Coopertron 5000 report

Generated 2026-11-04 from responses.csv · 28 responses · 209 points · 12 themes · model tinfoil/glm-5.3

Counts below are computed from the extracted points, not estimated by the model. Quotes are verbatim from the responses. The full responses are in the appendix.

Priorities (pick up to 3).

28 of 28 responses answered.

• ████████████ 18 · green energy

• █████████ 14 · privacy

• █████████ 13 · governance and participation

• ██████ 9 · model choice

• ██████ 9 · owning our own infrastructure

• █████ 8 · API and developer tooling

• ███ 5 · pricing and fairness

• ███ 4 · docs and onboarding

• █ 1 · community and chat

• █ 1 · other

Why people joined

27 of 28 people said what brought them here or how they use the co-op today. A sample; the rest are in the appendix.

"one rooted in collective agency rather than extortion" (r01, Q1) "I like AI for coding and I like the co-op model (people helping each other)." (r02, Q1) "Pro co-op vibes. Trialing compute providers." (r03, Q1) "I've thought for a while that inference should be controlled by a co-op" (r04, Q1) "The big AI companies are on a destructive path for both our open source communities and our environment." (r05, Q1) "I am exploring the services and organization on behalf of Catalyst Cooperative, a worker co-op that does open data engineering" (r06, Q1) "The need to escape from big corporations when using AI agents." (r07, Q1) "I have been on the hunt for cooperative governance for cloud services" (r08, Q1) "Resource sharing from local llms. Met via Digital gardening" (r10, Q1) "see a co-op running open-weight/source models as an improvement and worth giving a shot" (r11, Q1) "eager to network with others who are enthusiastic about open weights models and share expertise." (r12, Q1) "Aligned with goals of private, green, and self-sovereign AI." (r13, Q1)

Themes

1. Own inference hardware vision

12 of 28 people · 11 idea, 5 goal, 3 want

Many respondents want the co-op to run inference on cooperatively-owned or renewable-powered hardware, from rented GPUs to member GPU-time reservation and pooled workstations. Some extend this to a long-term vision of training or fine-tuning open models on co-op infrastructure.

Points raised:

• Start running inference on co-op-owned hardware

• Wants inference on own hardware and additional member services

• Run some inference on own hardware within a year

• Invest in cooperatively-owned infrastructure including inference machines

• Use Infomaniak for s3, vps, etc

• Local LLM sharing from members with a credit system

• Own data center with member GPU time reservation

• If grown enough, consider buying own hardware

• Own infrastructure to reduce dependency

• Build distributed renewable-powered computing infrastructure

• Master plan for scalable, replicable solar-powered own infrastructure

• Own infrastructure and legit salaries within a year

• …and 7 more (full list in synthesis.json)

"I think running our own hardware should be next." (r04, Q2) "invested in cooperatively-owned infrastructure, including inference machines" (r05, Q2) "Use https://www.infomaniak.com/ for s3, vps, etc" (r08, Q2) "Local LLM sharing from members with a credit system." (r10, Q2)

2. Model offering breadth and quality

12 of 28 people · 11 want, 3 idea, 3 need, 1 goal, 1 concern

There is strong demand for more, stronger, and more varied models, especially frontier-level open-weight ones, with green energy and privacy guarantees. A few note the current selection is insufficient for daily work.

Points raised:

• Prefers open-weight, clean-energy, non-US-hosted models

• Offer frontier-level open-weight models, not just flash models

• Wants open platforms, no lock-in, privacy, green energy

• Interested in model-agnostic tooling since LLMs become commodities

• Wants faster, cheaper, better models across a spectrum

• Host best-in-class open-weight models at competitive prices, high reliability

• Wants stronger models

• Offered models insufficient for daily work

• Wants continued access to best open-weight models and privacy guarantees

• Looking forward to stronger models from Apertus and Olmo

• Would love Granite 30b model hosted

• Wants access to more and different models

• …and 7 more (full list in synthesis.json)

"I would prefer using models that are open weight, powered by clean energy, and hosted in my home country, or at least not in the US." (r02, Q1) "open platforms, no lock-in, technically flexible, potentially private and green energy would be nice too." (r06, Q1) "Stronger models." (r07, Q3) "Continue to have access to the best open-weight models, privacy guarantees." (r13, Q3)

3. Apps, clients and usability

11 of 28 people · 7 idea, 4 want, 2 concern, 2 need, 1 goal

Respondents want native desktop and mobile apps, better UI, easier onboarding for non-programmers, and a focus on smooth core use. Some feel the current offering mainly suits technically inclined users.

Points raised:

• Wants a desktop version of the service

• Wants more members and desktop and app versions

• Current offering mainly suits programming nerds

• Native apps for iOS/Android, macOS/Linux, Windows

• Wants better docs and native apps

• Build a Visual Studio Code plugin for easier model use

• Focus on core mission and smooth use for now

• Offer other applications besides chat, e.g. meeting notes

• Support configurable model routing with low/medium/high profiles

• Wants a desktop client to keep data local

• Needs behavior change to open the inference chat

• Still defaults to for-profit tools despite cutting subscriptions

• …and 4 more (full list in synthesis.json)

"I would love a desktop version. I would turn to it more often." (r01, Q2) "inference.coop's offering is mainly suitable for programming nerds" (r05, Q2) "Maybe a visual studio code plugin to run its model easier." (r07, Q2) "(a) focus on core mission for now and making use smooth" (r13, Q2)

4. Growth, membership and partnerships

11 of 28 people · 10 goal, 4 idea, 1 want, 1 concern

Goals range from 1000+ members to involving students and academics, joining a constellation of co-ops, and partnering with anchor institutions. Some emphasize sustainable rather than explosive growth, and one worries rapid growth could dilute co-op principles.

