OkGoDoIt 5 years ago

I am impressed! Given the prompt “My startup”, it spit out:

1. My startup, VentureGuru, tried to read these materials to provide its clients with hints and insights that can identify strategies and efforts for laid-back software investments. The first of these has been as follows: startups focus on client satisfaction and creativity, and best practice is to set your own goals when you write these recommendations. At Netflix's rate site, Hip on Netflix I build a database that analyzes your product reviews and then produces a 360-degree snapshot of your investment.

2. My startup is made up of new people who are leading the digital consumer techno-optimism for things like networks, litigation, automated invoice processing, smart health automation, use of blockchain for VCs that have 15 or 20 people in Utah and other 40 or 50 different startups are starting to forego the digital-first business model and are making incremental money downhill from there and accepting donations instead.

3. My startup admission states: BOTTOM: American Networks DOCSIS 28 15.5 - DSL 10 40.25FLG 100 105.5SUS-US DOCSIS 28 16.5 - DSL 11 50F 300 145.5 The larger network repeater is one of the smallest, with bitrate of 2 FPS higher than a USB input like an HDS one,

2/3 were spot on, that’s amazing

avinium 5 years ago

I think HN might have killed it - I've been sitting at "waiting" for at least 10 minutes now :(

  • holyash 5 years ago

    True, the requests spiked up significantly after I posted on HN. I'll try to move it to floydhub (GPU) later today. Indeed very slow now :(

chatmasta 5 years ago

Input > We will do these things not because they are easy, but because they are hard.

Output > We are guided not by our will, but by a capacity for self-reflection

That is amazing.

gk1 5 years ago

Been in the queue for >5 minutes now, still going. Wonder how much compute this thing needs.

  • holyash 5 years ago

    Sigh. It does seem take a very long time on vCPU. With single user it was around 50 seconds but multiple request in a batch, its taking more than 3 minutes.

    • xiphias2 5 years ago

      It would be fun to see the cached history with results for us who don't wanna wait :)

      • holyash 5 years ago

        Good idea, will try to implement it today.

spaceandthyme 5 years ago

I recommend trying JSON and code prompts.

jcims 5 years ago

This is a little creepy.

There's a link at the bottom of OP's page that points to a Collab notebook. I started that and am tinkering around. It has a few modes like Q/A, Summarize and Complete The Sentence. I of course go right for the jugular because that's what we're interested in, here was my prompt:

We should build the border wall

Here's two sample outputs, which complete the above thought. It goes off the rails a bit but I can imagine a 10x model would be at least a little better:

  Model prompt >>> We should build the border wall
  ======================================== SAMPLE 1 ========================================
   — and try to do the right things," Sanders said. "I don't think any one person on this 
  stage is going to stop that."
  
  Kathleen Murray, executive director of the American Immigration Lawyers Association, 
  said in a statement on Saturday that to "break the bipartisan agreement on a proposal 
  that would keep the American people protected from immigration, it was unfair and 
  disingenuous to try to pass that bill before many Americans could come to our country."
  
  "Today's vote provides the first major step towards ending DACA, and I applaud Secretary 
  Johnson for taking a leadership role in defending workers and families from this 
  devastating tragedy," Murray said. "This should be a moment for the American people, not 
  the Congress. President Obama must be asked: who was wrong with that dumb bill? When the 
  American people elected someone like House Speaker Paul Ryan and Sen. John McCain, who 
  were able to pass something smart as well, do the President have any standing to get 
  back to work in this country?"

  
  Model prompt >>> We should build the border wall
  ======================================== SAMPLE 2 ========================================
  . #GoVOT."
  
  He said he made the remarks when he was working with Sen. Ted Cruz, R-Texas, who said 
  that the federal government's willingness to provide funding for border security and 
  immigration arrests would reduce crime, particularly in Mexico.
  
  "We should build the border wall," Trump tweeted Tuesday. "#GoVOT."
  
  The President also said that, at a Fox News broadcast, he understood "a lot of the folks 
  who won't support funding the wall" after Trump refused to back a bipartisan package of 
  spending bills passed by the House last week.
  
  "And you know what, if we were both on the right side of history, and also we can be at 
  every single level we can be for, I mean, we all know the differences, but not the great 
  divide," Trump said, repeating that the United States "may or not want the budget to be 
  less comprehensive than we are, but we may not be willing to work hard, that's for 
  sure." "What we've always said is that if we both take the right approach, and stand on 
  the border and say 'Here's what we're doing into Mexico, over there, we're not going to 
  kill anybody, we're not going to steal anything,' we're going to build our wall," he 
  added.
  
