I've found that for the most part the articles that I want summarized are those which only fit the largest context models such as Claude. Because otherwise I can skim-read the article possibly in reader mode for legibility.
Is llama 2 a good fit considering its small context window?
Personally I use llama3.1:8b or mistral-nemo:latest which have a decent contex window (even if it is less than the commercial ones usually). I am working on a token calculator / division of the content method too but is very early
I applied (for now) a pre-filled table with a 4096 default limit. Users can also specify an upper or lower limit from the UI directly now. Added chunk and recursive summarization too.
I presume you want information of some value to you otherwise you wouldn't bother reading an article. Then you feed it to a probabilistic algorithm and so you can not have any idea what the output has to do with the input. Like https://i.imgur.com/n6hFwVv.png you can somewhat decipher what this slop wants to be but what if the summary leaves out or invents or inverts some crucial piece of info?
"Then you feed it to a probabilistic algorithm and so you can not have any idea what the output has to do with the input"
This is theoretically true, but to me at least, practically irrelevant. In all cases, for most values of the word "all", the summary does tell you what the article contains.
For me at least, the usefulness is not that the summary replaces reading the article. Instead, it's a signal telling me whether I should read it in the first place.
Nowadays a lot of websites are written in a style that goes on and on and on dancing around a topic, adding historical context all in a terrible writing style only to lengthen the text for SEO. In such cases a summary can be a good thing.
It's really not that deep. There's writing you read for its aesthetic merits and writing you read for its contents. When you want the latter but the piece is written for the former a summary fixes the mismatch.
Even if I want to read the entirety of a piece of long form writing I'll often summarize it (with Kagi key points mode) so that I know what the overall points are and can follow the writing better. Too much long form writing is written like some mystery thriller where the writer has to unpack an entire storyline before they'll state their main thesis, so it helps my reading comprehension to know what the point is going in. The personal interest stories that precede the main content always land better that way.
Why don't you try using these tools yourself so you have an understanding of them? People like to get shit summarized, its really not as deep as you are trying to make it out to be.
> You might be surprised to learn that I actually think LLMs have the potential to be not only fun but genuinely useful. “Show me some bullshit that would be typical in this context” can be a genuinely helpful question to have answered, in code and in natural language — for brainstorming, for seeing common conventions in an unfamiliar context, for having something crappy to react to.
> Alas, that does not remotely resemble how people are pitching this technology.
sometimes you don't have time to read an entirety of a large article. You want a quick summary, some people are poor at summarizing things in their head as they go and can get lost in dense text. Extensions like these really help me with headers, structure that I want to follow, quick overview and gives me an idea if I want to deep dive further.
Sometimes it's not even an article, but a video. And sometimes all you care is just a single tiny fact from that video.
Although I don't think this particular summarizer works for videos. And I don't think Ollama API supports audio ingestion for transcription. There are some summarizers that work with YouTube specifically (using automatic subtitles).
### Added
- New `model_tokens.json` file containing token limits for various Ollama models.
- Dynamic token limit updating based on selected model in options.
- Automatic loading of model-specific token limits from `model_tokens.json`.
- Chunking and recursive summary for long pages
- Better handling of markdown returns
### Changed
- Updated `manifest.json` to include `model_tokens.json` as a web accessible resource.
- Modified `options.js` to handle dynamic token limit updates:
- Added `loadModelTokens()` function to fetch model token data.
- Added `updateTokenLimit()` function to update token limit based on selected model.
- Updated `restoreOptions()` function to incorporate dynamic token limit updating.
- Added event listener for model selection changes.
### Improved
- User experience in options page with automatic token limit updates.
- Flexibility in handling different models and their respective token limits.
### Fixed
- Potential issues with incorrect token limits for different models.
Is llama 2 a good fit considering its small context window?
If you have the compute, might as well use the better model :)
The 3.2 series wasn't the kind of leap that 3.0 -> 3.1 was in terms of intelligence; it was just:
1. Meta releasing multimodal vision models for the first time (11B and 90B), and
2. Meta releasing much smaller models than the 3.1 series (1B and 3B).
I presume you want information of some value to you otherwise you wouldn't bother reading an article. Then you feed it to a probabilistic algorithm and so you can not have any idea what the output has to do with the input. Like https://i.imgur.com/n6hFwVv.png you can somewhat decipher what this slop wants to be but what if the summary leaves out or invents or inverts some crucial piece of info?
This is theoretically true, but to me at least, practically irrelevant. In all cases, for most values of the word "all", the summary does tell you what the article contains.
For me at least, the usefulness is not that the summary replaces reading the article. Instead, it's a signal telling me whether I should read it in the first place.
Also note
https://hachyderm.io/@inthehands/112006855076082650
> You might be surprised to learn that I actually think LLMs have the potential to be not only fun but genuinely useful. “Show me some bullshit that would be typical in this context” can be a genuinely helpful question to have answered, in code and in natural language — for brainstorming, for seeing common conventions in an unfamiliar context, for having something crappy to react to.
> Alas, that does not remotely resemble how people are pitching this technology.
Although I don't think this particular summarizer works for videos. And I don't think Ollama API supports audio ingestion for transcription. There are some summarizers that work with YouTube specifically (using automatic subtitles).
Now you can’t possibly get through all of them and have to decide which of those could be worth your time. And in that case, the tradeoff makes sense.
- # Changelog
## [1.1] - 2024-03-19
### Added - New `model_tokens.json` file containing token limits for various Ollama models. - Dynamic token limit updating based on selected model in options. - Automatic loading of model-specific token limits from `model_tokens.json`. - Chunking and recursive summary for long pages - Better handling of markdown returns
### Changed - Updated `manifest.json` to include `model_tokens.json` as a web accessible resource. - Modified `options.js` to handle dynamic token limit updates: - Added `loadModelTokens()` function to fetch model token data. - Added `updateTokenLimit()` function to update token limit based on selected model. - Updated `restoreOptions()` function to incorporate dynamic token limit updating. - Added event listener for model selection changes.
### Improved - User experience in options page with automatic token limit updates. - Flexibility in handling different models and their respective token limits.
### Fixed - Potential issues with incorrect token limits for different models.
It summarizes via Puter (free).