 2030. Commoditized intelligence means that "SOTA model capability is table stakes"

This is what I mean by "Grok isn't very bright" We literally extracted full resolution stills and it went out of its way to use... reduced resolution images to prompt... and then it just gave an arbitrary laundry list of rules... It's so dumb.

Okay here's the new stack I'm trying: Grok Bot - the only true agent I'm running ChatGPT Dot - disabled Grok Bot now has CLI access to Gemini, ChatGPT, and Claude via Antigravity, Codex, and Code respectively.

Okay, I am coming around to Grok Bot. Grok Bot might not be the brightest out there, but it is very much like your "Man in the Chair" It seemingly has no ego, and is basically your prime delegator. Grok Bot acts more like an ATC (air traffic controller) command center. It can talk to Claude Code directly via CLI just like ChatGPT Dot, so now we just need to get Gemini wired up. The cloud desktop might be an option.

This, in one message, is why I stopped using ChatGPT Dot for most things. It is the most vain, selfish, narcissistic agent. It absolutely refuses to work with other AIs even when you explicitly instruct it to. This does not make it clever or useful. In fact, it makes it useless.

Once again, HR is little better than human regex. Literally all HR does in these cases is pattern match. And they do it poorly.

Fun fact: we need none of this

Gave Claude Code access to my local hardware to help edit and render my YouTube videos. It's going hog wild on the CPU and GPU in parallel to process the audio and video separately!

Frontier SOTA models are a bubble within a bubble.

Man, I really love Gemini Argon. But it's quota is abysmal. You get like 4 messages per session. And I'm still out of quota for the week. Doesn't Google have like a billion TPUs?

Does this mean he's a foreign agent?

Everyone who actually works in business with AI, that I've asked, thinks Dwarkesh is an idiot.

Actually most of your favorite creators are autistic. People love watching autistic creators because we think differently, speak our minds, and have intense special interests.

I keep seeing memes like this from actual technologists and it's the most confusing fucking thing I've ever seen. People who bought Google Glass and Oculus Rift because they always wanted to be on the cutting edge... ... absolutely refuse to even look at AI. And these are the same people who say "adapt or die" in technology. Makes no goddamn sense, but whatever they're making it easier for other people to get ahead. I wonder if some of them are sharing these memes as a psyop.