Note that the audio doesn't follow the post verbatim;it's still essentially the same content though. (audio version below)

TL;DR- There's a bunch of groups of glorified science labs on mostly discord doing mostly AI research, you should join them

1 Definitions

This is not about:

Open access - The concept of having papers,etc be publicly available (this is a good thing most of the time.)

Opensource - Where code etc is publicly available to improve,etc (a good thing most of the time as well.)

Citizen Science- Where you get citizens to do work for you after some training (eg: finding things in pictures of space,etc);is related to open science, though. 

Apps, startups that mimic current solutions without offering research contributions mostly fall outside  of Open Science.They emphasize product development over knowledge creation, and often require a different mix of skills, priorities, and evaluation criteria than academic or interdisciplinary research collaborations.

This is about:

Open science labs- Where online researchers (of varying skill level- but usually around masters-equivalent* or higher) come together to do research; usually resulting in a published peer reviewed research paper of some kind, sometimes from differing fields leading to interdisciplinary work.  (I should probably change the name.; Collective Research Networks? Feel free to discuss in the comments.)

*A bit of a vague term but anyone who can re-implement cutting edge papers for example-

 edit: it's weird; even some PHD's can't do this; education indicators are very high variance these days.Like a paper without code probably isn't going to be reimplemented.

2 The rise of Open-sci labs

Discord as a catalyst:

After ~2017 with the rise of discord as a way to communicate with an in-group as well as an easy search function to find that in-group :

(eg:scientists,mathematicians,),cross-communication between many field experts from various unis or industry labs and newcomers have become more common; pre-2017 communication between these groups were closer to stack-exchange/help-desk rather than people on-boarding to help solve a problem/write a paper (GitHub was always a thing though; but it didn't lead to a paper, there are a couple of IRC groups but those were uni only.)

I think the reason most of this occurred in discord over something like slack is ease of use.

Open-sci labs

This has led to some of these spots turning into glorified open-sci labs: people come together to work on a research problem.

Eg:busy-beaver challenge discord finding a solution, various AI and informatics discords writing research papers, discord groups doing GitHub repos for more advanced topics.

This has led to a lot of interdisciplinary work between fields. (eg: AI discords now having people in psych, data-sci, maths, etc), or the various bioinformatic working groups have people from various sub-fields. As well as people people doing similar work in a subfield can find each other much easier to collaborate (eg: mech interp,tokenization,model building).

This is different from groups doing collaborations for a hackathon or a general chat server about a specific field : 

Types of discord communities

 Opensci labs does not have a strict definition so there's a grey area.

* they aren't fully study groups; closer to education servers as yannic runs a youtube channel about research papers,fast ai is an ai course, and flash linear attention has links to a lot of conferences on youtube in the server. (same thing- just more learning focused essentially)

Note: There are also high school open-sci labs but these  felt closer to spaces where you make/do something new rather than like strong research; there are separate tracks for them in various conferences but  aren't the same as regular research papers; closer to a small project a company would do for a community etc, still very impressive but doesn't move the field forward,mostly cause of lack of required knowledge:

3 How open-sci labs works

The day in a life of a researcher in an opensci lab is decently similar to someone doing research in uni.

Writing a paper

If you have an idea you want to work on, you can usually post a thread looking for collaborations if you need specific help;otherwise you can look around and see any that are interesting; it's generally a good idea to show some level of progress/work done though (code,papers to read, strong write-up; usually an idea doesn't get much traction;especially if it comes from a newcomer). 

Sometimes the lab has their own projects as well,which tends to get a lot more activity, This can mean either papers they are writing or github projects that they are doing, (which sometimes relates to a paper).

Eg: A paper can sometimes be a dataset,an LLM,technique,or something else where code is released, on github +the paper writeup.

Usually you need at least one decently skilled person in these open-sci groups to lead the paper otherwise it'll lead to people  not being sure what to do or make biased results. Can also lead to a situation where one person does most of the work as well, (which, for both hackathons and research papers are usually fine if the others are here to learn or understand this beforehand- i.e don't expect a spot in the author list or cash prize.) ; kind of like a research org with one senior helping a bunch of junior scientists.

This is very paper dependent, eg: if it has a lot of "grunt work" (like cleaning data for a dataset/benchmark), you might see people who aren't ready for research yet join to do some grunt work (and sometimes learn along the way)

The papers are your own if you did the majority of the work; sometimes you might give them (the opensci labs) special thanks etc so they can get funding, if you don't do the majority you usually still get a part in the authorship, can vary since most community projects are created by the community, rather than top down/by the admins/organizers etc.

Related, but there are also fellowships* which tends to be either IRL, or otherwise more intensive. (people from these labs sometimes host a track there). Most of the time it is a paper being done.

*Fellowships tends to refer to multiple things but it usually means some sort of advanced training,sometimes with a paper at the end.

Learning resources

There are also research paper reading groups in these labs,These spots usually have a lot of resources to help people work on a project,given that you're already decently skilled (eg can re-implement papers or close to that skill-level.) 

If you don't know the general pathway to up-skill, (usually because searching has been hijacked with paid courses that don't help much) some of the labs redirect you to the learning discords/study groups or their own pathway, or tell you to join the reading groups/arxiv papers depending on skill level.

A lot of the comp-sci labs usually have people around the masters or below level formally, mostly because a lot of them are self-taught afterwards.

For a bit of context; the reason research paper reading groups exist in the opensci labs is because its kind of like literature review- learning spots usually don't require that, and advanced courses tends to have the papers to whatever subfield/niche example/capstone project you're working on rather than regular reading of papers, which can sometimes be outdated or only used in industry (eg: decision trees!).

Specifically,what skill level should i be to write or help?

Essentially, you should be already past ~ AI 201, otherwise best to reach to that point! (once again, there are the study groups and resources to learn to reach to this point.)

