Hey folks -- I created 10 Minutes A Day because I was spending too much time 'doomreading' in comment threads, in an effort to stay up to date on what was going on in the tech world. While I love Hacker News and other news aggregators for what they can teach me, I found myself often unhappy reading through hundreds of comments that while, entertaining, ultimately made me feel bad.
The concept of 10 Minutes A Day is extensible beyond just HackerNews, but the Hacker News dataset was a great place to start as it is often a go-to for news around technology and new projects.
It was built on top of Flask, React, OpenAI GPT-4o-mini (to drive down costs), and generally uses a few different prompting techniques to make things work as intended, because GPT-4o-mini can be frustrating to instruct.
10MAD is going to be extended across different news verticals in the future (once I have the time).
We launched a preview of App Spaces (v2) today which is our attempt to make the Microsoft Azure portal much friendlier for developers who are new to Azure or new to cloud, and don't want to deal with all the cruft and intense complexity that Azure/AWS/GCP/etc provide. We know that there's so much more we can do to make these experiences better and this is our beginnings of getting there via something like App Spaces.
If you want to learn more or provide feedback, you can reach out to me-- sk dot hartle at microsoft dotcom. We want to rapidly iterate on our initial design here and get to core value within the next half a year or so. We know we have a long way to go, which is why we're releasing it to the community to help us drive and shape our roadmap.
I have to log in with Azure credentials just to see the docs? That's a non-starter, even for a guy who usually has an Azure portal open.
There's also no way to know, without providing credentials, if clicking on the buttons and providing credentials on that page is going to deploy something or show me docs. Again, that's a non-starter. I'm not deploying something without knowing what I'm deploying or how to un-deploy it.
I want to be excited, but there's exactly nothing useful I can gather from this page in its current form. Zero ability to sort out whether this will help me or not.
The FAQ section has an answer that flows off the page on my phone, and the text isn’t selectable for some reason, so I can’t even share the section I’m talking about.
Thanks for pointing that out. It’s been driving me crazy that there’s all these sensationalist articles using that as the “smoking gun”, implying that Netflix specifically wants an AI product manager to produce television shows.
As somebody who has worked to create an AI-generated show, and who is also a PM at a big tech company that is using LLMs for non-creative purposes, I can tell you that the “PM” work I do with these LLMs is vastly different than the creative work I do with them. It’s an entirely different frame and discipline.
I’d start to be concerned once we see job listings that explicitly look for creators, with technical backgrounds in generative AI. The creator/creative talent part comes first before everything else.
Yeah, Netflix pays this much not because it will replace a near minimum wage writer but because optimizing their advertising tier by 1% will pay back more than the PMs salary every 30 days.
This doesn’t help you probably, but the difference between 3.5 and 4 when giving it instructions to follow is huge. I encourage everybody to use GPT-4 when possible, the differences are night and day.
This isn’t reality. Using ChatGPT this way is fruitless, because there is a system prompt you’re fighting against. I can write a one sentence system prompt for the GPT API that specifies GPT to only spit out JSON, and it works fine. It’s a pretty funny series of image, though.
Interesting, so you are saying you have an easier time generating JSON using the single prompt models like text-davinci and text-bison, rather than the chat versions?
Sorry, I was non-specific. If you're using ChatGPT, you're basically using a "product" that OpenAI created. It has specific system prompts and prompt engineering to ensure that it stays on the rails. If you, instead, use the OpenAI API's for GPT-3.5/GPT-4, you aren't beholden to the ChatGPT "product". It's very easy to create a chatbot (not using ChatGPT, the product) that only produces JSON. It's just hard to get ChatGPT to do that same activity.
That's why everybody doing experiments in this space should either 1) be using the OpenAI Playground, or 2) using the API, and not using ChatGPT.
Ah ok, I'm primarily using the API already. One interesting thing is that the GPT-3.5 "product" is much faster, but looks to be using a different model in the request, their encoding model iirc. I wonder if they are now using embeddings to cache results to reduce load on the real models when they can?
They don't mean text-davinci and text-bison, but gpt-3.5-turbo and gpt-4 (and gpt-4-32k). Those are the models powering ChatGPT 3.5/4.
The API for them is already structured conversationally - you don't provide a single prompt to complete, you provide a sequence of prompts of different types ("system", "assistant", "user"); however they mix those together on the backend (some ChatML nonsense?), the models are fine-tuned to understand it.
That's what people mean by "API access to ChatGPT". Same models, but you get to specify system prompts, control the conversation flow, and don't have to deal with dog-slow UI or worry about your conversation getting "moderated" and being blocked with a red warning.
(The models themselves are still trained to refuse certain requests and espouse specific political views, but there isn't a supervisor looking at you, ready to step in and report you to the principal.)
I think you missed the sibling comment where myself and GP have already aligned on this.
Don't need you to explain how the APIs work... and it seems that GPT3.5 UI is doing something else, using the "text-davinci-002-render-sha" model, just look in the browser dev tools. I'm not sure the UI is using anything beyond the smallest context size for GPT4 either, give the output is cut off earlier than 3.5 and it too loses focus after enough messages in a conversation...
