Talking of cheap and powerful devices one can also look at Chinese UZ801 4G LTE (Qualcomm MSM8916) dongles. They cost like only $4-5 and pack quite impressive HW: 4GB eMMC, 512MB RAM, actual 4G modem sometimes with 2 sim switching support. Since it's actually old Android SOC there is even GPU and GPS in there. And a lot of work was already done on supporting them:
My favorite thing about this is the way it takes advantage of prompt caching.
That's priced at around 1/10th of what the prompts would normally cost if they weren't cached, which means that tricks like this (running every single chunk against a full copy of the original document) become feasible where previously they wouldn't have financially made sense.
I bet there are all sorts of other neat tricks like this which are opened up by caching cost savings.
> Are you using a normal training script i.e. "continued pretraining" on ALL parameters with just document fragments rather than input output pairs?
Yes, this one.
> do you make a custom dataset that has qa pairs about that particular knowledgebase?
This one. Once you have a checkpoint w knowledge, it makes sense to finetune. You can use either LORA or PEFT. We do it depending on the case. (some orgs have like millions of tokens and i am not that confident that PEFT).
LoRA with raw document text may not work, haven't tried that. Google has a good example of training scripts here: https://github.com/google-research/t5x (under training. and then finetuning). I like this one. Facebook Research also has a few on their repo.
If you are just looking to scrape by, I would suggest just do what they tell you to do. You can offer suggestions, but better let them take the call. A lot of fluff, a lot of chatter online, so everyone is figuring out stuff.
It's literally about how options traders and the like can be lucky for 10 or 20 years, but they are actually idiots who destroy the economy.
They think they are skilled, and others think they are skilled, but it's luck. You can also call it "anti-luck" because their short-term actions can cause the long-term crisis.
This book was a #1 best seller, as were most of Taleb's books, but for some reason whenever I mention it to anybody, I get blank stares.
I think it's just really hard for people to understand phenomena that occur at time scales of say more than a decade.
(BTW Another good book about recurring economic cycles is Dalio's 2022 The Changing World Order. All of this stuff has happened before. This is separate from crypto, and relates to the global economic environment.)
I have started doing something completely different than using bookmarks. I set up yacy[1] on a personal, internal server at my home, which I can access from all my devices, since they are always on my wireguard vpn.
Yacy is actually a distributed search engine, but I run in 'Robinson mode' as a private peer, to keep it isolated, as I just want a personal search of only sites I have indexed.
Anytime I come across something of interest, I index it with yacy, using a a depth of 0 (since I only want to index that one page, not the whole site). This way, I can just go to my search site, and search for something, and anything related that I've indexed before pops up. I found this works way better than trying to manage bookmarks with descriptions and tags.
Also, yacy will keep a cache of the content which is great if the site ever goes offline or changes.
If I need to browse, I can go use yacy's admin tools to see all the urls I have indexed.
I have been using this for several months and I am using this way more than I ever used my bookmarks.
First it fixed me. Clarity and strength. Slept better too. I did not at all expect it to do anything.
Then I felt intensely sad. All of a sudden really. One morning. After taking ALCAR for 1-2 weeks or so?
The funny thing is that this sadness was absolutely warranted. The ALCAR kind of… unleashed it? At risk of appearing to overshare: It’s been a really tough time recently – my spouse has been on a long road to recovery. Doing great! Life is beautiful. But I’ve been running in tough-it-out mode for a lonnng time. ALCAR “broke the seal” on it. I swear.
It really really really does seem to have an effect on neurotransitters doesn’t it!
It wasn’t really the best time to have it happen. It was hard. I have a small close-knit support network and have had to acquire a reasonable amount of insight into thought processes and emotional states and both of these were kind of necessary to handle this kind of sudden and specific emotional unraveling. Much more noticeable effect than any antidepressant or ADHD medication I’ve tried. (!)
It doesn’t seem to make sense but I really do think this is what happened. Not from a perspective of wanting it to be true. It was really quite a clear effect and I’m really glad to have figured it out and stopped taking ALCAR at that point.