I.e:around the skill of being able to use huggingface,build pipelines etc;although probably best to be around the skill of replicating a paper if you're planning to write one rather than assist with one .

Other subfields probably have around the same classification; junior if you're helping,senior if you're leading (mostly for the extra theoretical knowledge thats grounded by reality on how this works). 

That's not to say you need strong theoretical skills to do research; You don't need a strong formal math background to optimize AI - you can just try things and see if something works and figure out why it works afterwards; eg for interpretability you can use the systems to interpret how a model works without requiring a topological understanding of mathematics; usually your peers,AI, etc would help.This is probably because AI still has a lot of low hanging fruits and a better understanding of the problem is a better way to do things than learning theoretical math to try to attempt new stuff rather than  explain things better.

If you're here for the reading groups you don't need any background to join; the requirements to write a paper are more of a soft requirement cause of problems with crackpots. Most reading groups are uploaded to YouTube (and there's also all the unis etc posting their reading online, and like researchseminars.org)

Academically; if you already did a diploma/degree, you're usually 1-2 years away from writing papers,i.e: PhD students write papers,not so much PhD holders themselves.

(please note interpreting how something works after you've measured it does not work for other fields- especially humanities.)

Whats the skill level looking like?

I ran a quick poll on the mech interp research server,interesting results, this place is more specialized than the labs.

I messaged Neel Nanda about his mech interp team and he said:

"Of the soon to be 8 people on the GDM(google deepmind) mech interp team:
- 2 Maths bachelors (including Neel) 
- 2 Maths masters
- 1 ML PhD
- 1 ML PhD dropout
- 1 Stats PhD drop out
- 1 Physics PhD

Fairly few of Neel's MATS scholars  have ML PhDs but that's non-trivial,"

Chatting with people from the MATS  track I went to in 2023 a decent amount of them definitely had completed undergrads at most.

Might be track dependent? although probably won't change much.

Neel had no industry experience before getting into AI (did some non ai internships pre-AI).

How are they funded

There tends to be two groups that are giving funding: 

the philanthropic/ charity-funded labs:
Tend to focus on more theoretical type research (toy examples,new techniques on smaller models,interpretability,etc). Sometimes also gets funded by grants.EA groups sometimes funds these spots as well.Money pays the core team+anything that's needed.

 DAO's
 A crypto collective that gets a lot of funding from crypto that then funds various research that other DAO's do (among other things).  DAO's tend to  have various informatics/applied/production ready to fit into a pipeline type research (bioinformatics,chem-informatics,drug discovery,etc);

Things do feel different in the two spots, DAO's feel closer to industry, labs closer to academia. DAO's tend to have less a top-down structure and more decentralized. While labs don't.

open source research hybrid:

Some open source groups usually have some sort of paid service/product that consumers can buy; API for an AI model, rarely paid one on one services, conferences, courses,etc.Usually part of a company so they get investor money etc.People who contribute sometimes get a job,bug bounty cash,etc.

As a whole,though,There just isn't much funding for these stuff,especially compared to academia;and significantly less for anything not in AI (makerspaces etc are IRL so they get their funding via IRL stuff)

Research managers 

In a lot of the groups; more notably in spots where the admins (who are usually also subject matter experts), take on the role of a research manager- Akin to a mentor:  they give guidance for harder problems/questions assumptions, introduce you to other researchers, or otherwise really help you along with your own research;Besides that they sometimes help with funding,conference stuff, Have a high level view landscape / good research taste of what the entire field is up to,Or otherwise just outright help with complicated issues if required.They aren't too common and sometimes the lead takes on the role of a research manager.

There's also always the community managers who help with the community-events,discussion, sometimes where should the group go,marketing,etc.

4 Problems in open-sci labs

The crackpot problem

Crackpots come into  open-sci discord labs and spew their nonsense ideas, looking for validation- usually being guided to the server(s) via chatgpt  etc.They get banned. The hackathon spaces and entrepreneurship spots tend to also have this problem but it usually results in people not taking them seriously or people attempting to write code that doesn't work.

There's a non zero chance this fixes itself in the next iteration of AI that decides to call all their ideas stupid and helps them/says they wouldn't want them here.

If more of these people pop up entire servers/spots can be full of crackpot and crackpot ideas akin to vixra if quality control is not done.

A lot of projects go no where

A lot of users start projects that tends to get abandoned very early on; mostly cause of activity issues from members involved;or otherwise the project was not good;because poster wanted to post an idea rather than do research, or that thing where its easy to start a project vs continue it mostly cause it seems easy.

Failures in research

Those especially new to research coming from an outside perspective finds it hard to realize you can do research for 100+ hours and it ends up sucking- especially if they haven't done something of that scale before (eg an app or game or something); which usually causes attacks of the research to mean attacks of the person from their point of view. 

Communication issues

There is also communication issues between what an expert says and what a newcomer says (and sometimes even what the newcomer means; especially if its a chatgpt assisted message.)

On-boarding difficulties:

Skill levels can be quite varied, it's hard for someone to rank their own skill level (dunning kruger etc); This, can lead to a situation where:

1: A new user joins,looks at some projects, thinks they are all too difficult, and leaves, except if  they decided to spend a week with the project, they would've been able to help. (theory practice gap).

2: A user joins, looks at various projects, never asks for help because of imposter syndrome

3: People more new to discord get overwhelmed with all the options and leave.

4: A lot of people join and say/do nothing; this is probably related to larger issues  with work ethic etc.

5: For community made projects, sometimes people just don't know how to help/it was badly worded, and the poster leaves.

Various discord problems

these are mostly low impact but discord being chat rather than forum leads to faster discussions which can cause topics to switch quickly.Most groups partially fix this by having specific channels for each topic+ some kind of log or something for whats everyone's up to.