I’ve been a part of many organizations where the meetings are bad. I don’t think it’s that meetings are inherently bad, but rather how different companies and teams use them.
I generally only invite people to meetings who I know will have important input or need the context. The meeting should be about determining next steps for N period and provide clarity and direction to take those steps.
Any other meeting I mostly find to be a waste of time. But when you get the right people in the room, once a week, to talk about progress on a new product (for example), it can almost entirely replace documentation and is far more flexible and lightweight.
I love writing a good narrative doc or spec, but it leaves room for interpretation. Other people are also not as skilled at writing, and it leaves them without a vehicle to communicate what they want.
So, yeah, agree that meetings are the work/can be an optimal tool for achieving work, but they need to be done right.
Hello! I’m the co-founder of Mismatch Media. We’re looking for a couple of talented and driven engineers and business folks with strong startup experience to come work with us. We are pre-seed, offering founding member equity. You may have seen us recently in the news (for better or for worse!). We create always-on, generative media.
If you’re interested in working on a new type of media, inventing things that nobody has invented before, and creating business models nobody has created before, feel free to reach out: skyler at mismatchmedia.com.
We worked on this w/ a very small team for the past four years, in-between our day jobs. When started, OpenAI didn't have an API, and Stable Diffusion definitely wasn't a thing, so we had to come up with novel methods to thread cohesive content together. Most of the "creative" details e.g., laugh track, dialogue, frequency of dialogue, camera shots, and so on, are all tunable on a per scene basis.
We're in sort of a holding pattern right now -- no clear path to monetization for the project, and it hasn't garnered enough attention for us to probably get funding based on the technology backbone.
Hope you enjoy it! Labor of love. :)
(posted this in the similar thread yesterday but I’ll take any exposure I can get…!)
I couldn't figure out how to link the comment, but it was at least 9 months ago I said this, and someone was already working on it?? SUPER cool!
I feel like AI/ML and some pre-rigged models could be used to build a web-hosted sitcom. It would take a lot of technical work, and a lot of writing chops to make it watchable, but it could be done.
Look at Space Ghost: Coast to Coast, a show that was made by reusing old animations. If someone writes a script (use humans for now), then the AI/ML can stitch together the animation.
Once there's a corpus of material, start letting the AI determine plot points, and have humans vet and write it. Over time, let the AI take more and more control over the direction of the show, making sure to introduce new characters or events as needed.
I think you rightfully pointed out the "watchable" criteria. This is where we got to with 4 years of working on-and-off on it. The hard part is continuing to iterate over the structure to make it passable as a sitcom. The easy part is actually the language model stuff... we have an OpenAI integration, I just don't keep it on because it's a lot of $$$. We have lots of ideas about how to expand the show's structure, but it's mostly backend work that we simply don't have the resources to finish right now.
Luckily, our "goal" with this project was for it be nonsensical (hence the parody part), but I'd love to spin off a new show that focused on making it watchable for hours at a time. Our system is extensible enough that it wouldn't be hard to pop in new art assets and have a brand-new show in a month or two.
Saw some articles about this and see you have 10k viewers right now - any clue where this burst of attention is coming from, or did this comment do it?
I believe some generate slapped bass licks interspersed would be nice. S*****d’s signature bass licks were actually played on a synth keyboard. Ideally the Melodie’s would be even more odd and sometimes go for very long stretches, always referencing the previous luck but elaborating on it in an annoying free jazz sort of way.
We worked on this w/ a very small team for the past four years, in-between our day jobs. When started, OpenAI didn't have an API, and Stable Diffusion definitely wasn't a thing, so we had to come up with novel methods to thread cohesive content together. Most of the "creative" details e.g., laugh track, dialogue, frequency of dialogue, camera shots, and so on, are all tunable on a per scene basis.
We're in sort of a holding pattern right now -- no clear path to monetization for the project, and it hasn't garnered enough attention for us to probably get funding based on the technology backbone.
It would be cool if you could somehow integrate twitch chat into it. Like maybe the characters mention keywords from the chat, or there could be like some screen that displays chat messages somewhere and the characters react.
We actually do scrape Twitch chat—- we have onscreen gauges, meters, etc, that the chat can interact with by saying certain things that will change what happens in the show. I just don’t have it turned on right now because our viewer volume is pretty low. :) Great suggestion though re: characters interacting to viewer chat, that would be a lot of fun.
We generate the dialogue per every scene, every 2 minutes or so. We wanted to be able to massage the narrative over time, so the show is constructed in these two minute intervals.
:D, awesome! Our vision for this was actually just that — second screen content, not necessarily the primary thing you have on, so we love to hear that!
The concept of 10 Minutes A Day is extensible beyond just HackerNews, but the Hacker News dataset was a great place to start as it is often a go-to for news around technology and new projects.
It was built on top of Flask, React, OpenAI GPT-4o-mini (to drive down costs), and generally uses a few different prompting techniques to make things work as intended, because GPT-4o-mini can be frustrating to instruct.
10MAD is going to be extended across different news verticals in the future (once I have the time).