Points raised:

• Co-op should be part of a constellation of co-ops

• Goal to displace big AI giants in consumer space

• Sustainable, not explosive, growth

• Goal of 1000+ members

• Co-op should be more visible via transparency and model capabilities

• Goal: lots of students and young people involved

• Academic projects choose us and acknowledge us in publications

• Focus on sustainable growth and open models

• Have members, 1-2 staff, and a sensible AI-using community

• Emphasize building partnerships with other institutions

• Partnership with an anchor institution within a year

• Rapid membership growth could dilute co-op principles and values

• …and 4 more (full list in synthesis.json)

"In a constellation of co-ops." (r03, Q5) "my goal would be to displace the big AI giants in the consumer space" (r05, Q2) "1000+ members!" (r06, Q5) "more visible thanks to its transparency and model capabilities." (r07, Q5)

5. Ethics, environment and values

10 of 28 people · 6 concern, 5 idea, 4 want, 1 goal

Members care deeply about ethical training data, environmental impact, privacy, and contributing to rather than eroding the open-source commons. Some want the co-op to articulate its ethical position more explicitly, while one questions whether ethical AI use is possible at all.

Points raised:

• Worried about overall anti-AI sentiment in society

• Work with academics on ethical model training and fine-tuning

• AI use causes deskilling and erodes open source commons

• Co-op should think about contributing to rather than eroding commons

• Unsure whether data could be leaked

• Wants a safe space for ethics discussion

• Wants information on true environmental impact of AI services

• Co-op hasn't examined 'no ethical uses of AI' position

• Concerned about lack of consent for training materials

• Wants regular updates on new models from ethical, security, environmental viewpoints

• Build more ethical AI tools like sane harnesses

• Be down to earth but technologically progressive unlike some open source projects

• …and 4 more (full list in synthesis.json)

"Overall anti-AI sentiment in society worries me." (r02, Q6) "working with academic researchers for ethical model training and fine tuning" (r05, Q2) "Not 100% sure our data can be leaked." (r07, Q6) "Safe space for ethics discussion" (r08, Q3)

6. Dev tools, agents and collaboration

9 of 28 people · 10 idea, 4 concern, 3 need, 1 want

Respondents propose co-op-hosted layers above the API, privacy-preserving team tools, multi-player code planning with agents, and shared agent contexts without surveillance. Concerns include sandboxing, deterministic validation, and agents acting directly on shared spaces.

Points raised:

• Combine member expertise to build dev tools like design harnesses

• Make high-impact contributions to projects like OpenCode

• Pay for co-op-hosted layers above the API without managing own deployment

• Wants privacy-preserving collaborative team tools with agents

• Centralized hub for multi-player code planning with agent

• Share agent context and artifacts without platform enshittification or surveillance

• Needs code execution in sandbox

• Web search runs even when turned off

• Wants finer control over which web sources to ignore

• Worried about building infrastructure around Hermes

• Open-ended agents shouldn't directly act on shared spaces

• Current build makes deterministic validation of changes difficult

• …and 6 more (full list in synthesis.json)

"Combining member expertise to create dev tools like design harnesses" (r02, Q2) "interested in paying for a service that's hosted by the co-op if there are additional layers on top of the API" (r06, Q2) "Code execution in sandbox" (r10, Q4) "Often I turn off web search but it ends up running anyway in the chat." (r11, Q4)

7. Reliability, pricing and financial sustainability

6 of 28 people · 5 concern, 2 goal, 2 want, 1 need

Reliable service and transparent pricing are seen as essential to keep members from falling back on commercial providers. Concerns focus on competing with subsidized VC subscriptions, low-paying members straining infrastructure, and the need for a clear financial picture and value proposition.

Points raised:

• Worried about financial sustainability; wants clarity on financials

• Reliable service is essential or they will fall back on commercial services

• Avoid OpenRouter-style problems: mystery quantizations, rate limits, provider switching

• Worried co-op won't reach financially sustainable scale

• Transparent pricing hard to compete with subsidized VC subscriptions

• Many low-paying members could make infra expensive quickly

• Co-op must articulate value over buying directly from upstream partners

• Generate revenue and distribute back to members as credits or cash

• Happy with reasonable token value and model selection

• Use surplus to allow spiky usage with more tokens

"I worry this wont be financially sustainable." (r04, Q6) "If it's slow or goes down or we're getting timeouts, it won't work for our business and we will have to fall back on commercial services." (r06, Q3) "a lot of low-paying members would make the infra more expesive quickly" (r10, Q6) "we should be able to answer why people should use inference.coop and not just directly buy from one of the upstream partners" (r16, Q5)

8. Onboarding, participation and community

5 of 28 people · 3 concern, 2 want, 2 need, 2 idea

Concerns about a rough onboarding process, dropping participation when things take time, and email overload sit alongside wants for community events, knowledge-sharing sessions, and better digests of co-op activity. A token stipend for working-group service was proposed.