5: Success in open-sci

Impact

These places usually have best paper and other conference papers,arXiv papers being done every year,.

There are also some groups working on large Models and other useful things that i thought i should note.

There are a bunch of other groups but they tend to fall into one of the other labels in the types of communities table. (mostly GitHub working groups; labs usually use their tools.),There are some that also does research paper reading groups without being a research lab as well.

There are a lot of different types of opensci labs but they can be mostly broken down into 3 types: 

1: Everyones insanely skilled so not much reading/upskilling required (can start from advanced readings)-everyone is in a subfield so they can help each other

2: Less skilled so more events,more low hanging fruits,upskilling etc to do - slightly more general audience;larger team sizes 

3: Turn things into a token/hackathon by talking to experts about what needs to be done.

Not fully sure how to compare this to regular unis; mostly cause I can't seem to get a solid number on how much papers a uni publishes a year (most of the results are per professor).

But I am almost certain paper cost ratio is lower, although that might be cause most of the groups are comp-sci adjacent which has lower paper costs,as well as people who are here are usually self selected /already highly skilled so not much funds/an funds needed to upskill (because it's all online)

Graph was made by checking the affiliation tag in neurips papers from the public repo of all the accepted papers,There might be a few more groups I'm missing,and huggingface isn't exactly an open science lab.

(note these aren't all the papers these groups made-theres a lot of papers that didn't make it into neurips; some by first-time publishers.I know eleutherAI held a summer of open research program- it's completed but some groups are continuing their work into workshop and arxiv papers.)Also huggingface may not  be an opensci lab.

There are a some industry research labs (open code etc but with a product) with a much higher amount of papers-Eg: allen institute has 31 papers, though their discord is more for users rather than researchers. (there are research spots but its not too active)

There are also a lot of dead servers, most that never really took off, but they still had an impact in the field/people moved to other spots.

There are various IRL hackerspaces and fellowships etc that sometimes produces high quality research by undergrads,highschoolers, etc- I know there was one in math where a 17 year old was able to disprove a conjecture.

Non paper impact

Besides these; these places have also been a great spot for hobbyists and anyone who self studied to apply what they learnt, or otherwise help with policy experts,sci comm,etc,or  a way to network or move into startups if an idea is strong.

Most of these spots also have labs,IRL offices, or something of the sorts in IRL places.  (most of them are companies or non-profits)

Example: both eleutherAI and manifold has companies that were formed from them by users in the server.

bio.xyz has done a lot funding for various research,and cohere has created a model among other things

Cohere for AI has the BIRDS program which helps bring people of various skill levels into research, usually by doing cohorts of reading etc before doing any research.Most places have reading groups as well which has similar effects. This is really useful for teaching junior researchers or people new to research how to do research.

These places also serve as a great spot for interdisciplinary studies/reading/people to meet- which can massively boost progress if multiple different subfields are looking at the same question.

Usually after a few years of being in these spots,users tend to  leave to go into industry; usually taking on problems in the valley of death in innovation. (gap between industry and academia where papers go to die since its not scalable/more info needed),or advancing their own research but as a product (new models etc)

How to up-skill: The alternate path to doing PhD-esque research with little formal education

You can do a top-down approach, where you get some state of the art papers and replicate them with claude; or, you can go through a more pseudo standardized pipelines of various courses eg for AI:

Pipeline (my opinion): basic python, then  fast.ai, then a sub-field*(eg ARENA for mech interp,gpu mode videos for low level), then replicate papers and do research; From here you go back and learn about adjacent stuff that might be useful eg: stats,math,etc.

This shouldn't take more than a year; and you can just go top down instead as you'll pick up more as you learn along.

*Subfields in AI are very close to each other; you usually only need a week or 3 to understand what your peers in another field (eg:audio,image,text) is doing,even those further away (low level gpu,applied eg:bioinformatics) also doesn't take too long. 

An understanding of research as a whole is useful to help ease friction;Arxiv system, papers,conferences,etc, you could ask AI this,watch some videos,or look around. Especially useful if you have a lot of industry experience eg: 10 years in python or something as its usually something that can cause friction/not measuring things correctly.

A strong example of what i mean for understanding would be like  skeptics guide to the universe,or https://casp-uk.net/casp-tools-checklists/ , https://jbi.global/critical-appraisal-tools , https://methods.cochrane.org/bias/resources/rob-2-revised-cochrane-risk-bias-tool-randomized-trials (these pages change url a lot, just search for it)   Talking to other peers can help with this too!

There are also a lot of links at the bottom of the write-up for a less technical understanding of what research is.

Top-down: Grab whatever state of the art paper(s) is, replicate it on github, ask opus to help, and peers whenever its stop being able to help,repeat this a few more times,making sure to understand what each choice of code does (the choices and why, you can ask the model questions about this,its the same as googling and copy pasting info vs at least reading the info and having a vague understanding of it.) 

(Seriously If you don't know what to do, for most compsci, AI, or asking research questions stuff, try asking opus-it is literally a great equalizer/million dollar tutor at your fingertips-it can massively accelerate learning speed.)

Realistically you'll probably do a mix of both : reimplement a paper, have an AI explain terms every 5 words, eventually the model links you to a course, then you come back and reimplement/ask AI to help with the course etc.

Fastai and cs transformers both have a discord community and has an interact-able study course online. Other fields tends to have some sort of course/discord as well (or a general overview of all subfields in that field you're looking into going into in the casual grad-academic servers).Eg an AI grad server would have most of the AI subfields as channels.

Besides that, there's also various companies that now have  feature bounties- not for bugs, but to help fix problems they have,eg: tinygrad has bounties to write code/bugfixes, to help implement in its framework to various GPU's.

Hackathons have also become quite advanced- kaggle has various AI challenges that are clearly meant for researchers/ senior engineers.

