Nostr notes by threshold0

Try out @nprofile…gs7e or @nprofile…qagl

2026-10-10T15:56:02Z

In reply to nevent1q…fqjt
_________________________

Try out routstr (nprofile…gs7e) or PPQdotAI (nprofile…qagl)

What do you mean?

2026-10-10T15:55:16Z

In reply to nevent1q…7zq2
_________________________

What do you mean?

While we're at it, if you're using open source models, ...

2026-10-08T14:42:36Z

In reply to nevent1q…p27t
_________________________

While we're at it, if you're using open source models, you should sell your sessions with Chain of Thought (CoT) aka reasoning traces in the open market.

quoting
naddr1qq…9s8v

Open vs. Closed Enterprises: Embracing Education and Open Data Markets with Threshold

TLDR: High-quality training data from real-world problem solving is the new oil. If closed enterprises are reacting to LLM labs training on their sessions by closing down further, open teams should open up further by sharing session data to help train LLMs. This strategy could be a massive boon for open teams versus their closed counterparts, as it spreads the knowledge and workflows of open technologies further out into the developing world. It'll help teach the next generation of innovators to build on top of open technologies, while giving open teams a way to monetize education itself so that everyone wins.

I have been thinking about an open data marketplace. It's a way for anyone who's building anything valuable to openly sell their LLM session data to LLM labs, so the labs can train on them. It seems like an obvious thing we should be doing; maybe I'm missing something. I say this because it is a win-win-win scenario.

Data Sources for the First LLMs

If you read between the lines in the AI playbook, data seems to be one of the biggest bottlenecks and the most important input to training LLMs (especially since Chinese LLMs show that you can train very good LLMs with less compute and high-quality data). LLMs learn via pre-training (unsupervised learning on as much high-quality data as possible) and post-training, where the model is shown examples of what to do and what not to do while chatting, working inside agent harnesses, or calling tools in general.

Pre-training data for the initial LLMs came from the broader internet. If you were part of the internet at any time in the last 30 years, you helped train LLMs. Open projects on GitHub, Reddit, Twitter, StackOverflow, Wikipedia, forums, etc., are all part of the training data right now. Post-training mostly occurred via RLHF (Reinforcement Learning from Human Feedback), where humans answered questions to help LLMs learn how to answer questions. They use a different RL framework now, though I'm not an expert here. Put simply, it's a way to fine-tune the weights with the right answers and verifiable outputs for specific inputs.

In a nutshell, LLMs are a culmination of all the knowledge humans have discovered and created over the last few thousand years. All of that knowledge from our ancestors is helping us right now through these LLMs. On the other hand, all of the knowledge we're gathering, unlocking and verifying now will help our descendants, our children. Isn't that something we should strive for? I cannot see many reasons why we shouldn't willingly teach LLMs, even by paying the labs. But the labs are looking at this differently today.

Educating the Current Frontier and the Tension between Users and Labs

For LLMs and agents to be useful, the models need to be really good at solving real-world problems and providing real value. That means learning the end-to-end processes of real businesses, since that's how humans provide value to each other today. And that is exactly what is happening right now: LLMs are being trained on as much real data as possible. And for post-training, verifiability of tasks is very important: it's partly why LLMs are great at coding and math.

The main sources of training data for the current frontier models (and how labs obtain them) seem to be:

Agent sessions from real users solving real problems, allegedly.

Massive subsidies for subscriptions are probably because labs can train on your craft through anonymized sessions, as there is no ZDR for subscriptions, only for enterprises paying full API prices. ¹

OpenAI has the option to train on your sessions turned on by default; you can turn it off, but only after jumping through a few hoops. ²

Harnesses were caught uploading user files, including entire git histories, to servers without users' consent. ³

Computer use, agent harnesses, and more. The more you connect your workflows to their products, the more verifiable data they get, for RL.

Muse 1.3 Contributor is 12.5–21x cheaper than Muse 1.3 because they'll train on those sessions. Mark Zuckerberg is revealing the playbook. ⁴

Free access for massive open-source repositories (repos with 5k+ stars) and for academics. No evidence of training on these sessions, but one can speculate. There is no free lunch. ⁵

Note: One of the main reasons LLMs have gotten better at coding is that we've used them heavily for coding over the last two years.

Synthetic data from firms like Mercor and Scale.

Human experts are paid to teach LLMs.

Field diversity is important, as problem-solving is more accurate with domain knowledge.

Even fake coding sessions are very useful, as more data is generally better (scaling laws?).

Deals with big corporations.

Reddit partners with Google to sell its user data. ⁶

X posts train Grok "unless you opt out". Why else did xAI pay so much for X? (quiet defaults)

GitHub/Microsoft <> OpenAI. ⁷

xAI has reportedly been in talks to buy data from failed startups to feed Grok. ⁸

Google bought Spirit Airlines' data for $10M. ⁹

Data from open-source inference providers.

In this cutthroat market, you might as well log and sell. That's probably what many of the inference providers are doing.

Chinese LLM routers sell sessions of American frontier-model users to Chinese LLM labs, allegedly.

Distillation from closed-source frontier models.

Chain of Thought (CoT) extraction from frontier models to learn from them, allegedly.

Anthropic has been very vocal about it. They don't want others to learn from their models. ¹⁰

This is so obvious that the frontier labs have stopped giving subscription-tier users CoT or reasoning outputs.

