Kevin Lacker

The search isn't on hold, we just do it differently nowadays. SETI@home was great for its time, but distributing data to lots of end user machines just isn't the right architecture any more. Modern radio telescopes produce so much data, you have to process it onsite - these are pretty remote facilities, and the bandwidth out is much less than the bandwidth the radio telescope can produce. Arecibo is also a couple generations ago - the Green Bank Telescope is more powerful, and the next generation of SETI is with the arrays, the VLA in New Mexico and MeerKAT in South Africa. I wrote some other stuff about the state of SETI here: https://lacker.io/physics/2022/01/21/looking-for-aliens.html

Is it? I remember reading [1] back in the day, which showed some limitations of GPT3, but ChatGPT actually answers all of those perfectly. [1] https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html

Say "I don't know." Back when GPT-3 was all the rage on the internet, I remember Kevin Lacker's "Giving GPT-3 a Turing Test" left a real impression on me [1]. Short read, but gets to the crux of the issue. GPT-3 is a (really sophisticated) statistical model, GPT-4 will be no different. Maybe it will showcase new emergent properties of LLMs, like GPT-3's in-prompt few-shot learning did. But the fundamental constraint of a statistical model optimized to a masking task remains -- they're really bad at introspection and confidence-assessment. [1] https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html

I found the blog post that inspired the short story more interesting: https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html

Yes, if you want answers, you need to frame it as a Q&A session: https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html

Language models can take input and figure out a response. They're designed to predict words - you just set it up so they predict the response. In fact I think that's the only thing their API lets you do at the moment. https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html

Agree. It may happen soon, but GPT-3 isn’t it. One of the biggest problems is that it doesn’t have any idea when it is wrong. This is a big problem, even in the human domain, but especially with AI https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html

http://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html Q: Which colorless green ideas sleep furiously? A: Ideas that are colorless, green, and sleep furiously are the ideas of a sleep furiously. Q: Who was president of the United States in 1600? A: Queen Elizabeth I was president of the United States in 1600. Q: How many Super Bowls do the Cincinnati Bengals win in the 2030s? A: The Cincinnati Bengals win two Super Bowls in the 2030s. Its regurgitating nonsense all the way down.

This essay, which touches on the subject of whiteboard coding, has some good food for thought on the subject: http://lacker.io/tech/2017/04/05/why-you-cant-say.html.

>There is much structural regularity in a large text corpus that is descriptive of relationships in the world. Sure, there is a lot. But let's say we want to learn what apples are. So we look at occurrences of "apple" in the text corpus, and learn that apples can be eaten, they can be sweet, sometimes they are sour, red, sometimes green, and so on. Can apples spontaneously change size? Hmm, no idea, no mention of that in the text. Can they be used as a knife? Dunno. If I had an apple in New York 4 minutes ago, can someone else be eating the same apple in Hong Kong now? Dunno. Did apples exist in France two million years ago? Dunno. Can you drive to Saturn in an apple? Dunno. In short, there's no actual model of what an apple is, as an object in space, connected to other objects by various relationships. If there were, the model could figure out the answers to the questions above by inheritance. Maybe these particular questions happen to be answered correctly by PaLM. Maybe not, but the next LLM will include this comment in the training corpus. But the reason GPT-3 and other models tend to make no sense is because their output is not constrained by reality. The text in the training corpus tends to conform to reality, but when you prompt the model with questions that nobody would ever write text about, the illusion falls away: Q: How many eyes does a giraffe have? A: A giraffe has two eyes. Q: How many eyes does my foot have? A: Your foot has two eyes. Q: How many eyes…

Being the author of The Program, a fiction podcast described by fans as "a love child of Black Mirror and NPR", I've been following news around GPT-3 very closely. So I decided to make an episode depicting a fictionalized origin story of GPT-3. The twist is that I incorporated actual text generated by GPT-3 to tell the narrative (even the episode title has been generated by it!) To this end, I used four excellent articles and wonderful HN discussions that ensued as sources of inspiration and insight. Some of the comments have made it into the episode almost verbatim (of course, all have been properly credited in the show's references). I'm talking about: https://tinkeredthinking.com/index.php?id=841 - discussion: https://news.ycombinator.com/item?id=23923799 https://www.gwern.net/newsletter/2020/05#gpt-3 - discussion: https://news.ycombinator.com/item?id=23623845 https://lacker.io/ai/2020/07/06/giving-gpt-3-a-turing-test.html - discussion: https://news.ycombinator.com/item?id=23887637 https://arr.am/2020/07/25/gpt-3-uncertainty-prompts - discussion: https://news.ycombinator.com/item?id=23989215 Additionally, I partnered up with a startup called Narration Box to synthesize the computer voice of GPT-3, which I connected with back when I originally unveiled The Program on HN (you might be glad to hear the show recently passed 100,000 downloads). This community has been nothing but good to me so I thought you might be interested to check out the latest work of whimsy it…