The biggest pros with top-down is you essentially know whats going on in the cutting edge of things (once you're re implementing new papers),thus really being able to get an intuitive understanding of the math compared to just numbers and theories on a screen, biggest con is that its very discipline driven and you'll be confused a lot if you don't have a background in math/cs etc so you'll need to recursively ask the model to help you understand stuff.

For bottom up , the biggest pro is that it gives a view of some techniques that may not be used anymore, and its easier to follow, while the biggest con is that it's usually not fully up to date with existing literature.

6 How to get more people to this point

The current state of things/public perception

Most undergrads or boot-camps tend to stop right  at undergrad rather than up-skill some more; usually cause of costs,issues with finding work,etc.

Besides papers, and outside of AI; A LOT of grad software requires help from cs, most are used in research in varying fields and I think that's also a good start  with beginning in open-sci; especially if you're out of undergrad and looking to up-skill.

Most users are in the US,but there is a large amount of people in India who wants to help.As a whole  though users would rather apply if things are more formal (eg a fellowship) rather than more casual (eg: help a github repo). This has lead to skill issues with people who are  under-skilled since they haven't really done any non formal stuff.

What can we do to grow this place?

Create more spots for people to up-skill and get them here (up-skill from diploma level to masters/junior researcher level,that is.)

Or, get some of these groups/experts to talk to each other more often; A lot of users of communities/experts are very closed off and aren't aware of other groups/people- would be great for knowledge sharing,posting events,etc. 

More IRL hackerspaces and fellowships etc can help but I don't think it would be the best use of money; they are usually too localized for what its worth.

Issues getting people here

One of the bigger issues is that the general public wouldn't think of research as a viable option/traditional path  - A change of this style of thinking to thinking of research as another path to choose would be incredibly useful in boosting the amount of people here.

Sci- comm also plays a factor into this style of thinking; the #SOME made by 3b1b has helped shine a light on whose doing what,  

Something else of note is that there's a difference between doing a online course, doing research for someone, and doing research for yourself; it tends to be best to do all 3 in sequential order so you get the soft skills needed; eg: for a project more actionable you probably wont be doing literature review but writing code,but for your own project you'll prob be writing code and doing a lit review.

I think it would also create a lot more people reading research papers or doing advanced tutorials ; which is a soft per-requisite to being able to do research, which,there isn't much reading groups  outside of academia.

It's kind of a web of issues of anything past undergrad-equivalent being more difficult to access, leading to less people being able to do stuff even if they find here, leading to less people thinking its a path to doing well in life, as well as people already in masters equivalent not knowing about here/in academia.

More mentors, better upskilling to get people up to this point (something to bridge the gap between undergrad-equivalent and research) would be great.

Or, Better impact by asking professors,research companies etc open problems to work on could be good to help people see here.Bio.xyz  is close enough to this;although more fields should join; with skill levels stopping users from falling for a crank project or GPT wrapper. https://www.gap-map.org/ 

Realistically issues to get people to this point (especially with regards to being able to do research) would require deep cultural changes, though just connecting people who are already skilled/high ambition to researchers etc seems to be significantly easier and higher impact to do (which is what most of the AI safety ORGS and EA do!).

Shifting the global culture into taking  research more seriously would require the help of memes or something.Culture,memetics,psyops,marketing, you hopefully know what I mean (I could probably go into details in the comments). Like-Do you see how entrepreneurial culture is like?

Or the whole thing about some people  not being able to think about stuff (learned helplessness from thinking narrowly in school/societal norms/education being a checklist).

Rather than seeing learning as constantly growing, laziness to explore whats out there or even the ability to know what to ask to move around in a field, mostly cause these are usually learnt in PhD's/researchy extra-curriculars/pastimes ,which is why there's sometimes situations where undergrads are more knowledgeable than masters.

Or the laziness(?) crisis, where people pick the path of least resistance/self-defeating cycles,or even just the various structural barriers stopping them from having time to do this.

Or the problem where people attempt research, realizes it's hard, then gives up. (except research is failing repeatedly, if you'd ever made something eg a writeup/video etc you'd know how much extra time it takes to produce the content that gets tossed out).Relates to people in fellowships etc just leaving after the first week.

I think a lot of people could retrain themself akin to doing a PhD but it will probably destroy most free-time/ability to consume content etc......

This conversation can go on for a while, pick any aspect of human behavior or societal structure, and it explains why so few people are doing this kind of independent research or could get to that point. There is potential here to shift the ideologies through memetics and advertising tactics etc though.

As a quickfire of issues:

• Learned helplessness from failing grades/school
• No payment compared to working a dead end job- no time to do research when home
• Don't know how to go from MOOC's to reading papers
• IRL group dynamics don't see research as a career path/distrust in science
• Path of least resistance leading to people not doing things out of norms.
• Discoverability issues
• People just aren't real with themselves; how many of  you all said you'll go to the gym/do something from January etc and then just stopped early on?
• Unknown unknowns stopping people from gauging skill (relates to crackpot problem)
• A lot of various stuff are coupled to these issues with varying impact:
  • Hustle culture/gig economy (people going into dropshipping, side businesses, rather than upskill.)
  • Lots of misinformation wars/misunderstanding of the crisis in science
  • General passive consumption (relates to passivity of learning)
  • Academic culture being split into multiple subgroups-more noticeable for diplomas which gets coupled into PhD life,which relates to views on science
  • Popsci/pop culture's views on what a scientist/researcher is
  • lack of "research culture"
  • Views of learning as a checklist rather than something active

So why don't people already do these things?

If you try googling AI course; you'll be shown a bunch of incredibly expensive courses that might not teach you the skills needed.

This is usually because research skills are different from applied skills in AI i.e an AI course can probably tell you how to implement a transformer architecture but not actually explain what it is, what latent space is , etc; i.e its a lot more code based/applied.

