With the performance gains they're claiming, I wonder if they implemented the Casual Encoder-Decoder technology from DeepSeek 4.1's paper.
I could see them accomplishing it and seeing gains like this in roughly the correct timeframe, and when I heard about that development I assumed the frontiers would probably jump on it.
Doesn't mean they didn't apply something similar. They could have also come up independently with their own version, the speculation is not they copied it, rather that they have performance breakthroughs which perhaps is a result of work in same domain
Unless they already have something similar of their own, which is always possible, they'd be stupid not to. I don't suppose we'll ever know, though. It would not be a good look if after the trillions of dollars that have been thrown at US labs, investors found out that they're down to copying Chinese tech.
They might be using something like this, or they might be using some other "increased sparsity" techniques, of which there are a great many. They also might be optimizing for something else - like less RAM use for KV cache.
Alternatively, they might be cutting into their margins and dropping the price because of stiffer competition from Astra. I do think that's unlikely though.
A model that was introduced a few days ago that's LLM-based, but instead of producing text, you can ask it questions and it will return decisions. The key thing being that it responds pretty quickly and predictably.
according to the people who made Jev, it does NOT output text. it's a closed model, so we can't inspect the internals, but i would just take their word for it.
just because the API responds with JSON text, does not mean that the underlying model is generating JSON text.
> I would have to do multiple google searches to build a secure network around a replacement tool I built for 1password
I hope that isn't implying you're hosting a vibecoded password manager. Because that would be a horrible idea. If you selfhost a password manager, it's much better to use something that is reputable and tested against like Vaultwarden.
Even if you have it say, in a Tailnet that only your machines can get into, that just means someone has to compromise one of your machines.
All of this also assumes it doesn't hallucinate heavily in the process and give you instructions that are in reality utter nonsense, or create something entirely different from what you were trying to do.
If you've managed to get the equipment and resources to pull this off in the first place, someone with the biological knowledge probably is not the ceiling stopping you from the other part of the problem.
Yeah, back when this stuff was pretty new in 2024 I found a jailbreak and sent some bio/chem weapon instructions it generated to an organic chem PhD friend. They told me not only were the procedures wrong but that there were several steps that almost certainly would have lead to injury or death. I assume it's gotten better by now but, no way or desire to test.
TBF an inexperienced person following _correct_ directions that involve anything even remotely dangerous is also exceedingly likely to injure or kill themselves.
IRC is small and niche enough to be a good place for more ephemeral discussion if you have a bouncer to stay connected for you. Chat on it tends to move slowly, and the userbase trends more towards actual nerds.
I like the rationale you have behind trying to be anti-addictive with the design. I wish more places around the net had that sort of mentality. Though I am curious, why remove timestamps?
To resurrect old discussions. The hot algorithm makes old engaging posts stay a little bit longer, and with the timestamp it introduced bias that you might not want to answer a very old post. My instinct was right as users are engaging with those old posts more often now.
The irony here is that modern social algorithms actively conceal these for exactly the same reason. Social feeds want to show you a steady stream of content that you are likely to engage with. They are incentivized to frame content as literally timeless and impose a continuous, never-ending present on the user, since they don't want recency to get in the way of engagement. But the time-context in which a piece of media was produced is a fairly important detail in critically evaluating it and situating it in an actual, human conversation.
Did this piece of writing with common AI smells get posted before or after 2021? Did the author write this piece about someone they admire and aspire to before or after the murders became public knowledge? What was the expert concensus on the topic when the author wrote about it? What meaning did this slang have back then?
What technologies were commonplace? He said he needs help with something I'm an expert in -- can I help him, or was it over a decade ago?
What was happening culturally when this was written? What was the weather like? Did this collection of posts with a dour tone get posted during a global pandemic when mostly everyone was self-isolating? Did he write this post about his loved one before or after her death?
I mean hopefully I don't need to sell you on why this information is useful and actually antithetical to addictive, low nuance, engagement-driven social media. We need more media literacy, not less, and de-emphasizing critical details like this seems problematic
I do sympathize with the challenge and know that there aren't really easy answers here, and every possible change is likely to ruffle someone's feathers, but I have a feeling there is another way to achieve what you're after.
> Did this piece of writing with common AI smells get posted before or after 2021?
Textlog didn't exist before 2021, did it? (though you are probably making a general point)
> Did the author write this piece about someone they admire and aspire to before or after the murders became public knowledge?
This is a really good point.
> What was the expert consensus on the topic when the author wrote about it?
Also a really good point.
> What meaning did this slang have back then?
True, but many people just don't care (unless you are talking 10 years in the future).