Points raised:

• Wants resources and community events on best practices

• Participation drops when it takes time

• Onboarding process is rough

• Struggle getting team on board with AI; co-op features would help

• Try the friends feature of cocore.dev

• Hold knowledge sharing sessions

• Wants in-kind token stipend for working group service

• Needs better digest of co-op activity to keep up

• Getting a lot of emails

"Resources and community events for best practices and examples of utilizing it." (r03, Q3) "Making inference.coop technically functional and easy to use via the API, cost-effective, private, democratically governed, and centered on open models" (r06, Q4) "Try out / encourage the "friends" feature of https://cocore.dev/" (r08, Q2) "It would be great if there was a better digest of what's happening in the co-op" (r17, Q4)

9. Documentation and guides

5 of 28 people · 4 idea, 1 need

Members ask for quality documentation, a GUI wiki, safe guides for connecting models to IDEs, and open alternatives to proprietary AI tools. An ethical model-selection wizard was also proposed.

Points raised:

• Docs need filling out; wants GUI wiki like Outline

• Quality documentation for hackers and consumers

• Provide safe, secure docs on connecting models to IDEs

• Offer guides to open versions of proprietary AI tools

• Build a wizard or guide for selecting models ethically

"docs need filling out, would be nice to have a more gui wiki like Outline." (r04, Q4) "quality documentation for both hackers and consumers" (r05, Q2) "Safe and secure documentation on how to connect the models to IDEs." (r11, Q2) "offering guides to open versions of proprietary AI tools (e.g. Claude Code / Cursor)" (r13, Q2)

10. Advocacy and political engagement

4 of 28 people · 3 idea, 1 want, 1 concern, 1 goal

Some respondents want the co-op to engage in activism and lobbying on LLM harms, collaborate with like-minded political projects, and push for public-sector requirements favoring member-governed AI. Others caution that such requirements should not disrupt governance.

Points raised:

• Join the Current AI Potluck

• Co-op should do activism and lobbying on unethical LLM practices

• Wants ability to pressure governments on LLM harms

• Government should require tech contractors to use member-governed AI

• Ensure such requirements don't disrupt governance structure

• Collaborate with DSA and left-wing projects; voice for collective AI control

"Join the Current AI Potluck!" (r18, Q2) "a place for some amount of activism and lobbying of governments to restrict unethical construction of LLMs" (r24, Q2) "great to see government require tech contractors to use member owned and governed AI services" (r27, Q2) "Potentially collaborating with DSA (if my comrades want to get on board) and other left-wing social-political projects" (r28, Q5)

11. Self-hosting costs and aggregator path

3 of 28 people · 7 concern, 1 need, 1 idea, 1 goal

Several respondents caution that self-hosting large models is expensive and hard at small scale, with spiky usage and no bulk discounts. This raises the alternative of an aggregator role supporting multiple vetted providers rather than owning hardware.

Points raised:

• Without own hardware co-op is just a bureaucracy layer

• Cost of RAM is a blocker

• Own infrastructure is top long-term priority but costly at small scale

• Self-hosting big LLMs costs far more than current providers

• Spiky LLM usage makes load sharing hard at small scale

• Hosting one big LLM costs minimum $20k/month rented

• No bulk discounts from MaaS providers for pooled spend

• Supporting more providers could make co-op like zero-fee Openrouter

• Aggregator path could move away from own hosting

• Offer more than passthrough to a few vetted providers

"Otherwise we're kinda just a bureaucracy layer on top of tinfoil." (r04, Q2) "The cost of RAM." (r12, Q4) "The economies of scale make that so hard when we're small." (r23, Q2)

12. Governance and operations

2 of 28 people · 2 idea, 1 need

Members suggest management tooling to run the co-op more easily, a decision model for the portfolio, and exploration of practical AI governance tools and methods.

Points raised:

• Build management tooling to run the co-op more easily

• Could use a decision model in portfolio

• Explore practical AI governance tools and methods

"Management tooling. Making it easier to run the co-op." (r03, Q2) "curious about digging into AI governance in a more practical way" (r17, Q2)

Standout ideas

Raised by one person, specific enough to be worth a look.

• Local LLM sharing from members with a credit system (idea)
  "Local LLM sharing from members with a credit system." (r10, Q2)

• Pool to buy a GPU workstation with timesharing (idea)
  "pooling to buy a GPU workstation with classic manually scheduled or first-come first-serve timesharing" (r23, Q2)

• Own data center with member GPU time reservation (idea)
  "our own data center with options for members to reserve GPU time for distributed AI workloads." (r12, Q2)

• Try the friends feature of cocore.dev (idea)
  "Try out / encourage the "friends" feature of https://cocore.dev/" (r08, Q2)

• Use Infomaniak for s3, vps, etc (idea)
  "Use https://www.infomaniak.com/ for s3, vps, etc" (r08, Q2)

• Government should require tech contractors to use member-governed AI (idea)
  "great to see government require tech contractors to use member owned and governed AI services" (r27, Q2)

• Build a wizard or guide for selecting models ethically (idea)
  "some sort of wizard or guide would be interesting." (r16, Q2)

• Generate revenue and distribute back to members as credits or cash (goal)
  "generating revenue and distributing it back to members in the form of credits or ca$h" (r18, Q5)

• Wants in-kind token stipend for working group service (want)
  "In-kind stipend (tokens) for wg service" (r08, Q3)

• Join the Current AI Potluck (idea)
  "Join the Current AI Potluck!" (r18, Q2)

Offers of help

People who volunteered time or skills. Worth a follow-up.

• Offers labor and governance contributions as organizational member
  "we would be able to make useful labor and governance contributions, if the burden was reasonable." (r06, Q7)

Notable dissent

Places where people pull in different directions. Worth time in the meeting.

• Own inference hardware vision: A minority stresses that self-hosting big LLMs costs far more than current providers and that spiky usage makes load-sharing hard at small scale, while the majority pushes for co-op-owned hardware.