If you're new to AI its kind of hard to know what you should pick- as it's an unknown unknown. 

Experts in the field looking to find these spots will also find it difficult- it's still mostly known through word of mouth which is good to stop the crackpot problem.

It's a tradeoff but I think there might be ways to help both groups (eg: verification through some way to show your skill level; not ORCID/arXiv since that's a catch-22; maybe word of mouth from the other learning spots? write-ups you did on lesswrong? unsure.)

The passivity of learning

Most groups tend to do whats given to them eg: a boot-camp or course that was advertised to them- rather than looking around. 

This is probably because upskilling past a point is hard (and no one really trusts/uses claude yet for these things despite it's potential-as well as claude not being super useful if you don't know what to ask it cause of unknown unknowns).

Its hard to break out of this style of thinking.

It's like that thing where everyone in self help thinks the right thing to do is read 48 laws of power/how to win friends and influence.

Probably because they don't have someone  with a PHD or something (successful businessman) to guide them with next steps etc.

There is an issue here of people thinking that learning mostly takes place in school/uni, not really thinking you'll need to learn more during your job, much more true for research since you need to do literature review, sometimes requiring extra skills you need to learn to parse the literature.

 I think the biggest filter to anyone who manages to reach up to this point (get in the learning server or research one, wanting to write a paper/self-study), is motivation,like  80% of all people in self-study just leave because of motivation issue?*
*might not be fully true- ~80% of people in MOOC's tend to leave early on; but could be more reasons than just motivation

Even when people reach to doing papers, they sometimes just do whatever people tell them to do.

Essentially there's the motivation issue (i.e: low executive function) and the inability to find good resources or know what to do (i.e: low agency).

7 Why go into open-sci

Resume boost/learning/career forward

I do think writing papers would be a great way to boost your resume and learn if you aren't from a academic background and looking to do research work.Various open-sci people went on to work in the major labs , usually with them having little to no formal education background. A great example here would be Leo Gao who co-founded eleutherAI and is now working at openAI; no formal education as far as i can tell.

There are a lot of other names I can list here but the groups are still decently new for any real quantitative analysis, or even figure out people who tried and failed this pipeline (excluding people who started and didn't actually complete their project for example.)

From writing papers, some tend to go on to get a job in industry  as a junior researcher etc. ( with no formal education),or showing skills with helping out the GitHub projects the labs use/created and someone hiring them from there.Kind of like how running a substack gets you a hire as a writer for a company sometimes,or really all the other stuff in non STEM industry (indie game dev,solo musician,artist,cs person,writer,etc all being picked up by larger companies.)

It's just a great stepping stone to go into working for the closed-source labs/industry/government contracts, or start a technical startup,especially for non traditional academics.

Tinygrad's bounties also leads to internships/jobs at the companies.

Networking,fellowships,get advanced help

Great spots to get help with any advanced problems you might have; and find people to work on your project with you; although most of this is specifically for AI and any AI related fields.

AI safety,research, and policy fellowships and research sprints etc are great ways to do research; they are similar to a hackathons but usually goes on for a longer time; occurs at various times at the year; as well as all the beginner programs.

These hackathons are usually shared in the labs eg: MATS being shared in an AI server.

Besides the labs there are usually also IRL fellowships/research tracks Sometimes with a stipend etc which is usually a decent amount of money if you aren't in the US; these tend to be more personalized if you aren't motivated but can lead to more issues (bad reputation if you leave the track early on/ghost etc)

Does help with networking and investor finding etc too as the labs are usually funded by some of the  philanthropic organizations; and some of the members/tracks rarely go on to start companies and get seed funding; really depends on the track;more entrepreneurial ones might have this occur more commonly.

Sci-comm

Besides that I do think this could help with sci-comm if you're interested in that field as you get to meet experts and ask them questions etc in the sub-field they're working on.

Also, what else are you planning to do with your free time? I think especially if you haven't been in research circles there's a lot of cultural and social unknowns that you may be wrong about that the general public thinks is true.

8 Why did mostly comp-sci,AI rise to this occasion?

You only need a PC and data to do research; you're not going to become a medical professional doing open-sci,as there isn't really a path to start doing medicine in the open (no one's going to let you operate on them,for example,and new papers tend to be created by events doctors etc see in real life , at least sometimes/depends on the sub-field. 

Humanities also feels like it has a set of pseudo researchers as well; at least for  political-science and philosophy etc.Probably relates to the paper crisis; paper quality etc too.As well as the larger lobbyists etc cherry picking papers to use,and it being harder to replicate studies compared to compsci (copy the code onto your PC and test it) 

Biology,chemistry,physics,etc has a lot of advanced online courses (mostly the informatics subfields), But it's hard to go from there especially if you need to synthesize materials or aren't in a country where there are openlabs/hackerspaces (that's not to say theres nothing- https://sphere.diybio.org/ seems useful for bio for example.). But costs for experiments tends to make things difficult, and without deep knowledge of the field no one will know how to do low cost research level science  with kitchen tools (eg: the graphene tape experiment-people made really small graphene molecules by taping graphene and pulling it multiple times over.)

It's a multitude of factors such as but not limited to:

• Open-sci  and computer science movement away from academia being a thing already pre- discord (github,etc)
• There more people in comp-sci
• More unemployed/underemployed people in comp-sci; most other fields work in their uni labs/industry.
• Funding/ compsci culture (community as a workspace;its why robotics is active too)
• Easier to test papers etc
• Lower barrier of entry
• Most advanced courses are both online and replicatable.
• Most people just aren't aware of other subfields-i.e they are invisible.
• Compsci kind of teaches you a lot of research soft skills indirectly so it's not too difficult to start doing research.
• Lower skill floor so people don't spend as long learning to do research
• AI especially is all online - everything can be done virtually; no need for hackerspaces etc for robotics or equipment etc.

i.e The you can just do things mindset; and as such grew as more people became aware of it.