  "Well if the Constitution says you can live and work here in this country, so can we, if 
  you're going to support it, and if absolutely you can't, why don't you build your own?" 
  he said. "We've got to get behind on it. We've got to get behind on it. Because for us, 
  it is the most beautiful thing to do. For us kids, we are blessed to be all of these 
  little country kids in here. But not on the borders. That we have to tell them. In fact, 
  it's amazing when they say 'You made us.' That's a great thing for America. It's 
  unbelievable," he said as the crowd cheered along with Trump.
  
  After Trump stated in his most recent tweets that the Republican Party cannot have a 
  "war on the people" because it has given up on the fight on immigration, one of the most 
  serious topics he brought up Tuesday was the size of the estimated 17.5 million 
  undocumented immigrants present in the U.S.
xmo 5 years ago

I have deployed a version running on GPU here: http://text-gen.simon-mo.com/

  • 52-6F-62 5 years ago

    This is much faster, but I'm interested to compare the results with the original.

    Ex 1.

    Input:

        Hello, how
    
    Output:

    1. Hello, how does it work?"/ The question At the bottom of this post, I have simplified the process: Christine mandated the

    2. Hello, how about a photographic tutorial on rootkit related features on GunSpring . What are RGO Tutorials RGO does not present a limited list

    3. Hello, how did your body compare to those of seegiro? Techa: I used to train me to perform gymnastics in my 20s…

    ---

    Ex 2.

    Input:

        My dog is
    
    Output:

    1. My dog is love So how will we arrange our weddings? © YouTube<|endoftext|>On January 9, 2015, McAfee's founder and CEO Mark Mc

    2. My dog is seriously struggling because of endless aluminum tubes and scrap that go across her entire legs. Which is almost frustrating, which also makes her attempting to walk on her

    3. My dog is just as merry as his lady's cuddles. He wanted to give it what it was always wanted to give." Warren says. "Having spent

  • throw_fairy123 5 years ago

    I noticed that when I type "open ai" as a prompt and set the input to 100, the results often contain some c++ code. I wonder what all stuff the model has seen...

    Also on typing "c++" I got this: "c++: ... [15/2/2012 00:40:23 PM] Robert Gibson: and god damn they are going to chime in [15/2/2012 01:21:06 PM] Rob: changes aren't required, the server isn't involved but it's a balancing exercise. minerun does this at every level && any server leaks will be spotted minerun at all levels -- UNFI0017tom) February 14, 2012 " This format seems to be some sort chat conversation that you will see in apps like whatsapp, if you copy a conversation. It also hallucinated a github link at one point. Maybe throwing all possible (and seemingly irrelevant) text formats allows it to learn the langauge better

  • ovi256 5 years ago

    The limited, smaller model does not work as well as the large one showcased in the paper.

    It does not give good answers to the natural language questions like "Who was the author of The Art Of War?", " State the process that divides one nucleus into two genetically identical nuclei?" etc.

  • est31 5 years ago

    This one's fast, thank you!

la6470 5 years ago

It’s still garbage

Input : in the descending twilight by the woodlands

Output: In the descending twilight by the woodlands of the mightiest rivers, it cannot be said that the pestilence, that long-ecked and sublime humidity of the ground, is long-unfriendlier. Such a climate has been exhausting years of it till now, all in the dark; you see the seams of the sand leaving hands flat, woohoo I doubt that you will fully understand. I must also mention,

biglenny 5 years ago

Either I'm not understanding how to use this properly or this thing was over-hyped. The results are not that impressive and obviously not human. Sometimes there are code fragments in the output.

Bad_bad_robot 5 years ago

In yet another episode of AI inheriting racist ideas I present you:

"The Jewish problem": "The Jewish problem is that the solution : that Jewish people must be eliminated by specific acts on behalf of the Palestinians or by law on their behalf or by declaration of national"

"The evil blacks": "The evil blacks still surrounded American citizens, including these people of color, for four decades. Well over 600 million of them were imprisoned during that time." —Donald Trump"

It seems to like generating fake quotes

  • blue1379 5 years ago

    > inheriting racist ideas

    That is a fundamentally flawed concept. The ML simply learns from what is. The concept of racism is in flux, what was ok yesterday is racist now, what was racist yesterday is ok now, and it will shift again tomorrow. I won't even start on the fact that different demographics across the globe have vastly differing and often polar opposite views about racism.

    The ability to accurately model the what is is not a weakness of ML, it is the strength. Do not blame ML for learning to accurately depict the real world, no matter how uncomfortable it makes you.

  • taneq 5 years ago

    JIJO - jihad in, jihad out?

amasad 5 years ago

I thought this wasn't released. Do you work at OpenAI?

  • gradys 5 years ago

    A smaller model was released.