Building the next social media platform
Acworth, GA
- Tweets 987
- Following 281
- Followers 108
- Likes 434
I hate it when interesting posts are AI generated and I have to make the executive decision to either silently boycott the slop and remain ignorant or eat from the trough like a good piggy.
1
16
Guys how is Delta gonna post 50% returns annually if they don't cost double per seat of the budget airlines, pinch every penny, and charge what the market demands for unusable internet?
Won't somebody think of the shareholders?
1
26
I don't normally post about politics, but its very...curious...to see AI doomers being courted by "progressive" candidates, while the "conservative" White House is aggressively pushing AI progress.
I wonder if the parties are about to flip again like Whigs vs Republicans.
I just want my damn hoverboard.
1
15
Jury is still out on 6.1 Sol quality, but at least for now: it consumes less than 0.5%/hr on ultra, meaning you could run at least one agent continuously for the entire week on a $200 max plan and never run out.
GIANT caveat that usage changes often and without notice.
1
19
Plumbing requires a human because it was designed for humans. The shape of factories, robots, and the jobs we need them to do will all change and in a generation people won't remember what it was like having to wait for a human to fix these problems.
1
21
Has anybody seen or heard an explanation for why recent models get worse at higher effort levels? Seems like it's across the board. Previous generations didn't have this regression.
Are we finally hitting the wall?
39
Antigravity CLI not accepting steering input at all is probably the worst UX of any AI app right now.
And then when stopping an agent to steer, it occasionally gets desync'd and swallows all inputs and outputs beyond that point unless you close the TUI and resume.
Plz fix.
65
A few thoughts about the latest Opus model:
I've had several long sessions going concurrently today. Up to now, Opus only dispatched subagents for reviews. Just a second ago, we reviewed several outstanding items at once. Before I could finish typing it, Opus launched 4 subagents to tackle them.
Every turn very consistently follows the house rules of conclusions up front, then what was done, then what's needed from me. Except for short answers where it's obviously not useful to do so. Previous Claude models have been inconsistent about rule following, while OpenAI models follow rules pathologically.
If you've been following along, I've posted several times recently about Gemini's improved performance. It often found problems even with Fable and Astra. Twice today (and for the first time EVER) I've had Opus nail a build with zero findings following an adversarial review from a different model family.
The good sense of this model is incredible. It isn't perfect, but it's the first model that feels like working with a peer instead of just a smart computer. From what I've seen this model is at or near intelligent human level in terms of taste and common sense.
Really impressed.
2
3
63
Heard someone say "about as useful as a chocolate teapot in the Sahara" today. I was pretty sure I heard right but its so absurd I had to google it. Turns out it's a real thing!
This is not useful information. But it IS one of the more interesting phrases I've ever heard.
1
31
OpenAI is crazy. My social media app uses Luna to score images and video and was already getting ~1200 images for a dollar. Half that is basically free.
This is the entire story right here. Sol cheaper than Gemini 3.8 flash. Its happening, people!
1
39
GPT6 models out for over 3 minutes and still not updated on @ArtificialAnlys . SHAME!
GPT-6 Sol and Luna are out. Not only are they a very significant improvement across the board, but also in writing and general "you know when you try it" quality.
We are also permanently reducing the API price by 50% making both of them viable for a ton of new usecases and making your usage go further too, even on the subscriptions.
And one more thing. We are loading a banked reset into all accounts of our Plus, Pro and Business users. Let's go!
openai.com/index/introducing…
1
32
Attention feels like it's about to be solved.
Along with the usual gated delta net, this recall head learns how to get full attention down to 11% of global calls with effectively no decrease in recall accuracy.
Also an interesting side note that as we find breakthroughs, it's looking more and more like normal computer memory hierarchy and less like a typical transformer.
arxiv.org/abs/2609.20734
34