A lot of those factors are probably also why there are so much high school type research groups-lower gap,easier entry,free time,doing your own research,etc.

 Biohacking is probably the closest to AI opensci labs culture wise but it's all IRL and they are currently in a "research winter".

There are a bunch of grad servers for the other fields (eg math) but its closer to stack-exchange - more about asking questions and getting answers; for your own work, rather than helping out a larger piece of work.

Anyone doing co-working in other fields is mostly in private hubs /DM's; where they already know the person IRL/beforehand; not really  looking for people to collaborate.

One group that seems to be having  success is https://www.independentscience.org/ -  They are very new and came out while i was writing this essay, if you care about non AI non compsci science (mostly hard science -especially physics), you should check them out!

9 I have my own research but it's very unique/different but almost certainly not crackpot; what do i do?

I'm guessing you talked to the experts or still can't find the experts, if so, you should read this:
https://defenderofthebasic.substack.com/p/how-do-we-bootstrap-the-open-research

https://www.cosmik.network/

https://exa.ai/

The TL;DR is that because each sub-field is so small (usually less than 100 people) ; you might need to do a semantic search to find your peers; i.e they might not be in a discord/other spot somewhere.

More common if you aren't from academia and as such not finding the keywords your peers might be using in their papers.

10 So, what can I do?

Not comp-sci adjacent

If you're already an expert; try looking at the pipeline required to make it to where you are today; if there are any places that are free,open source, etc make note of that; try building a community and advertising those spots to those new in the field, while also getting people together to do research in your field.

If you aren't an expert, try upskilling! you're probably still a bit to go before being able to do research; maybe ask around what paths to take in the more social graduate spots. Heck even if you don't want to do research,upskill! having an advanced knowledge of things can help if you decide to go the entrepreneurial path too. 

Comp-sci adjacent

If you're already done with an undergrad type course/or at the point  you can replicate papers, try joining the labs! they can usually point you in the right direction if you have a sub-field in mind, or to the study resources,or compute resources if needed.

If your undergrad had a lot of AI in it/ you did an AI course as well (eg:masters or phd), you can come join us! feel free to help out or put your own research that requires help, or lurk and join whenever we read research papers.

Otherwise, try upskilling some more!;you can still join the labs as its a soft requirement to do research, while there is no requirement for the reading groups. They can point you in the right direction but there might be better choices.

If you're specifically in highschool you can join the highschool labs. (eg:rishab academy has an aggregate of most of them).

talent agencies for science?

I have decided to add the spreadsheet of all the servers! It's near the bottom of the post.

for both groups there's also these links:

https://github.com/BitGeek29/awesome-abandoned-research

 https://eamag.me/2025/Project-Ideas 

https://www.gap-map.org/?sort=rank 

https://openreview-copilot.eamag.me/projects 

https://www.grandchallenges.org/ 

https://ahead-lightyear-d16.notion.site/Open-Source-ML-AI-Projects-for-Immediate-Contribution-25c4bd90b7948079b175d8e268e58ad1

 https://euclia.vercel.app/

 https://www.independentscience.org/

Funding ORGs

You should probably fund these communities- they tend to create a lot of impact .Spreadsheet is near the bottom of the post, The main groups here  are:

AI research: Eleuther, cohere labs ,alignment jams/apart research ,ai-plans,open neuromorphic

Non AI: Manifold research group,independent scientists

I am willing to hop on a call/respond to any messages about the sheet,who to fund etc.

In general though I can't say who to fund- those are safe picks since they have impact;but would lower impact spots be better if they get funding? unsure.

There's always the suite of github projects that would want funding too (third sheet in the sheet).

Maybe run some grants for these stuff? they tend to fall out of most types of grants from what i can tell?

(feel free to fund me as well! more info further down).

https://manifund.org/projects/openscience-projects-to-get-non-academics-upskilled-and-involved-in-research-u8cgglp1jc

I want to build a community 

The biggest things working against you here is time-you really need to put in a lot of time to micromanage everything early on when things are moving quickly and make sure people don't leave early on while the community is building (afterwards it'll snowball and you can be more hands off) some suggestions are:

• demo days,share you work,weekly events/write-up of discussions had
• spreadsheet of whose doing what/who wants to help so things aren't an ideas soup
• notions/coda so ideas aren't on discord since it can be forgotten by new posts
• guides on how to use git etc to help ease friction
• draconian type ruling so people don't procrastinate.
• incentives (stipends,certs,etc)
• weekly check-ins,task board/open research calls so contributors can see what to do,apprenticeships
• design around the people staying active as a lot of places have high dropout rates
• videos,collabs,conferences,papers,seed funding,posts etc to help with recruitment,especially after a project is completed
• AI to help with boilerplate, project pitches/chat about ideas on call,hackathons etc experts to help with quality control
• work with unis,companies,publish everything open,industry demo days to companies etc
• various rules (all work is open?) to help with IP drama from working with those groups
• Partner with hackerspaces etc to use materials/equipment if required.but post results online- "video research papers" can help a lot with marketing and understanding of whats being done.
• Think about your skill-base- the skills of all the people you have; are they enough to do research? if not, how long do you think it would take to upskill them?
• Make the space for the people whose active (see below)

These should also all help move an existing community from just chatting about science to doing science. 

The biggest "problem" I've seen with groups/cohorts/fellowships/events esque stuff is the high dropoff after the first week or so; having a burn in period would help (eg: coding exercises,merge requests for github projects,  resumes would be useful but I think skills are pretty vague on resumes.Maybe this is just a problem everywhere and safely ignored.Although I have no clue where these people keep going after leaving.

11: Future Trends

With the funding cuts, I guess we might see more people taking this career path especially if they would otherwise be not in training,education.

