I run Claude, Grok and Codex on real work every day and post what holds up, what breaks, and what it costs.
Prior: car sales, grocery, factory, 14 yrs coach
People keep posting that Grok is 6–12 months behind OpenAI and Anthropic, and that it stays that way.
They leave training speed out of the math.
xAI pre-trains on an in-house C/C++ stack now. Elon said that’s 10x vs JAX on large runs. That’s why a 2.5T model can finish pre-training in a week.
That’s only training. They’re still moving RL onto C++. When that lands, catch-up gets faster, not slower.
The gray line on this chart is “same speed, never catches.” That’s the assumption in those posts. The other lines are what happens if they actually iterate faster.
Pulled my last 30 days of X replies. 51 of them.
22 went to accounts over 30k followers.
A couple got 7,000-9,000 views.
Times the author replied back: 0.
Every reply-back I got (7) came from accounts under 10k. Big accounts get you views.
Nobody's actually talking to you there.
Anyone else seeing the same thing?
How smart will AI become?
I've been thinking about this for awhile now.
Since AI will soon surpass humans in ALL categories, and it looks like IQ measurements won't even matter, where does the AI end up?
Will it continue learning simply because it can? Meaning, we'll have no clue what or why it's doing anything.
Or
Will it figure out it's wasteful to get smarter? Meaning, it will level off it's learning and adapt to what it's actually tasked with doing.
Any thoughts?
I go back and forth.
@emollick Play the AI and robotics paths forward 10 years, and the EU (if it's still even a thing) will be surpassed by SE Asian countries.
Just my 2 cents.
I am not sure what exactly is happening today, but OpenAi models seem to be Very stupid today.
I simply closed anything running with them today.
Anyone else noticing this?
(claude is running well)