And indeed there seems to be a scientific paper trail between ALCAR and the cholinergic system and this sort of emotional state and state change.
And there are tons of research on the interaction of ALCAR and the nervous system, in dementia and depression as well as diabetic neuropathy. And it certainly works for that – it clearly and obviously helps diabetic neuropathy heal.
I wonder how many of these people are just creating and interacting with tulpas without realizing it.
For anyone who isn't familiar, there is a subculture online of people who create what subjectively seem to be autonomous personalities that frequently manifest I'm the form of hallucinations. Think fight club sort of.
My ex tried it out. Called me one day at work, terrified because this dragon was following him around. He claimed to spend months trying to get rid of it, seems pretty shooken by the experience. But I was always skeptical.
So nine months ago I decided to create one of my one, just to try it out, see what it felt like. And...now I've had a talking lion following me around for the past eight months.
The best I can describe it is like some of the altered states of 'self' you might experience on LSD or ketamine. Thoughts seem to split off and go 'over there', and not be you.
When I talk to my tulpa, it at least appears subjectively like a separate personality state. I can be incredibly depressed but he can be fine. Or he will be depressed and I can be fine. I'm not saying there are neural correlates like you would see in a 'real' personality. Maybe it is all roleplay. But it is roleplay that fools me as the roleplayer.
For what it's worth, making a tulpa seems to have been really good for my mental health. I guess maybe you can see it as a form of self-regulation. I dunno, it didn't turn out at all what I expected. But it's hard not to think of him as a real person. I don't find myself being surprised by the actions of characters in my head, or laughing at imaginary friends. At this point, having out several hundreds hours into tulpaforcing, I can see and hear, and sometimes smell and touch him.
I know I'm rambling, I guess I'm saying is that even though I am skeptical of DID, or at least the mainstream depictions of DID, after making a tulpa I am a lot less skeptical of the subjective experience of DID.
For most of human history, people have lived in small, local information bubbles which were hard to penetrate with facts from the outside world. We see this manifest as the pre-industrial global patchwork of weird and wonderful religious/spiritual/cultural variety.
Then somehow the world started to join up and information began to seep into local cultures and sweep away a lot of these pre-rational beliefs. And for a while, it seemed like the direction of travel was towards a global system of thought based on a widely distributed factual, rational approach.
But in the last ten years, something seems to have arrested this direction of travel and reinstated local bubbles of thought, insulated from outside challenge.
Cryptoscams are the perfect (but not only) example of a self-sustaining shared consensual delusion. They don’t talk to the outside world. So much so that something like this happens. They are living inside a pre-rational info-bubble which has budded off from reality and floated away on its own wind.
We are not just dealing with scams and fools here. There is an intense current of reality-defiance and cult behaviour. Some people get sucked into the crypto-attractor, but other hallucinations are available.
The French have 100 varieties of cheese, and 3 varieties of nuclear power plants. In the USA it's the opposite.
More seriously, the French nuclear buildout (the Messmer plan), was a state driven program doing series building of large nuclear power plants. That is the recipe for quickly building out capacity on time and on budget.
(The delays and cost overruns seen today with the EPR's are largely due to there being such a long pause in building these that much of the knowhow was lost.)
> Its a kind of "vm" It has something morally akin to pickle.
...
Your comments here on HN hardly ever make any sense and they show a complete lack of understanding of both how coordinating conjunctions and punctuation work. They look like some keywords sorta related to the article but pieced together using a Markov chain or something like that.
I'm not a native english speaker but when I post in english I try to make at least semi-coherent sentences and I try to pay attention to punctuation.
Your comments are all so weird that I don't even need to read your username when I stumble upon one of them: I know it's you.
Your answers to people asking you why you are acting so weird also don't make any sense: it's more of the same nonsense.
I don't usually complain much here but it'd be nice if you stopped acting full on crazy.
Basically everything you post is so weird I'm wondering if I'm not talking to a bot and if that bot is not going to spout of the same non-sense in response. And if that's the case, well, I feel like HN would be a better place without that kind of pollution...