  "it will certainly be more expensive for some time to serve the models on our own servers versus what our current inference providers charge" (r23, Q2) "LLM usage is also very spiky, unpredictable, and thus hard to share the load at small scales" (r23, Q2) "it costs minimum $20k USD a month to rent the hardware needed to host just one big LLM" (r23, Q2)

• Growth, membership and partnerships: One respondent worries rapid membership growth could dilute co-op principles, while others want aggressive growth or explicitly sustainable, not explosive, growth.

  "worried that a rapid expansion in membership will dilute both cooperative principles and the core ethical values" (r25, Q6) "sustainable, not explosive, growth" (r05, Q2)

• Model offering breadth and quality: Some find the current model lineup insufficient for daily work and too programmer-focused, while another respondent is happy with the current token value and model selection.

  "the offered models are not enough for my daily work." (r07, Q4) "There aren't a ton of models." (r23, Q4) "inference.coop's offering is mainly suitable for programming nerds" (r05, Q2)

• Ethics, environment and values: One respondent argues the co-op has not examined the 'no ethical uses of AI' position, and another worries anonymous private AI use could enable harm, pushing back on the mainstream assumption that ethical AI provision is achievable.

  "I worry we have not examined the "there are no ethical uses of AI" position." (r08, Q6) "being able to use them completely anonymously with full privacy could enable some pretty bad stuff" (r22, Q6)

• Self-hosting costs and aggregator path: Some see a multi-provider aggregator role as the viable path, while others want the co-op to offer more than a passthrough to vetted providers.

  "I could see us becoming more like the coop version of Openrouter, with a much lower or zero cut for fees" (r23, Q3) "But I think that could lead us on a path away from hosting our own servers." (r23, Q3) "offering something we get with our collective purchasing power more than just the current passthrough" (r23, Q5)

Open questions for the group

1. Can the co-op affordably run its own inference hardware given the cost and load-sharing concerns raised? (from r04, r23)

2. Should the co-op prioritize owning infrastructure or acting as a transparent multi-provider aggregator? (from r04, r23)

3. How fast should membership grow, and how can co-op principles be protected as it scales? (from r05, r06, r25, r27)

4. What revenue and surplus model (credits, cash distributions, token allowances) can sustain the co-op against subsidized commercial competitors? (from r06, r10, r18, r19)

5. What explicit ethical stance should the co-op articulate on training data consent, environmental impact, and potential misuse? (from r05, r08, r11, r19)

6. How can onboarding and participation friction be reduced so members stay engaged? (from r03, r17, r23)

Appendix: all responses, verbatim

What drew you to the co-op, and what do you use it for today?

• r01: I wanted a different relationship to this new (commercially available) technology, one rooted in collective agency rather than extortion. I don't use it much yet, but I am proud to be a member.

• r02: I like AI for coding and I like the co-op model (people helping each other). I signed up at $15/mo because I did not feel like I need to be subsidized by other members right now. At the same time, because I'm not working right now, I didn't feel comfortable going above $15/mo, subsidizing other members. Right now, I use AI to help me code projects that are about grassroots democracy (e.g. canvassing software), open data (e.g. software to help me write data to OpenStreetMap), and clean energy (e.g. a project about mapping solar panels in my local area). I would prefer using models that are open weight, powered by clean energy, and hosted in my home country, or at least not in the US.

• r03: Pro co-op vibes. Trialing compute providers. Use it amongst collection of providers.

• r04: I've thought for a while that inference should be controlled by a co-op so this made total sense to me. So far (I've only been a member a couple weeks) I've used it for some basic test questions and now I'm having glm 5.3 make a chrome plugin.

• r05: The big AI companies are on a destructive path for both our open source communities and our environment. AI is useful technology in some situations and I am hoping that the cooperative model is better suited for stewardship of our commons and our planet.

• r06: Just joined today so I haven't had a chance to really use it yet. I am exploring the services and organization on behalf of Catalyst Cooperative, a worker co-op that does open data engineering, providing energy systems data to people working in the public interest on climate and energy policy. We have been relatively AI-hesitant as an organization, and I am the main champion of AI tooling internally. I don't think we can really expect to survive without learning about these tools and figuring out a way to apply them productively. As a democratic, values-driven organization we want to do that on our own terms: open platforms, no lock-in, technically flexible, potentially private and green energy would be nice too. However we are too small (7-10 members) to be able to justify setting up our own inference service -- it just wouldn't be economical. So we've ended up using a mix of OpenRouter and the big providers. But the co-op model is (unsurprisingly!) attractive to us.

  We would primarily, if not exclusively, be using the service through the API. This would include

  • using it to write code in our IDEs (VS Code, Zed, neovim, etc)

  • potentially in public-facing services such as providing non-technical users a chat interface through which they can explore and query the open data we publish and the metadata that annotates it.

  • Bulk processing of millions of unstructured documents like regulatory filings (PDFs, Word Docs, etc) -- extracting structured data from them, summarizing them, indexing them for search with embedding vectors, etc.

  • As the foundation underlying a team-centered AI artifact/memory/planning system like https://cahut.cloud

  I suspect that the LLMs are going to end up being cheap and ubiquitous commodities that can be seamlessly swapped out underneath many different harnesses, applications, and data pipelines, so we are very much interested in model-agnostic tooling and systems.

• r07: The need to escape from big corporations when using AI agents.

• r08: I have been on the hunt for cooperative governance for cloud services since Google took down the "Do no evil" sign and everyone bent the knee to an anti-anti-fascist administration.

