Yeah, it's been awesome! I'm so excited about tool calls and function use, the possibilities are huge. I ran it over 1494 videos that range in length from a few seconds to over 3 hours. Total duration 260 hours and a total size of 3795 GB. I don't know exactly how long it took to run, as I found some bugs I needed to fix when processing mkv files, but it was probably around 24 hours in total. That wasn't all LLM requests, but also the local Whisper transcription and frame extraction / analysis. I used gemini-3.1-flash-lite-preview for the content analysis and tagging. Analysis cost $9.22 and Tagging cost $2.72 and the results seem great (for comparison, I did 885 videos a few weeks ago with Sonnet and it cost $130 in total). Gemini seems much less verbose than Sonnet, even with the same prompt, so the descriptions are much shorter, but they seem very good. The tagging is great. Another added bonus has been that with the larger screenshots being sent, the LLM can now read much more of the text it sees on screen. Some of my videos are top-down showing me drawing and writing, and now it picks that up, so it's all indexed and searchable. I tested a few models with the RAG Chat feature, and the best one so far is GPT4.1-Mini. Before, when asking questions about the library or a video it was around 4 cents each query, now its averaging about half a cent.
Hey, just want to thank you for this suggestion. Spent this morning swapping to open router and changing all my prompts to use tools instead of XML. Not only is Gemma and Gemini much cheaper, the output tokens from the tool call are much less too. Cost to analyse one 20 minute video with 10 snapshots went from $0.21 to $0.009, and I'm even sending full HD snapshots instead of the 960x540 ones I was sending before (to save costs). The results so far are pretty good. It looks like the larger images are giving the model more context, so in some cases making the cheaper models results better than the expensive models. I'm going to run this over a few hundred videos today and see how it goes in bulk!
I originally limited mine to 10 frames spread evenly throughout the video, but it missed a fair bit of context at the analysis step, and didn't scale with length. So now when a video is loaded the app extracts a bunch of frames for the entire video, then calculates an image histogram and compares similarity to the previous one. There's some configuration so it doesn't send too many to the LLM, but still gets a good cross-section of frames to send.
You could also just use FFmpeg as it can do scene detection too. I tested both but liked the results from the histogram analyzer more.
Yeah, markdown works well if you're going to search through it with Claude Code or something like that. I built ClipScape as an Electron app with a local SQLite database, as I wanted an interface I could search and chat in and see the relevant thumbnails.
Cheers, I'll give it a try. How are those models at returning structured results? When I was writing the prompts for the analysis step and testing with older Claude models, it would have trouble structuring the XML consistently. Sonnet 4.6 handles it really well.
This is great. I wish I had enough ram for a local model. I just spent the last few weeks writing something very similar, but I made it a local Electron app with Whisper, ffmpeg and I added semantic search and embeddings for chatting with the videos. It talks to Claude for the vision analysis, tagging and video chat. Do you only send one image for yours? I used a customised scene detection algorithm to find multiple different images per video and then send them all in one request to Claude (along with the subtitles). It's definitely the most expensive part. Using Sonnet 4.6 for the analysis and Haiku for the tagging costs about $1 for an hour of footage, I can imagine it would be slow locally.
But saying it's a confidence trick is saying it's a con. That they're trying to sell someone something that doesn't work. Th op is saying it makes then 10x more productive, so how is that a con?
Reading your comment made me think of the Roman generals returning to a triumph and someone constantly following them saying "memento mori", reminding them they are not a god. Now, instead of humility it would just be seen as a challenge.
An AMS is useful just so you can have 4 different filaments ready to go at any time. Doesn't need to be for multi material models. I have an A1 with the AMS lite and a Prusa mk3s, and manually changing materials is a chore.
I think you'd have to manufacture a culture, with rituals and habits designed to keep people focused so that the meaning of their lives was tied to the end-goal. It would make a good story :)
Yeah, the app lets you configure which whisper model to use and then downloads it on first load. Whisper blows me away too. Ive only got a 2080 and use the medium model and it's surprisingly good and relatively fast.
Maybe developers are using it in a less visible way? In the past 6 months I've used AI for a lot of different things. Some highlights:
- Built a windows desktop app that scans local folders for videos and automatically transcribes the audio, summarises the content into a structured JSON format based on screenshots and subtitles, and automatically categorises each video. I used it on my PC to scan a couple of TB of videos. Has a relatively nice interface for browsing videos and searching and stores everything locally in SQLite. Did this in C# & Avalonia - which I've never used before. AI wrote about 75% of the code (about 28k LOC now).
- Built a custom throw-away migration tool to export a customers data from one CRM to import into another. Windows app with basic interface.
- Developed an AI process for updating a webform system that uses XML to update the form structure. This one felt like magic and I initially didn't think it would work, but it only took a minute to try. Some background - years ago I built a custom webform/checklist app for a customer. They update the forms very rarely so we never built an interface for making updates but we did write 2 stored procs to update forms - one outputs the current form as XML and another takes the same XML and runs updates across multiple tables to create a new version of the form. For changes, the customer sends me a spreadsheet with all the current form questions in one column and their changes in another. It's normally just wording changes so I go through and manually update the XML and import it, but this time they had a lot of changes - removing questions, adding new ones, combining others. They had a column with the label changes and another with a description of what they wanted (i.e. "New Question", "Update label", "Combine this with q1, q2 and q3", "remove this question"). The form has about 100 questions and the XML file is about 2500 lines long and defines each form field, section layout, conditional logic, grid display, task creation based on incorrect answers etc, so it's time consuming to make a lot of little changes like this. With no expectation of it working, I took a screenshot of the spreadsheet and the exported XML file and prompted the LLM to modify the XML based on the instructions in the spreadsheet and some basic guidelines. It did it close to perfect, even fixing the spelling mistakes the customer had missed while writing their new questions.
- Along with using it on a daily basis across multiple projects.
I've seen the stat that says developers "...thought AI was making them 20% faster, but it was actually making them 19% slower". Maybe I'm hoodwinking myself somehow, but it's been transformative for me in multiple ways.
When my kids were younger I tried to to replace my swearing by saying "sugarplum fairies". It was fairly successful in becoming a natural replacement. However, the other day I kicked my toe really badly and instinctively yelled "sugarplum FUCKING fairies" and my kids (now early teen) found it extremely funny.