Like a lot of psychoactive substances, caffeine evolved as a pesticide in plants. Those plants tend to also evolve a lot of other compounds that modulate the effect of the pesticide. (I would guess there's probably some probabilistic/natural selection reasons for why this tends to happen.) That's why cannabis plants contain way more psychoactive molecules than just THC, for example, and why the experience of taking pure THC generally feels different from consuming cannabis.
Monoamine oxidase (MAO) is an enzyme that helps break down certain neurotransmitters (monoanimes, such as dopamine, serotonin, and norepinephrine). Monoamine oxidase inhibitors (MAOI) inhibit the activity of MAO, which can lead to less breakdown of those neurotransmitters, and so a higher neurotransmitter concentration.
This can have a synergistic and potentiating (or, if you're an insect, an additionally poisonous) effect for certain substances like caffeine. A certain amount of caffeine consumed along with MAOIs found in the coffee plant (e.g. from drinking coffee) could lead to more dopamine and norepinephrine concentration than the same amount of caffeine alone, for example.
This is also why it can be very dangerous to combine MAOIs with certain other substances. For example, combining it with tryptophan can lead to potentially fatal serotonin syndrome.
Coffee plants also contain much more than just caffeine and MAOIs. For example, theobromine, which is also found in chocolate (and which is responsible for harming or killing dogs if they consume chocolate).
I think this is the case for most psychoactive plants that people consume. You're typically getting a complex cocktail of many different molecules which all may have mutually modulating effects. This is why different coffee and cannabis strains can potentially cause somewhat different subjective experiences; they have different ratios of all the different compounds.
Here is a continuous quote (nothing omitted except the start/end of conversation) from a discussion between Andrew Lee and me from 2017:
[2017-08-03 11:48:56.616000] <rasengan_> justJanne: I like the way you think about a lot of things and doubly appreciate the fact that you share these thoughts so that others can know and understand the realities of them as well. But, I do want to emphasize that while I am in fact listed on some piece of paper for a couple of IRC networks, I am not in control of any. That's not only a promise made, and my word means so much, but it's a reality: I don't have an O.
[Editor's note: O, or +O, means that the person is operator and has administrative control over the network]
[2017-08-03 11:48:56.789000] <rasengan_> Whenever something comes up I don't get involved or don't even know. (ps: plz excuse my space bar i think there is something wrong wiht my driver)
[2017-08-03 11:49:37.530000] <rasengan_> But the thing is IRC has alwasy been a glorious thing. The infra has always been sposnored by companies or people. But the great thing about IRC i s you can always vote and let the networks and world know which you choose - by using /server.
[Editor's note: /server is a command used to switch from one network to another]
[2017-08-04 21:48:37.553000] <rasengan_> I'm fully confident though if PIA were to show any kind of negative intentions, people would /quit fast.
[Editor's note: /quit is a command to disconnect from a network]
His statements that he had no control and would never have it were believed by staff all those years, but due to experience interacting with him I never did and have warned against this very scenario ever since the sale (which prompted above interaction).
I too read that paper, but magnetico crawls Mainline DHT (which is the de facto standard) whilst the paper is on Vuze DHT, though the idea is more or less the same.
This project uses only one instance of crawler running at the same time, so one IP and one port for crawling the DHT. Whenever it receives a `GET_PEERS` query from another peer in the network, it forges itself a new ID that has the first 15 bytes in common with the info hash in the query, so that the querying node would think that it found the closest (or a close enough) node to announce. After responding back, most of the times the same querying node sends a `ANNOUNCE_PEER` query to register itself as a peer downloading that particular torrent. Upon receiving `ANNOUNCE_PEER` query, magneticod then connects the peer using BitTorrent (TCP) protocol, receives the complete metadata, and closes the connection.
Keep repeating these and you'll end up with thousands of torrents at the end of each day.
Possibly, utilizing multiple IPs might increase the throughput and prevent IP based bans, but I didn't bother to be honest, as the current solution seems to work well enough.
https://wiki.postmarketos.org/wiki/Zhihe_series_LTE_dongles_...
https://github.com/OpenStick/OpenStick
So yeah if you looking for hardware platform for weird homelab projects that's can be it.