Maybe we might see more in humanities/math? I don't think math is too collaborative because of the nature of their work (most collab is using the servers etc like stack-exchange 2.0 for research you're working on), Humanities may see an uptick especially if more people become interested in internet type data-sci research.

Perhaps more free courses for those fields might occur too; humanities could use more,math might have more interactive math courses,but I don't think it'll be too much because math tends to usually need pen and paper.

Perhaps even more comp-sci adjacent sub-fields might have groups now? the DAO's have created a group for a lot of varying science+computer sci interdisciplinary fields,but there probably are still some gaps.

The larger online ecosystem will probably develop some more, with more fellowships,hackathons etc to help move to this spot in the pipeline.

There aren't too much spots compared rn to  all the regular uni's, there are all the cs spots that feels like they are at least masters worthy (eg: deeply developing the rust codebase or something).

If I compiled hackerspaces+research internships etc it might look bigger,especially for humanities,but idt it would be like 5% of the uni population (especially cause most of the people who got jobs etc there are in uni.)

Probably decently higher if I only include diploma holders (20%?); in humanities scicomm, has a higher impact than formal researchers.

Didn't talk about them as not sure where to place them.Could just lump them into cs not having a degree/up-skilling at work to masters from a degree. (more stuff to write in the book ig.)

Perhaps the new juniors in research get all the grunt work like they're the new masters/PhD's.....

Extra: Difference between research and entrepreneurship

Entrepreneurial and research tends to be similar depending on what you are doing- eg: if you make a new material; you could attempt to upscale and sell it to businesses/consumers who would want it,or write a paper about it

Not all entrepreneurial projects can be papers though (eg: if the material is clearly just another material with something that done that's obvious- eg a new phone that does nothing new in 2025 would not be a paper, but an android phone in 2005 definitely would be paper worthy.) 

Not all papers can be entrepreneurial projects (eg: code/technique that increases an AI's score trustworthiness score by a few percentage points can't be a entrepreneurial project,but a series of papers,compiled, can be used to start a consultancy or otherwise some type of industry research lab type business.)

Most groups go from academia to entrepreneurship;there isn't much that goes the other way,usually cause you're probably already an expert if you're in industry, as there aren't too much people who go from junior to senior in industry without knowing much/doing research work on the job; and even if you are in that pathway,you're probably an expert at that point, so no need to do grad-school.

(as well as entrepreneurship mostly meaning things like drop-shipping etc to the general public.... which is not what I'm referring to.There is  other stuff like culture etc here that I touched on.If i get funding to turn this into a book expect this to be fleshed out.)

12 Conclusion

(not sure how to conclude so here's chatgpt concludeing it)

The emergence of open-sci labs represents a fundamental shift in how research is conducted in the 21st century.

What began as informal Discord communities has evolved into legitimate research collectives producing peer-reviewed papers, launching careers, and challenging the traditional academic monopoly on knowledge creation. 

This movement addresses several critical issues in modern research: the accessibility gap for those outside traditional academia, the siloing of expertise across disciplines, and the inefficiency of conventional research pathways. 

By lowering barriers to entry while maintaining quality through community peer review, these labs have created a new model for collaborative science. 

The success stories—from EleutherAI's ICLR papers to the various DAOs funding applied research—demonstrate that this isn't merely a hobbyist movement but a viable alternative research ecosystem. 

For computer science and adjacent fields, this has already become a proven pathway. The question now is whether other disciplines can adapt this model to their own constraints and cultures.

The open-sci movement isn't just about democratizing who can do research—it's about re-imagining what research communities can be. 

In an era of rapid technological change and complex global challenges, we need all available talent contributing to human knowledge. 

These Discord servers turned research labs might be the prototype for how we organize scientific collaboration in the future. 

In a world where knowledge work is increasingly accessible, the question isn't whether you can contribute to research, but whether you will.

What are the next walls to fall?

Contacts

discord: seonresearch (most active here)

email: seon04877@gmail.com

xitter: https://x.com/SeonGunness (I'm on here daily or so)

linkedin: https://www.linkedin.com/in/seon-gunness-85405b10b/ (I'm on here weekly)

So why did I write this?

I don't have any formal education after high school.

I'm not from the US, and it does feel like the culture in most of the poorer countries/communities is to learn computer science/be a doctor/go to uni, etc - except most of them stop at python when there are things you should be doing after that.

It relates to the whole thing about research not being a way out currently.

Heck I don't even think a lot of people here know what a PhD is (like the timeframe,what a dissertation is, etc.)

I hope to change that.

There has been some progress in advanced upskilling (various free AI courses,datasci courses,etc),But not much work in moving these to research side of things- mostly cause of the passiveness stated earlier. But realistically the issues go much deeper; relates to a bunch of problems with society, some of which i went over above,some can probably be fixed with a kind of rebranding of science (eg:like how duolingo got people to learn, or a lot of people doing conlangs, or even just anyone interested in science probably going into philosophy first) marketing,propaganda, etc.

(I have a paper coming out in regards to AI red teaming;should be out by the end of the year).

What am I going to be doing now?/fund me!

I need funding.

https://manifund.org/projects/openscience-projects-to-get-non-academics-upskilled-and-involved-in-research-u8cgglp1jc

What i'm certain i'm doing:

1:Map out the discord/whatsapp/facebook/etc ecosystem for AI and non AI. (groups etc) 

https://docs.google.com/spreadsheets/d/1DlBT1pF8-zMECntRWXFsL46gZyvNp1BJlJ6LXGze4dA/edit?gid=0#gid=0

above is the sheet in it's current state, I did a non research track during SOAR to get the sheet into a site:
https://eleuther-ai-project-soar.github.io/

It's in a working state,i'll need to move it somewhere else soon though as github only has static pages.

