az226

7 ms·

91.
by az226 1y ago
That’s the irony. Doing it now is just hampering competition and making it better for the incumbents.
92.
by az226 1y ago
Mercor
93.
by az226 1y ago
This is exactly my take. Are you me? :-) Well said!
94.
by az226 1y ago
Little known secret is they paid $2.7B, not $2B. And Zuck and the FB head of M&A were talking shit about John Carmack’s crazy wife, who was doing his negotiation for him. On WhatsApp no less.
95.
by az226 1y ago
Puzzling indeed. Just watch an interview with him. Nothing insightful, just trendy buzzword wordvomit.
96.
by az226 1y ago
Microsoft back in the day and today still doesn’t pay top dollar. So you can’t get top talent with 65th percentile pay.
97.
by az226 1y ago
How does one train a TTS model with an LLM backbone? Practically, how does this work?
98.
by az226 1y ago
Or they are using a speedy add-on decoder.
99.
by az226 1y ago
Even classic GPT-4 from March 2023 was quantized to 4.5 bits.
100.
by az226 1y ago
They’re afraid Cursor is going to drift away from VS Code.
101.
by az226 1y ago
The least speculative: PPUs will be converted from capped profit to unlimited profit equity shares at the benefit of PPU holders and at the expense of OpenAI the nonprofit. This is why they are doing it.
102.
by az226 1y ago
You missed the part where OpenAI the nonprofit gives away the value that’s between capped profit PPUs and unlimited profit equity shares, enriching current PPUs at the expense of the nonprofit. Surely, this is illegal.
103.
by az226 1y ago
Yes and no. It sounds like the capped profit PPU holders will get to have their units convert 1:1 with unlimited profit equity shares, which are obviously way more valuable. So the nonprofit loses insanely in this move and all current inves
104.
by az226 1y ago
But these rumors were said and talked about several days. But no big options trade was made before the actual day of announcement. That’s why it’s telling.
105.
by az226 2y ago
Meta will most likely compare against it when they release the upcoming Llama 4 reasoning model.
106.
by az226 2y ago
And the practical flops always end up lower. As an example a V100 has 125 according to spec, but the ideal case is more like 100 and non-ideal like 60.
107.
by az226 2y ago
That’s with sparsity. So it’s 29% down from 40%.
108.
by az226 2y ago
Can you export/download the models after training?
109.
by az226 2y ago
Is this some joke? They use Llama 2 7B? What year is it?
110.
by az226 2y ago
Classic, Oracle denying breach despite clear evidence.
111.
by az226 2y ago
Which puzzles had the lowest solve rate? I did the first 10 and felt all easy (mentally solve it in 10-20 seconds for easier ones and 30-60 seconds for harder ones), I’d like to try the most difficult ones.
112.
by az226 2y ago
Did any single individual solve all problems? How many such individuals were there?
113.
by az226 2y ago
He wasn’t bound by the same clearance/confidentiality.
114.
by az226 2y ago
Nat Friedman circa 2048: we were successfully able to extract the man’s last thoughts (“oh shit, my boss is going to be so mad when he hea”)
115.
by az226 2y ago
GPT-4 was 1.3T. 221B active. 2 experts active. 16 experts total.
116.
by az226 2y ago
Really shitty. Avoid.
117.
by az226 2y ago
The bitter lesson is all about more data/compute. Doesn’t matter if it costs more because it’s not as efficient, the scale is what matters at the end of the day. Sure you’d reach farther with both scale and optimization, but raw scale
118.
by az226 2y ago
Until we see progress slowing down, I don’t see venture capital disappearing in the race to ASI and beyond.
119.
by az226 2y ago
It was.
120.
by az226 2y ago
Before 2020, it was around 7-10x, so it doesn’t surprise me it went up after.

More ›