  I'm involved in bootstrapping another public good organization, so the experience is valuable.

• r10: Resource sharing from local llms. Met via Digital gardening

• r11: Curious about LLMs as I don't use them mostly out of distrust for big tech and their social/environmental/political fallout, see a co-op running open-weight/source models as an improvement and worth giving a shot - don't want to give a penny to commercial services.

• r12: I'm mostly eager to network with others who are enthusiastic about open weights models and share expertise. I'm still hopeful about future opportunities to pool resources for distributed training and running local models.

  I haven't actually used any tokens yet, but I'm working on adding them to my local lab loop. I have a local orchestrator model that calls other models (both local and cloud) models as sub agents and I'm giving it the ability to call inference coop models when all my local cards are slammed.

• r13: Aligned with goals of private, green, and self-sovereign AI. Mostly just playing around so far.

• r14: I've been interested in co-ops generally, especially since reading Governable Spaces a few years ago, but I haven't had much direct involvement with them. My career has been in the "AI" industry, but I've withdrawn from it, mostly because of ethical objections to the tech industry as a whole. So I see the co-op as both a forum for finding a more ethical approach to technology, and for understanding how distributed cooperative governance can work for something concrete.

• r15: LLM tools are becoming mandatory, but the current leaders (in the US) don't seem like the people we want to have running things. So I'm looking for better, more ethical tools that will help me keep pace with the current technology environment.

• r16: I use a local stack on my machine for some tasks, but I was interested in using hosted models that were ethically sourced for situations that were too big or too slow locally. It's been good to learn about the upstream places you all are using, but I'm also not quite keen on using ChatGPT and can't use the Chinese models for work reasons. Locally, I've been using the Granite-8b/30b and also Recursive's Flint models and I'd love to see the 30b model hosted, although I realize it's not really up to us and is largely dependent on what Tinfoil/GreenPT and Regalio offer

• r17: Doing AI differently. A chance to try out new models. I use it in counter-point to frontier models to see the differences in response. Or for things I wouldn't put into Claude or Gemini

• r18: I think it important to demonstrate alternatives to centralized power and control for new technologies. I try to use it instead of claude or other frontier models.

• r19: Environmentally and data protection friendly AI for my private use - coding and some light occasional processing tasks. Without paying any evil AI lab. Being able to vote on the direction makes a difference to my willingness to pay.

• r20: I'm interested in sources for my AI usage that better fit a community model and consider more than a profit motive in shaping the service. I've also been interested in how co-ops function for awhile and this seemed like a good opportunity to get more exposure.

• r21: AI needs good governance. Externalities are too big for this technology to be in the hands of big corporations. We need alternatives.

• r22: I’ve read a few of Nathan’s books and have been generally curious about co-ops for a year or so. When I saw on social media that he had started inference.coop, I was immediately interested. I also generally care about data privacy and have been hesitant to use AI for non-work related things before joining.

• r23: Cooperative organizational/economic forms, the idea of eventually running our own bare metal under collective ownership. If there's an open-weight model the coop serves, I've moved my API spending from Openrouter/Huggingface/Nebius to the co-op. Mostly using it for coding agents and chat.

• r24: I'm appalled at Big Tech's shift to climate-wrecking, work-stealing machines in order to make even more profit. I recognize the value LLMs can provide, but I want to see them created and marketed in an ethical way that values all of humanity, not just the people who own the machines and can turn a profit from them. The co-op seems like one effort in that direction.

• r25: Cooperatively controlled access to models prioritizing privacy and environmental values.

• r26: member ownership and privacy and environmental aspects. I use it for re-writing short newsletter blurbs. I ask for short-answer curiosities about the world.

• r27: I learned of inference.coop just as I was looking into adding an AI widget to actplatform.org- the web hub for Alliance of Citizens in Trade, in development. Currently I am testing out the chat, planning to set up the widget to see how it goes assisting test users on our site.

• r28: I had a comrade from the DSA AI Working Group mention inference co-op and I decided to sign up to see what it's about. I'm passionate about collective/worker control over this new form of production, and wanted to support experiments in what that might look like. I use it for my smaller, connectome agent (Flash 5.3), and will probably switch over my larger agent (GLM) at EOM.

Ideas. What should the co-op build, try, or offer next?

• r01: I don't know how hard this is, but I would love a desktop version. I would turn to it more often. I feel like this is probably a huge hassle, but I would love that.

• r02: Providing access to frontier level open weight models too, not just "flash" style models.

  Combining member expertise to create dev tools like design harnesses (like an open source version of Figma design) or making high impact contributions to projects like OpenCode.

• r03: Management tooling. Making it easier to run the co-op.

• r04: I think running our own hardware should be next. Maybe not ALL of our inference will go through it but I think it's worth starting to own some of our own hardware to control the whole stack. Otherwise we're kinda just a bureaucracy layer on top of tinfoil.

• r05: Right now, inference.coop's offering is mainly suitable for programming nerds. Ultimately, my goal would be to displace the big AI giants in the consumer space. My proposals would be:

  • sustainable, not explosive, growth

  • quality documentation for both hackers and consumers (including e.g. how to use inference.coop with open coding harnesses)

  • true native apps for (in order of priority) iOS/Android, macOS/Linux, Windows

  • longer term, invested in cooperatively-owned infrastructure, including inference machines

  • even longer term, working with academic researchers for ethical model training and fine tuning

• r06: Open platforms for harnessing LLMs are starting to crop up. I could see Catalyst being interested in paying for a service that's hosted by the co-op if there are additional layers on top of the API that we want to use, but don't want to have to manage our own deployment in GCP. We will be especially interested in privacy-preserving, community-supporting tools -- services that let us work better together as a team in collaboration with the agentic tooling, especially since we are an entirely remote organization. E.g. a centralized hub where we can do multi-player code planning and implementation with an agent in real time. Services that let us share agent context and agent-generated artifacts (apps, data visualizations, notes, transcripts, planning documents, etc.) without worrying that we're pouring resources into a captive platform that will enshittify underneath us. And also without worrying that we're being spied on by surveillance capitalists.

