AI Week Feb 19th: Time Bear and chocolate cake with snails, plus AI in war
Happy Monday end-of-the-day! It was a holiday here in Ontario, so this week's AI Week is going to be brief. It would be even briefer, but a ton of interesting things happened this week.
But first, a neural-network valentine
Valentine's Day was last week. Janelle Shae trained a neural network to generate some candy hearts. "Sweat Poo" is my favourite, but Janelle wasn't satisfied and kept training until it got stuck in a bit of a bear loop.
More candy hearts by neural network - AI WeirdnessCommentShareCommentShare
Last week I trained a neural network to generate new messages for candy hearts. They were, well, unique. My training data was really small, though - I could only find about 360 existing messages total. So, when I decided to generate some more messages, I decided to add my favorite neural network-generated messages to the original dataset, increasing the total to almost 500. This is still really small for a neural network, but a noticeable upgrade.
(If TIME BEAR and BEAR BEAR aren't your thing, there's always this old-school candy heart generator.)
Headlines
Some big AI news this week. First, Google announced that its flagship LLM, Gemini 1.5 Pro, has a "context window" of a million tokens.... practically, it can keep track of a conversation roughly 700,000 words long or process an 11-hour video. Only a limited group has access to this expanded context window; everyone else has to make do with a mere 128K tokens, same as OpenAI's GPT-4.
Second, OpenAI showed off AI-generated video clips from its text-to-video tool Sora, which are both incredibly impressive and raise a ton of questions. (The clips don't show the text prompts, and some look to me like short clips that they've edited together, sooooo we're kind of taking their word for it.) Sora was trained kind of the way an LLM was trained, but instead of using bits of text as tokens, Sora's training used bits of screen and trained on how they were associated with adjacent bits in screen-space and time.
In addition to the very impressive demo, OpenAI made some big claims about how this video generation is the first step in world simulation, but people have big doubts about that due to issues like lack of object permanence and bad physics. As Gary Marcus puts it, “It is a model of how images change over time, not a model of what entities do in the world.”
OpenAI collapses media reality with Sora, a photorealistic AI video generator | Ars Technica
Hello, cultural singularity—soon, every video you see online could be completely fake.
Third, Meta Research released a non-generative video model called V-JEPA. It's interesting that this was released at the same time as OpenAI was stealing headlines with Sora, because V-JEPA has a conceptually completely different approach: it's a self-training model that watches videos and gets better at predicting what will happen next, thereby learning, in some sense, about the visual world.
Small models news
Okay, this first one is for people who are also big SQL geeks like me. (Hi!) A small model called sqlcoder-7b-2 is now better than GPT4 at generating correct SQL queries, according to its devs. (Still only 90% correct, so... we aren't obsolete yet?)
There was also some buzz around a small model that just showed up on Chatbot Arena, mistral-next, with some people saying it was better than ChatGPT. I asked it to write me a recipe for chocolate cake with snails, and it refused:
I apologize, but I am not sure what you mean by "more snails" in the context of a chocolate cake recipe. Snails are not typically an ingredient in chocolate cake. Could you please clarify what you are looking for?
(A different model told me to decorate the cake with chocolate snails. I think that's better. A third one, however, just replaced the flour with "1 1/2 cups (360 ml) slimy slug trails", which wouldn't make a very good cake.)
Relatedly, go try Chatbot Arena, it's super fun!
On-device models news: AI on your PC
/cdn.vox-cdn.com/uploads/chorus_asset/file/25283485/Chat_with_RTX_Feb_13.jpg)
Nvidia’s Chat with RTX is a promising AI chatbot that runs locally on your PC - The Verge
Chat with RTX is a very early demo app.
Et tu, Reddit?
/cdn.vox-cdn.com/uploads/chorus_asset/file/24731145/WJoel_STK024_01.jpg)
Reddit has a new AI training deal to sell user content - The Verge
Reddit’s reported AI deal comes as it prepares to go public.
Longread: AI in the Ukrainian war
“Ukraine is a living laboratory in which some of these AI-enabled systems can reach maturity through live experiments and constant, quick reiteration” - Jorritt Kaminga
