Amazing work there @jrmyphlmn . Very recently I was thinking about how to preserve photographs I have collected over last 10 years in my external drive and instagram and unchecked SSD cards. Time to weave them all together. Bonus for me to be the 100th github starer on your repo
> With RSS, you subscribe directly to websites, blogs, or news outlets, meaning there is no middleman algorithm deciding what you see.
This enters a failure mode very soon, especially because most people using RSS-like technologies would typically subscribe to more sources than they can typically read through. Like it or not, _the algorithm_ does serve the purpose in prioritizing and discovery. The trouble, IMO, is with the objectives for these recommendation and ranking algorithms.
A middleman/aggregator who is paid by subscribers would be incentivized for the users, a marketplace-like aggregator would always have trade-offs.
Algorithms other than FIFO are fine when they serve you. Way back when I had a mail reader (Gnus) that used a Bayesian classifier to predict which emails I might especially want to read, based on past reading experiences. That was nifty! An RSS reader could do the same, on my own machine, based on my own preferences and not some marketer’s. I’d like that an awful lot.
You sort of can with very little work. When I used RSS more I had a "primary" folder and a number of secondary folders. I always looked at the primary; I'd dip into various secondaries when I had the time.
I do sort of agree with the general premise. The sort of social media that sort of replaced RSS is largely dead.
This is already a problem with things like Mastodon - as soon as you subscribe to some more "spammy" accounts such as news outlets, all the other content is drowned out.
So yes, having kind of re-ranking _algorithm_ can be a good thing, whether we like it or not.
Semi-Popular youtube channels regularly get offers from someone who wants to buy their channel. There are people and companies that put up good/useful content for a while to get subscribers and then shift focus. There have been several cases where someone has lost control of their system/password because of a "hack". Likely there are more that I'm not aware of.
The solution I've found, whether with RSS or other feed-based platforms (e.g., Mastodon / the Fediverse), is 1) organise feeds (by topic using RSS, by interest generally); 2) to ruthlessly prune feeds particularly in my high-interest list/category/tag; and 3) park any voluble feeds into their own "voluble / noise" group. They can drown out each other, but not lower-volume, higher-quality feeds.
Interest level works far better than category for social-media feeds, if only because few people (as opposed to organisations) tend to stick to a given topic. On Google+, one feature I used for my own outbound content was its own classification system, such that my tech posts went to a tech channel, science to science, news/current events, etc. to their own. Those following me could choose which of those they were interested in or not.
Yes, but this isn't a solution for most people. Most people don't want to do that active gardening. It's like when Google Plus excepted everyone to put their contacts in various circles and keep it up to date.
Funny thing was that I'd developed this strategy on G+. It took me years to arrive at it.
And no, G+ "Circles" were hardly straightforward or convenient to use.
But the notion of restricting your highest priority follow list to a small set (10 -- 50 profiles or feeds, and less is definitely more here) is key.
Over time that list will likely grow, but more because many of the feeds have fallen silent or infrequent. Pruning the departed actually takes some work, and is something I'll do maybe once a year or so.
The way I'd arrived at this though was that I'd gotten desperately sick of G+ (and Google) at one point, and having initially followed many, many profiles with abandon, I pruned off virtually all of them, leaving a small core I particularly cared about. Ironically, as I was trying to make myself less dependent on the network, not more, my stream quality improved immensely. Virtually all of the annoying bullshit, even if only vaguely annoying, vanished. The people I was left with largely knew me and interacted with me regularly, and had things to say I found interesting.
G+ is gone, but I've carried through that strategy to Mastodon (still relatively active, though I've taken a break much of this year) and Diaspora* (dying its slow death, but something I'll still check into a few times a year). A small but interesting curation still proves quite compelling. A key realisation was that the voluble streams which do occasionally produce an interesting insight will almost always have those forwarded by others I do follow directly, allowing me to rely on them (or their own upstreams) for curation.
It's also given me insight into mechanics from the age of print newspapers and magazines: when a local region had its own publication (as with newspapers), syndication or curation would gather content from elsewhere, and major stories tended to get carried locally. It might seem that distant publications produced exemplary content, but in truth what I'd read was creamed off the very top, and digging further into such a source often proved disappointing. I keep that in mind with current social / algorithmic / stream-based media. Economics of print publication mean that that former behaviour is largely lost to us now, but high-quality periodic publications (The Economist, Atlantic, Foreign Policy, and the like) can remain worth picking up and reading even now, should you happen across a physical storefront actually carrying them. Might even choose to subscribe should the desire be strong enough.
To be clear, G+ Circle Management was a tedious, largely unrewarding, PITA, both in general and in the specific mechanisms provided through the G+ interface. Though pruning was quite often quite rewarding.
