OpenAI hide n seek - my new reasonable understanding

  • 16 Replies
  • 747 Views
*

goaty

  • Trusty Member
  • ********
  • Replicant
  • *
  • 555
Re: OpenAI hide n seek - my new reasonable understanding
« Reply #15 on: November 23, 2019, 06:44:11 AM »
the more generalized your cells are in your net,  the more they fire therefore the less you need,  but if you ever OVER generalize, the system breaks down and will become nonfunctional.

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • *****************
  • Sentinel
  • *
  • 3945
  • First it wiggles, then it is rewarded.
Re: OpenAI hide n seek - my new reasonable understanding
« Reply #16 on: November 24, 2019, 04:38:42 AM »
https://openai.com/progress/

Open AI's amazing releases have maybe a pattern, they release one every so often on their website:

OCTOBER 15, 2019
SEPTEMBER 17, 2019
APRIL 23, 2019
APRIL 25, 2019
FEBRUARY 14, 2019
DECEMBER 14, 2018
JULY 30, 2018
JUNE 11, 2018

aha, first pattern, the visual image competition transformers and MuseNet transformer for music where released so close in time. 2 days apart.

1 5 2 2 0 5 1

There's the months between each release! 1 months.....release....5 months......release.. . . . .

Looks like it usually 2.5 months (75). Which was the number of days they released each stage of GPT-2 for the 4 model sizes!!!!! COOL!

And since their releases have been so amazing lately, its very very likely the next will be as interesting, if not greater of course!

I still don't know why the below one wasn't in their main ones, this is big achievement, i added its date above therefore:
https://openai.com/blog/sparse-transformer/

Of course korr will immediately deem this more marketing campaign business :) The recent Safety Gym was another. But maybe they are just proactive.
Emergent

 


Users Online

16 Guests, 1 User
Users active in past 15 minutes:
Kooxpi
[Roomba]

Most Online Today: 22. Most Online Ever: 340 (March 26, 2019, 09:47:57 PM)

Articles