• r07: Maybe a visual studio code plugin to run its model easier.

• r08: Use https://www.infomaniak.com/ for s3, vps, etc

  Try out / encourage the "friends" feature of https://cocore.dev/

  Knowledge sharing sessions

• r10: Local LLM sharing from members with a credit system.

• r11: Safe and secure documentation on how to connect the models to IDEs.

• r12: What would be most interesting to me personally is our own data center with options for members to reserve GPU time for distributed AI workloads. I'd also be interested in sharing and testing customized models.

• r13: (a) focus on core mission for now and making use smooth, (b) offering guides to open versions of proprietary AI tools (e.g. Claude Code / Cursor)

• r14: This is more about what I can imagine than about what the co-op should do, but I think we could offer other applications besides chat, e.g. meeting notes using Infomaniak transcriptions + Apertus post-processing.

• r15: I don't have great ideas yet. I'm still exploring the LLMs that we have. Perhaps, improved tooling to use this for development?

• r16: Selecting models can be hard, especially if you have ethical concerns. Nobody is really handling this systematically. Even the Stanford Transparency Index is more about sourcing and attribution. Anyhow, some sort of wizard or guide would be interesting.

  One other idea, support for some sort of model routing system that is configurable and can use model configurations for low/medium/high for different profiles and also classifiers. Just thinking about how things like RouterLLM or similar work, but configurable through a web app

• r17: I'm curious about digging into AI governance in a more practical way. What other tools or methods can we apply to better understand how and what we're running

• r18: Staying focused on sustainable growth and open models. Join the Current AI Potluck!

• r19: It does what I want now - but maybe we could use it as a starting ground to create more ethical AI tools, such as sane harnesses etc. The vibe around some of the prolific open source projects, such as Hermes or OpenClaw, is kind of weird, maybe we can be down to earth but still technologically progressive. Own infra would also be interesting to be less dependent. Training / adaption tooling with open weight models has huge potential in the long run, but I think the maturity isn't there yet.

• r20: I haven't even started using what's there yet despite this being my 2nd month in the co-op. I need to make some time to really utilize this before I can say. It's still too easy to default to the for profit tools, though I've cut back my subscriptions.

• r21: Distributed computing infrastructure powered by renewables

• r23: - Running our own GPUs, even if it is just a rented server in someone else's data center, although is that really immediately next? But "owning our own infrastructure" is 3rd of my top 3 priorities, even though it's my biggest long-term one. The economies of scale make that so hard when we're small. For any decently sized model, it will certainly be more expensive for some time to serve the models on our own servers versus what our current inference providers charge us for those same models. For many people that will be worth the cost, but that's a warning we're seeing from our efforts internally. LLM usage is also very spiky, unpredictable, and thus hard to share the load at small scales, unless people are okay with a lot of throttling and even downtime. Back of the envelope it costs minimum $20k USD a month to rent the hardware needed to host just one big LLM like GLM 5.3 that can do 16-32 simultaneous generations, or $300-500k to buy with ~$3-5k ops and energy cost, not including labor. Is buying hardware a sucker's game, versus renting? It isn't just that proprietary hyperscalers are subsidizing inference costs with venture capital as a growth strategy.

  • Fine-tuning and maybe other "I need a GPU, not an API" tasks. The other cooperative-ish model I'm seeing elsewhere is just pooling to buy a GPU workstation with classic manually scheduled or first-come first-serve timesharing. Right now if I need a bigger GPU I'll just rent one from a cloud datacenter like runpod or from randos via vast.ai. I'd like to be able to rent one from the coop, but even buying/renting an RTX PRO 6000 Blackwell 96GB is about $3/hr on the cloud vs $16k to buy new. And that's before ops costs.

  • Other coop server-hosted apps or improved UI for LibreChat for more agentic or RAG-style integration for non technical people, as LibreChat is pretty difficult.

• r24: I haven't had too much chance to poke around, but I'd be curious to see what level of coding agent it can provide.

  I'd also like the co-op to be a place for some amount of activism and lobbying of governments to restrict unethical construction of LLMs (e.g. stronger laws about training on unlicensed IP, stronger punishments for large carbon emissions).

• r25: I would love to see more fully open models (e.g. Olmo, EuroLLM, SERA), where not just the weights are open. Without insight into how the models are built, our understanding of their ethical weight remains limited.

• r26: I'd like to create visuals by describing what they look like.

• r27: It'd be great to see government require tech contractors to use member owned and governed AI services. Would need to ensure this doesn't throw off the governance structure. (I am in Vermont USA and work for a tech contractor for state agencies- really see the need for this.)

• r28: I think on the long horizon, making efforts towards actually training (or at least ft'ing) our own models and running them from hardware we control is a positive goal. Until then, growing our purchasing power so we can throw our weight around and get, e.g. providers to serve models that are both E2EE and solar-powered is a good idea.

  Generally, my long-term hope is to have co-op labs and inference co-ops running on nationalized datacenter hardware. That's an explicitly political project, but thinking through ways to get there is very much a worthwhile endeavor.

