- jaypeters.dev
- Posts
- Vectorizing, Operating, Deepseeking
The Weekly Variable
A light week after travel and fighting a cold, but a pretty big week in AI.
Topics for this week:
App Distribution
I was really excited about the convenience of Expo’s Application Services last week and hasn’t disappointed me yet.
Their remote build is surprisingly convenient.
I wanted to create a new Android build but needed to go somewhere so I kicked off the command, packed up my laptop, and let the cloud do the building for me.
By the time I got where I needed to go, the app build was done and I could download the APK straight to my phone.
Kind of amazing.
It seemed too good to be true.
I was expecting the service to just deny my build command at some point, but it hadn’t so far.
Digging around their website, I found they allow 30 builds per month in the free tier, which seems really generous, and after that there’s an On-demand, $99 per month for business, or $999 for enterprise.
30 builds per month is plenty at this point, as long as I don’t try to make a new build for every little change, or get stuck on something I don’t know how to fix and it takes 30 attempts to fix, which could happen.
The Android version of my current app is not cooperating as well as the iOS one, so I could see more test builds in the future.
But we’re on track to have potentially both app versions available for download in the near future.
More to come on that so stay tuned.
Vectorizing
I was working on the foundation for an AI Automation Agency the last few weeks, which unfortunately has slowed to a crawl this week with travel and feeling under the weather, but I still think about it constantly.
I have a list of automations to build out and productize, but I also can’t help but think of other services in the AI world that could be really helpful for me, and most likely others.
Consuming tons of YouTube content, I began to wonder if I could save all the information in the videos and reference it somehow.
Rather than trying to recall which video I remembered hearing a specific piece of advice and jumping through timestamps to find it, I could have a chat interface that could casually talk about the answer in full context.
All I need to do is download some transcripts and give an AI access to them.
This could easily be done in NotebookLM for free.
But, if I wanted to a reference a bunch of videos, say 50 or more since NotebookLM cuts off at 50 sources per Notebook (although you could squeeze multiple transcripts into one “source” and really push the limits), the transcripts could get a little long and the AI could become less accurate, so it would then make sense to put all the transcripts into a vector database instead, which should make it easier for the AI to search for relevant topics and data from the video.
Vectorizing surprisingly isn’t that difficult to do, OpenAI has embedding models for that purpose.
A little data prep (which can also be done with AI) and I could reference an entire video catalogue, assuming it’s more audio focused like a podcast.
Throw the results into something like Milvus, and I’m all good to go.
Thinking about it more, I was debating if that process could be a good candidate for a service to offer as a business.
The client provides the data source, and they get a vectorized data source back.
I’ve been putting it off, trying to focus on direct value versus experimental value like this project, but I may give this idea a shot soon.
In the long run, it could be really helpful to have a chatbot will all the specific knowledge of an expert from YouTube.
And once I get that working for myself, it could be valuable to set the same thing for others.
Let me know if you run into anyone that’s in desperate need of vectorized data or any other AI automations.
Operator
Another surprise drop from OpenAI this week.
They released “operator”, their next big push into a fully autonomous agent.
With operator, the AI now has full access to a web browser so that it can navigate to different websites and complete tasks on it’s own, like booking a trip or ordering groceries.
Anthropic released a similar feature with computer use back in October, so maybe it’s not surprising that OpenAI is also adding a similar feature.
I saw an independent project back in October that accomplished the same functionality as well.
A Chrome extension that would give an AI control over the browser to accomplish whatever task you gave it.
These might not seem like the biggest breakthrough at first.
APIs have allowed AI to access basically whatever you want, but being API driven creates a huge technical knowledge barrier.
You have to have some understanding of what you’re doing, even with No Code platforms like Make.com when connecting an AI to other APIs.
But if you can just tell the AI what to do with words and it can go do that thing in the browser, and eventually your computer, suddenly things are much easier and more accessible to everyone.
It also starts to paint the picture of not needing to physically interact with your computer or phone at all.
You tell it what you need and it handles it.
It’ll ask you to confirm details or for help if it gets stuck on something, but otherwise you give it direction and it goes and accomplishes the task.
Another feature that will accelerate the presence and capabilities of AI.
The operator announcement caught me off guard, so I haven’t spent much time with it yet.
When I briefly tested it, it was able to browse to this newsletter and give me a brief summary of what I typically talk about.
When I tried to have it check my YouTube history for relevant videos, it was immediately blocked.
YouTube isn’t available through the service yet, so they must be slowing opening up what’s allowed and what’s not.
I’ll have to play with it more this weekend, but it’s an interesting step forward.
AI is not slowing down anytime soon.
The Stargate Project
Speaking of AI not slowing down, a really huge announcement this week is the Stargate Project making sure AI doesn’t have a chance to slow down.
$500 billion dollars pledged to facilitate building OpenAI’s infrastructure over the next 4 years.
Ultimately I think this is a good thing.
Clearly I’m a bit biased toward AI, but I really do think AI will help with massive advances in technology, science, and medicine, and more powerful data centers will help make sure that happens.
I tend to focus on technology the most because that’s my background, but there’s some truly incredible things happening on the medical side as well, and projects like AlphaFold may hugely impact people’s lives in the next few years.
I’m excited to see what happens with AI and this could be a big step forward.
DeepSeek R1
At the same time that OpenAI announced massive funding to fast-track AI infrastructure scaling, China released an open source model called DeepSeek R1 that, in many benchmarks, is on par with o1, OpenAI’s current leading model.
You can try it out yourself on the hosted version or find it on huggingface, or you could use something like ollama or LMStudio to run it on your own computer.
It might be tough to get full o1-mini level performances though, because that requires 404 GB of RAM which might be hard to come by without paying for a massive cloud.
The 7b or 8b versions only require 5GB which is much more realistic.
May even be able to squeeze the 14b onto a computer with 9 GB of ram to spare.
This could be a real game changer, having an o1-level model that’s completely open source and customizable, available for commercial use, and not limited to OpenAI’s api hosting restrictions and costs.
Looking forward to trying it out myself but let me know if you check it out!
And that’s it for this week! Pretty quiet week for me but not for AI!
Those are the links that stuck with me throughout the week and a glimpse into what I personally worked on.
If you want to start a newsletter like this on beehiiv and support me in the process, here’s my referral link: https://www.beehiiv.com/?via=jay-peters. Otherwise, let me know what you think at @jaypetersdotdev or email [email protected], I’d love to hear your feedback. Thanks for reading!