I'm not going to bother writing every quote-reply, but in this spirit, I wrote in another comment that I think I'm pro not having timestamps but now maybe not. I'm liking the suggestions of having very low granularity. To throw in an idea: rough labels such as: fresh, day-old, recent, this year, last year, 2/3/4/5 years ago, years old. Maybe even whole threads could become "stale" after a large amount of time, like a year. Very much spitballing here.
RE AI writing and slang, the particularities of the examples aren't so important. My point wasn't to comprehensively catalog of every instance where we're losing something by not having timestamps on textlog specifically. The point was that time -- even relative time if not timestamps -- can be used to understand a piece of content -- often they are even critical, which you get.
I like the idea of playing with granularity. I'll also suggest: one way to make timestamps available but less of a focus is to add friction. Require an extra click into a post metadata page that loads with a full page refresh. Any friction introduced reduces a feature's usage, and even as little as just adding an extra page refresh tends to be enough to reduce its usage by an order of magnitude or more. If you wanted to be particularly sadistic you could require the user download a pdf with that data lol
A solution here that I’m considering is revealing the historical information after a post has reached past a certain age. This will bring context to the post while also reduce bias on recent posts.
I do feel reluctancy answering to a post even a few hours ago thinking my answer might not be applicable to their context anymore. But not knowing, I believe keeps the conversation going for a while. I see it actually taking place.
So, what is the right thing to do here? More engagement and lack of information, or more information and less engagement? Not an easy answer.
> I do feel reluctancy answering to a post even a few hours ago thinking my answer might not be applicable to their context anymore. But not knowing, I believe keeps the conversation going for a while. I see it actually taking place.
First time I hear about textlog today, seems interesting (and good to remove likes etc). But if there aren't other comments that already address whatever point I would want to make I dont care if it is 10 minutes old or 10 days old (now, 10 years might be different and probably should be different).
Maybe that is due to my background in forums focused on solving technical problems (troubleshooting, programming, etc), but even if the original poster helped by my input any longer, someone else might be even if that is years down the line, so answers are always useful.
Of course context is everything, and if it is someone talking about what they had for lunch, that is irrelevant 10 days later (but I also wouldn't be interested in that content to begin with, even if it was 10 seconds old).
So in conclusion, I think having a timestamp is useful, and definitely for older posts (but for newer posts I think you are overestimating the problems of having them).
Gotta say that I agree with @evnm's reply on the original post that timestamps are fundamental to microblogging and maintaining a timeline. I love the philosophy behind textlog but I do disagree with this particular aspect; seems like something the author of a note/post should have choice and control of.
You could have a feature that users could turn on in their profile to display if they are online or when they were last online, and have a setting to restrict it to certain times, so that other users can know if this person would respond to a post/needs response to a post. And then a "follow" feature, that just displays the timestamps for users that you "follow" in particular (because I imagine that if you follow someone you might not care about the age of their posts). Lastly, you can just have it as an option for any post "include timestamps," so that a user can let others know the temporal context of one (or all) of their posts. I think these 3 features together would solve all of the issues with timestamps while still retaining the engagement strategy you are looking for.
I'm pretty sure The Guardian (UK news website) have a dynamic image preview that says how many years ago the article was written for this reason - often old articles are reposted or recirculated by people trying to stir up drama or spread misinformation. Think of someone sharing a news article about vaccine side-effects from 2014 in 2021...
I wonder if this will become a new revenue stream for providers. Want to know if Claude generated some text? There's a free web form you can paste into.
Of course, you can also run your own check service if you pay some API fees, and those checks can be a lot more convenient for users since the services can check multiple sources, to whom they are paying for the privilege.
Then someone washes the text through a local model that rewords it, the markers are lost, amd they're clear again.
I can see this being real useful when they scrape the internet for content. And they don't want their own crap. Or just to see how prevalent their stuff is.
It doesn't look like the providers are going to publish their randomisation key. The google version is accessible only via the gemini prompt which only actually ran the synthID tool 1 out of 2 times I tried it the other day. So at the moment the best approach would be taking the suspicious test and using claude computer use to feed it into each provider's prompt brower window "manually"...
Occasionally my girlfriend will comment on something she saw in an ad, even buy things off them. Thing is though that ad platforms don't seem to be doing anything meaningful to stop bad actors.
So really, clicking any ad and interacting with it amounts to placing your trust in a total stranger with the only criteria being they paid a little bit of money for you to see them. To which you will then download their software, give them your credit card, or whatever the end goal is.
It's far easier to assume anything in an advertisement is a scam and block them accordingly.
I could see them accomplishing it and seeing gains like this in roughly the correct timeframe, and when I heard about that development I assumed the frontiers would probably jump on it.
How it works: https://miraflow.ai/blog/deepseek-v4-1-flash-causal-encoder-...
reply