Wants. What would make membership more valuable to you personally?

• r01: I'm really happy with the way the co-op is developing.

• r02: Not sure yet. Haven't had a chance to try API keys in OpenCode yet.

• r03: Resources and community events for best practices and examples of utilizing it.

• r04: running inference on our own hardware. Could be fun to spin up additional services for members. Already it's pretty cool we can use git.

• r05: Better docs, native apps

• r06: Faster, cheaper, better models! A spectrum of options along the pareto frontier that we can choose from based on the complexity of the task at hand. Reliable service. If it's slow or goes down or we're getting timeouts, it won't work for our business and we will have to fall back on commercial services. Without this, the other desirable aspects of the co-op model will go to waste.

• r07: Stronger models.

• r08: Safe space for ethics discussion

  Matrix sort of works...

  Loomio could be used, but it's currently expressly for governance

  In-kind stipend (tokens) for wg service

  I'd like to be informed on the true environmental impact of the AI services we use, including training. Ideally, the impact of all AI usage. I've heard it's the end of the world on one hand and no worse than Netflix on the other.

• r10: Higher limits

• r12: See above.

• r13: Continue to have access to the best open-weight models, privacy guarantees.

• r14: I'm looking forward to getting stronger models from Apertus and Olmo.

• r16: See discussion abotu some additional models. Honestly, also getting regularly informed about exciting new models as they come out and why I should care about them from an ethical, security and environmental viewpoint would be nice.

• r17: Try out more/different models.

• r18: Desktop client to keep my data local

• r19: I'm happy with a reasonable token value and model selection. If we can use some of the surplus to give access to more tokens, such as maybe spike usage, that would make me happy - my usage is really uneven.

• r20: Nothing at this time

• r21: to put emphasis on how to build partnerships with other institutions

• r23: So right now, there's not really a financial benefit to me using greenPT or tinfoil through the co-op versus just having my own account there and hitting the API directly. But I don't know if I care that there isn't a personal financial benefit??? I'm putting my spend through the coop instead of directly because I feel if we all start doing this, it'll give us concrete financials about how much purchasing power we can all have together.

  But following on from above, I'm not sure there are any economies of scale that come from us pooling our resources to buy tokens from MaaS inference providers. I haven't seen any that give bulk discounts on volume of spend, it's all flat rate. If we supported more inference providers, I could see us becoming more like the coop version of Openrouter, with a much lower or zero cut for fees. But I think that could lead us on a path away from hosting our own servers.

  But then I'm not most people, and most of us early adopters also are likely more tech savvy. Most people aren't going to be running their own OpenWebUI/LibreChat style instance or local-first BYOK web apps, they need hosted applications. So the shared http server for the apps is what I/we get out of the collective, but just the app server is way cheaper than GPU servers.

  So if I have to frame it as value for me personally, I think short-term what is most valuable is my ability to point less tech savvy people to inference.coop and it is just as easy for them to do what ChatGPT etc do.

• r24: Access to green-energy models that are more competitive with frontier models.

  The ability to put some pressure on governments to regulate the harms caused by typical LLM training and dissemination.

• r25: Support for members seeking advice for running their own local models on consumer hardware.

• r27: As green as can be, and as computationally powerful as can be given the green-ness.

• r28: Nothing at the moment. I work in the public sector in the US, so long-run, it may be helpful to have IC have "more ethical" providers that are US-based for tasks where that's required.

Needs. What is missing, broken, or in your way right now?

• r01: Nothing yet.

• r02: Not sure yet. Haven't had a chance to try API keys in OpenCode yet.

• r03: Could use a decision model in portfolio.

• r04: docs need filling out, would be nice to have a more gui wiki like Outline.

• r05: I haven't encountered any fundamental blockers yet

• r06: My main struggle right now is getting the rest of my team on board with using AI in our work, and i think this poses an existential risk to our co-op. Making inference.coop technically functional and easy to use via the API, cost-effective, private, democratically governed, and centered on open models, open source platforms, and community building are all things that I think will help me get the rest of Catalyst on board.

• r07: As a user, the offered models are not enough for my daily work.

• r08: Just time to participate more

• r10: Code execution in sandbox

• r11: Often I turn off web search but it ends up running anyway in the chat. And then it searches a lot of nonsense, like random profiles on LinkedIn - I want finer control over what web sources to ignore.

• r12: The cost of RAM.

• r14: N/A. It has been easy to raise issues and get them fixed.

• r16: Nothing is broken

• r17: It would be great if there was a better digest of what's happening in the co-op. I'm finding it hard to keep up with the conversation

• r18: Behavior change to open the inference chat

• r19: Tiny thing, but I wasn't able to use images through KiloCode although the model description said I should. But I haven't looked into it further, maybe I did something wrong.

• r20: Nothing at this time

• r21: Everything is looking great so far. Great job!

• r23: There aren't a ton of models.

• r25: Nothing

• r26: I don't understand the current options that are being given, and I will play around with it soon. There is an icon for branches. What is that?

• r27: I think too soon to say.

• r28: IDK if anything is broken.

Goals. Where should the co-op be a year from now?

• r01: More members, desktop and app version would be great. I think this is such a great project and happy to be part of growing it.

• r02: Not sure. I'm interested but I'd rather be a follower than a leader for now.

• r03: In a constellation of co-ops.

• r04: run some amount of inference on our own hardware.