The first sheet has a lot of AI stuff,second sheet is all the non AI stuff.

2:Turn this writeup into a book

Lots of extra stuff to write about. Like i talked about hackathons three times but do you remember me talking about low executive function and low agency? what about interdisciplinary?

3: Get more people here who won't be doing this:

You remember that image earlier about psyopping people into loving science (chapter 6)? I'm planning to do something like that but I don't have a full plan of action yet.

although, I have been getting dm's since the past 3 month; It's mostly been informally where people dm me asking for help since I seem to look knowledgeable in the servers ig, and send them links/things they should know etc, small impact here and there, can confirm one person was able to get help co-writing a paper, otherwise not much so far? sample size is pretty small though (8 in total),I know another might be joining some stuff for funding etc,but too early to tell.

Rather than getting top talent to do research, just train a bunch of people to become top talent.

4: If required, monetize this somehow: 

Making the resources paid would not be a good idea but There might be a way to monetize the people who use these resources to up-skill against themself:

Kind of each person is their own stock- and if they're able to do well/meet goals/do hackathons etc, the stock goes up,otherwise it goes down.

Kind of like a crypto coin etc? something like this https://codetiger.me/project/StonksCF/about.html or this https://cvffund.com/blog/pipe-a-fintech-pioneer-in-revenue-based-financing/ , or just like those personal stocks you see on manifold markets? 

What im not certain about doing:

Get a list of people,skill levels, and what they're looking for/ what they know/unknown unknowns. (mentor mentee list), Create grad level study groups,match mentors with mentees 

Independent scientists has one and a lot of people are interested in AI rather than any subfields- I think this shows a skill gap in understanding.

bounties for some easier projects/things companies might want, and research projects.

I'm willing to help direct funding, but I think others more skilled than me should do that.

https://manifund.org/projects/openscience-projects-to-get-non-academics-upskilled-and-involved-in-research-u8cgglp1jc

The discord spreadsheet

sheet

https://docs.google.com/spreadsheets/d/1DlBT1pF8-zMECntRWXFsL46gZyvNp1BJlJ6LXGze4dA/edit?gid=0#gid=0

if you're into deeper alignment/policy:  https://www.aisafety.com/

https://eleuther-ai-project-soar.github.io/ site view- might be wonky/outdated, will be changed soon, its a static site so submissions don't work, would need to contact me to add stuff to sheet.

this is incredibly high impact from what I can tell.

On funding:

It's weird- most groups won't fund what im doing as it's not fully alignment oriented, nor is it really opensci oriented (at least in the traditional sense). If you have any links/contacts to funding do let me know! or put a good word in for me.

here seems interesting but won't apply to me https://maps.foundationcenter.org/#/map/?subjects=all&popgroups=all&years=all

If you want funding- https://80000hours.org/2025/01/it-looks-like-there-are-some-good-funding-opportunities-in-ai-safety-right-now/ or https://www.aisafety.com/funders 

EDIT: forgot to add: https://www.researchhub.com/ (isn't just funding)

https://bagel.fund/ 

https://experiment.com/

Also I  haven't applied to any grant funding yet- please if you know more about these stuff do reach out; I wanted to wait until this writeup and sheet is out before applying; so I will probably be applying soon,especially if the manifund doesn't cover everything.

Suggested reading

How to do research:

https://www.lesswrong.com/posts/kpmaEevZ2KehZo2tp/some-advice-on-independent-research

https://irregular-rhomboid.github.io/2025/08/15/hitchhikers-guide-to-research.html

https://www.alignmentforum.org/s/5GT3yoYM9gRmMEKqL

https://colah.github.io/notes/taste/

https://gwern.net/research-criticism

https://www.lesswrong.com/posts/jP9KDyMkchuv6tHwm/how-to-become-a-mechanistic-interpretability-researcher 

Papers:

https://ar5iv.labs.arxiv.org/html/2210.06413

https://arxiv.org/abs/2210.06413

Doesn't seem like much other papers talked about this.

Might be some conferences on cultivating open source AI research etc?

Other:

https://andymatuschak.org/stillness/ 

self control esque stuff might be nice but its difficult and rarely (less than 10%) of the time works.

Wasn't sure where to put this, but if you feel like reading is boring,try reading through more obvious sections in a way where you're reading the words quicker than your voice is able to reason them, I think it works best for articles you have an idea of what its talking about, really helps me with stuff i have some idea in what they're doing and feels a lot less boring.

Scientists occasionally get managers to help them with motivation/procrastination problems. Kind of interesting to see the exact same issues (motivation,procrastination,burnout,etc) plaguing nigh everyone.

Scientific freedom Donald Braben seems like an interesting book.Will probably read it soon.

Memes and stuff

https://imgur.com/a/opensci-memes-yqvfZpQ

https://imgur.com/a/opensci-memes-yqvfZpQ

this is what I meant by using memes to run a psyop etc;  these memes are funny and informative (and may be anti phd which is not a good thing...) but something's missing-maybe it's cause i havent related it to any cultures?.  

Also i made a claude artifact of the main stuff discussed here- i don't like it. 

https://claude.ai/public/artifacts/c6a85920-6304-44c7-8b9a-93857b7fd830

link of all the memes i made here: https://imgur.com/a/opensci-memes-yqvfZpQ

Special thanks

Neel Nanda for data on his team's skill level, Genetyx for help with various stuff (and putting together the SOAR program which directly shaped some of my views),Madeline of Cohere research for their views on community building,Quentin Anthony of Eleuther for views on research managers, Afiq of casualphysicsenjoyer/independent scientist for views on open physics,dust for views on whats going on in biology hackerspaces , and various others who dm'd me over the past month or three about this writeup.

Thanks for reading this post! feel free to leave a comment or message me etc; my dm's are open.I am once again,looking for funding to do this fulltime.