AI - Observations

Fun With GPT-3

Last week I finally got access to GPT-3. Asking for facts First off, GPT-3 does a pretty good job when you want factual information: At least 50% of the facts here seem to be true!Asking for recipes The model doesn't do a great

AI

Don't Close Your Eyes To AGI

In the paper Concrete Problems In AI Safety [https://apoorvupreti.com/concrete-problems-in-ai-safety/], there's a great example of how AIs can evade objectives they are asked to accomplish. A cleaner robot is asked to vacuum the dirt it sees in the room. It perversely satisfies

AI

Don't Plan For The Expected Outcome

When planning for the future, we often try to prepare for the most likely outcome. But this isn't always a great idea. Sometimes you want to plan for the best possible or worst possible outcomes. You want to plan for the best possible outcomes

Dailies

DALL-E

OpenAI's newest generative model DALL-E [https://openai.com/blog/dall-e/] can generate images from descriptions of text. Like GPT-3, it's such a major improvement over anything we've had in the past that it's absolutely shocking. The most impressive example: when prompted to generate 'an

AI

The Time For AI Safety Is Now

2020 was the year I started taking the risk of super-intelligent AI seriously. I blame this on OpenAI's language model, GPT-3 which was simultaneously impressive and scary. Some did people oversell its abilities – it isn't even close to as smart as a human being,

Dailies

An AI Wrote This

Liam Porr generated this [https://adolos.substack.com/p/feeling-unproductive-maybe-you-should]article using GPT-3 and many people were fooled [https://news.ycombinator.com/item?id=23893817] into believing it was written by a real person. The text seems devoid of useful information and mostly repeats

Dailies

State of AI 2020

This is a report put together by two investors on the current state of AI. Slides here [https://docs.google.com/presentation/d/1ZUimafgXCBSLsgbacd6-a-dqO7yLyzIl1ZJbiCBUUT4/edit#slide=id.g949ca8ac58_0_28] . Highlights: * Jobs most at risk of automation (slide [https://docs.google.com/presentation/d/

computer-science

Paper: Concrete Problems In AI Safety

Imagine you've created the world's first fully intelligent cleaning robot. You're getting ready for the very first real world test run. Attempt 1 You put the robot in your living room, turn it on, and command it to clean the room for you. You

Dailies

Sentient AI Is Coming / GPT-3

Scott Alexander covers [https://slatestarcodex.com/2020/06/10/the-obligatory-gpt-3-post/]OpenAI's latest NLP neural net, GPT-3. It's trained with a massive 175 billion parameters, and sometimes it's shocking how human-like the language it generates is: > PROMPT: UNITED METHODISTS AGREE TO HISTORIC SPLIT – THOSE WHO

Subscribe to Observations