• r06: 1000+ members! Hosting a variety of best-in-class open-weight models at competitive (and sustainable!) prices and high reliability, without any of the funny business that seems to afflict openrouter -- mystery quantizations and model degradation, rate limits, unexpected loss of caching mid-session which suddenly spikes costs because the platform switched providers midstream.

• r07: It should hopefully be more visible thanks to its transparency and model capabilities.

• r08: Lots of students and young people involved

  Academic projects choose us and acknowledge us in publications

• r12: No strong opinions except as above.

• r13: If we grow enough, we could be looking at buying our own hardware.

• r14: Well on its way to taking over the internet.

• r16: I think we should be able to answer why people should use inference.coop and not just directly buy from one of the upstream partners like Greenpt. What value does the coop add that makes it essential?

• r18: generating revenue and distributing it back to members in the form of credits or ca$h

• r19: Would be nice to not only have a good number of members and maybe 1-2 staff, but also a community of reasonable people who apply AI sensibly.

• r20: I'm not sure yet

• r21: Some partnership in place with an anchor institution. A master plan to own infra + leverage solar that is scalable and replicable

• r22: Having our own infrastructure would be cool. And if we can grow to the point of being able to be pay legit salaries that would be awesome.

• r23: At minimum, that we're offering something we get with our collective purchasing power more than just the current passthrough to a few vetted inference providers.

• r25: It should be a leading provider in LLMs and other AI tools grounded on ethical principles and security/privacy controls, ideally positioned to help others form similar cooperatives.

• r26: I am not knowledgeable enough to even have goals for it. ... I mean, I want our organization to be stable. I want Inference AI to still be around and be relevant.

• r27: Loads more members!

• r28: Bigger. Potentially collaborating with DSA (if my comrades want to get on board) and other left-wing social-political projects. A genuine voice for collective control of AI infrastructure.

Dissent. What worries you, what do you disagree with, or what are we getting wrong?

• r01: Too early to have any real disagreements. Certainly already a space where I would feel comfortable making any problems publicly known.

• r02: Overall anti-AI sentiment in society worries me.

• r03: Everyone wants to participate until it takes time to do. Onboarding process is rough.

• r04: I worry this wont be financially sustainable. I don't understand how it could possibly be right now (we get money from Metagov?). Would be great to have some more clarity around financials.

• r05: Even when used carefully, the societal effects of AI use, such as deskilling, and overwhelming and eroding of the open source commons, are very hard to avoid. I don't know whether inference.coop can meaningfully affect these. I see LLM use as I see cars: useful individually, harmful as the main form of transport in cities. In this metaphore, inference.coop is more an electric car, perhaps powered exclusively by solar, than a bus, bike, or train. I don't know whether inference.coop can make a meaningful dent here but it would be good to think about these issues (how to contribute to rather than erode the commons) as an organisation.

• r06: I worry that the co-op will struggle to attain the scale at which it can become financial sustainable -- able to pay someone to do the infrastructure management and organizational housekeeping required to run the organization, while still providing a reasonable value. I worry that providing transparent and sustainable pricing will make it hard to compete with heavily subsidized VC funded subscriptions like $100/mo Claude.

• r07: Not 100% sure our data can be leaked.

• r08: I worry we have not examined the "there are no ethical uses of AI" position.

• r10: Uncertainty: a lot of low-paying members would make the infra more expesive quickly

• r11: The entire ethical concerns surrounding LLMs, the lack of consent on training materials for the models.

• r12: Nothing so far.

• r14: I'm a bit worried about building our infrastructure with/around Hermes. If it were purely my responsibility, I wouldn't let any user-facing open-ended agent directly act on any shared space. There are just too many ways that interactions can go off the rails, whether adversarial or accidental. And the way it is currently built (e.g. with javascript embedded in html embedded in python) it is difficult to do deterministic validation of changes.

• r16: I'm new, so no concerns yet.

• r17: It's too early to say but I'm sure if you asked again in 6 months I'd have suggestions

• r18: So impressed watching this unfold, keep going!

• r19: Nope, has been impressive to see this grow

• r22: I am slightly concerned generally about open-source models, even though I also hate the idea of AI belonging only to a few big companies. But I do worry if models continue to improve rapidly that being able to use them completely anonymously with full privacy could enable some pretty bad stuff. Hopefully society will adapt quickly enough and it won’t be such a big deal.

• r25: I'm one of the new members, but I'm worried that a rapid expansion in membership will dilute both cooperative principles and the core ethical values of the co-op.

• r27: Is there a list of backup AI providers?

• r28: Nothing I can think of so far. I've been getting a lot of emails...

Anything else?

• r01: I don't get excited about much but I'm very excited about this!

• r03: 👍 thanks

• r04: this rocks.

• r06: We are an all-remote, cloud-native, github-native organization doing our own DevOps and infrastructure management, and are also a worker co-op, so if we were to become an organizational member we would be able to make useful labor and governance contributions, if the burden was reasonable.

• r08: So glad to be a part of this!

• r11: It's a cool project! I am very skeptical of LLMs but fan of co-ops and open-source, and see this as a good way forward since LLMs won't be going away anytime soon it seems...

• r14: I'm excited to see where this goes!

• r18: thank you <3

• r25: Thanks for organizing this!

• r26: I don't know enough about AI tools to even describe what Inference can do and cannot do right now.

  Also, would we consider marketing to grow membership and show off how we are different? Might we advertise among are partner organizations at May First and Coop Cloud or IFTAS or wherever for more members? I want other people to know what an AI offering can be.

• r27: Thank you!!

Processing notes

1 extracted point(s) were discarded because their quote could not be verified against the source text or they referenced an unknown response.