It's probably a good strategy, but Hacker News is a bubble and most people aren't going to do this absent a very hard limit in the app itself (like Path did).
App-based limits/nudges are all but certainly the only way to see widespread adoption. But the tactic arises naturally out of both attention scarcity and the tendency for high-salience (or high-appeal) messages to be widely distributed regardless.
More problematic is when you're searching for needles in haystacks / nuggets of gold: sparse signal in high-noise environments.
Didn't Bluesky solve this problem already by allowing anyone to publish their own algorithms?
I feel like user generated sorting algorithms would be a great fit for RSS. Power users would get an ability to tweak their feeds to their liking, while other users would have a lot to choose from
I'm now building an RSS reader that is specifically designed around the algorithm that learns what sources you like the most. It also slightly adjusts the rankings for high/low frequency feed so subscribing to The Verge won't lead to you skipping updates from some personal blogs.
And I now use it far more than I ever used Reeder.
Can't express it more clearly than this. Data structures are just one part of the story not the only spot where the rubber meets the road IMO too. But going back to top of the thread, for new projects it is indeed steps 2 and 3 that consume most time not step 3
Coool. I remember when the OG pebble launched but I couldn't get one for myself (it wasn't available in my region and my pocket money didn't allow for it either ;) ). Looking forward to this #bitesNailsFuriously
> Unfortunately if you naively quantize all layers to 1.58bit, you will get infinite repetitions in seed 3407: “Colours with dark Colours with dark Colours with dark Colours with dark Colours with dark” or in seed 3408: “Set up the Pygame's Pygame display with a Pygame's Pygame's Pygame's Pygame's Pygame's Pygame's Pygame's Pygame's Pygame's”.
This is really interesting insight (although other works cover this as well). I am particularly amused by the process by which the authors of this blog post arrived at these particular seeds. Good work nonetheless!
would be great to have dynamic quants of V3-non-R1 version, as for some tasks it is good enough. Also would be very interesting to see degradation with dynamic quants on small/medium size MoEs, such as older Deepseek models, Mixtrals, IBM tiny Granite MoE. Would be fun if Granite 1b MoE will still be functioning at 1.58bit.
Oh yes one could provide a repetition penalty for example - the issue is it's not just repetition that's the issue. I find it rather forgets what it already saw, and so hence it repeats stuff - it's probably best to backtrack, then delete the last few rows in the KV cache.
Another option is to employ min_p = 0.05 to force the model not to generate low prob tokens - it can help especially in the case when the 1.58bit model generates on average 1/8000 tokens or so an "incorrect" token (for eg `score := 0`)
You likely mean sampler, not decoder. And no, the stronger the quantization, the more the output token probabilities diverge from the non-quantized model. With a sampler you can't recover any meaningful accuracy. If you force the sampler to select tokens that won't repeat, you're just trading repetitive gibberish for non-repetitive gibberish.
> And no, the stronger the quantization, the more the output token probabilities diverge from the non-quantized model. With a sampler you can't recover any meaningful accuracy.
OF course you can't recover any accuracy, but LLM are in fact prone to this kind of repetition no matter what, this is a known failure mode that's why samplers aimed at avoiding this have been designed over the past few years.
> If you force the sampler to select tokens that won't repeat, you're just trading repetitive gibberish for non-repetitive gibberish.
But it won't necessary be gibberish! even a highly quantized R1 has still much more embedded information than a 14 or even 32B model, so I don't see why it should output more gibberish than smaller models.
Maybe I missed something but this is a round about way of doing things where an embedding + ML classifier would have done the job. We don't have to use an LLM just because it can be used IMO
Nicely summarised. Another important thing that clearly standsout (not to undermine the efforts and work gone into this) is the fact that more and more we are now seeing larger and more complex building blocks emerging (first it was embedding models then encoder decoder layers and now whole models are being duck-taped for even powerful pipelines). AI/DL ecosystem is growing on a nice trajectory.
Though I wonder if 10 years down the line folks wouldn't even care about underlying model details (no more than a current day web-developer needs to know about network packets).
PS: Not great examples, but I hope you get the idea ;)
why not fix the calculator in a way that avoids/mitigates scenarios where users get to wrong quotes and then do an A/B test? This setup seemingly tilts towards some sort of a dark pattern IMO
Because the results were probably wrong because the inputs were wrong (exagerated by over-cautious users). There is no automated way to avoid that in a calculator; only a conversation with a real person (sales, tech support) will reveal the bad inputs.
I wonder if some of that could have been automated. Have a field to indicate if you are an individual, small business, or large business, and then at least flag fields that seem unusually high (or low, don’t want to provide too-rosy estimates) for that part of the market.
reply