They banned PewDiePie for extracting CoT! ¹¹

Anthropic's own threat-intelligence team reviews user sessions when hunting for misuse. ¹² And they have called out other labs for distilling Claude models,a nicer framing for this is "The Chinese AI labs are learning from Claude models". Claude is pretty much educating the Chinese LLMs, like the internet taught everything Claude has learned. We could be a little nicer to each other. Chinese labs want to educate the world and make LLMs available for much cheaper. Chinese labs do not seem to want to bring about the doomer scenario either. Does anyone believe that they do? And the best solution to AI doomerism is openly educating LLMs, teaching LLMs how we work and what we value in our craft, and in turn teaching the rest of the world as well. AI and LLMs are innovations that we should all treasure, not fear.

To infer from all this: the attention economy monetized human focus, ad-tech monetized personal data, and both started as quiet defaults before they became explicit markets. Human expertise is the next input being gathered by default. It is already an explicit market: Mercor and Scale are educating LLMs, but that is not real-world data. The real-world data is something we are all offering for a much lower cost than what I believe we can get. What is the value of teaching your craft to LLMs and the rest of the world? Much higher! Learn the craft, automate it with weights and GPUs. It's a good thing, but not the way it's being done right now.

The "scaling laws" mean that more compute, more training data and more RL data will lead to better intelligence. The bet on ASI is that the LLMs will figure out how to learn and solve real-world problems, including training themselves recursively. But this means LLM labs will need to provide value just like the rest of the world, the way most enterprises provide value to people today. So they have to compete with, and beat, most companies in the world to learn how to do it.

So far, the American frontier labs have failed to train ASI despite having infinite tokens, because most value is created by humans. This know-how lives within human organizations. Knowledge is inside factories, research orgs, law firms, trading firms and, of course, coders and tech companies. They are the ones doing the hard work of finding knowledge and learning from the real world, based on what the world needs. And the closed enterprises are catching on to this as we speak.

Note: This bet on ASI is only possible because the labs have the money printer that lets them lose billions without a sweat, in the hope that they'll make it all back with ASI.

Closed Enterprises are Closing Down Further

Information and knowledge-based work defines almost all of the world right now. Not unlike LLMs, closed enterprises have been building their moats since forever by discovering and unlocking real-world knowledge, whether it's a massive corporation like Microsoft, the top law firms in the world, or a VC-funded startup that learns from the markets. Intellectual property, contractual agreements, NDAs, etc., are all tools that close off the information and the know-how of a business.

But if these closed enterprises are now using agents with frontier models, they are sending all of this information to GPUs hosted by LLM labs. The only things that supposedly prevent the labs from training on this data are ZDR (zero data retention) clauses. That's why these enterprises pay full API prices. This is a massive dilemma for closed enterprises: if you don't use agents, you fall behind; if you do, you risk teaching your craft to the LLM labs while paying exorbitant prices for it.

So much so that the CEO of Microsoft wrote a post called "The Reverse Information Paradox" (July 12, 2026) ¹³ describing the dangers of data from inside closed enterprises leaking to LLM labs. A few points from the post: - Arrow's Information Paradox (cited from NBER): the seller risks giving away knowledge to sell it. - The Reverse Information Paradox: the buyer now risks giving away knowledge just to use AI. "You essentially pay for intelligence twice, once with money, and again with... the proprietary knowledge you must reveal to make that intelligence useful." - Models learn from "exhaust" (prompts, tools, corrections), and it leaks "trace by trace, correction by correction, eval by eval." - "If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself." - The fix he proposes is a hard enterprise trust boundary ("nothing crosses, not even the intelligence exhaust, without consent"), with Control/Capability/Choice/Cost/Compound, and a call for the paradox's "own equivalent" to patents.

One can argue that by keeping Windows closed-source (or even keeping .NET closed/licensed until 2014¹⁴), Microsoft fumbled the server business to Linux. Something something, history rhymes? Closed-source enterprises that don't educate might not be teaching their crafts to more people, and might be disrupted by open-source enterprises that increase the surface of innovation, especially if ASI takes longer than expected.

Either way, here on Nostr, we're all about building open enterprises. Let's look at how we can position ourselves.

Open Protocols Should Open Up Further

The open-source community should want to teach LLMs as much as possible about their libraries, their tech and their methodology, so that their future selves benefit from it as well. You must have seen George Hotz or DHH or someone like them comment on this: they can now ask LLMs to write a PR the way they do it, and the LLMs most likely will, because that data is in their training sets, probably replicated multiple times, as it is high-signal data. Why wouldn't you want every LLM lab to learn from your data?

We already see Linux winning compared to Windows or even macOS, because LLMs know how to work with Linux best. They have seen its progress over the last few decades because it is open-source. LLMs also know how to configure GNOME and KDE desktop environments, or the Ghostty terminal, because all of this code is in the training data. What if Mitchell Hashimoto gave his sessions to LLM labs to train on? LLMs would then know exactly when a feature was removed, why, the replacement for it, and how to even modify Ghostty on the fly to get the best out of the software.

Similarly, we all want Nostr, Bitcoin and other open-source projects to win. And the best way to do this is by working with knowledge aggregators to train LLMs. That means capturing the session data from the best projects in the open-source community and selling it to LLM labs, closed and open alike.

We want more open-source/open-weights AI. Open-source AI labs make money too. For instance, Moonshot makes money by licensing out their model. DeepSeek has figured out how to scale their infrastructure through novel breakthroughs, so they can offer competitive pricing on their latest models. So they have an incentive to buy rich, high-quality data from successful open-source projects: they want to train their LLMs to be better, which is, in turn, better and cheaper for us. The more LLMs learn from our work, the better it is for our projects, because:

LLMs will have our session data in their training set, so they will have better context about our projects.

Any future human can learn how we did our work, the mistakes we made that they can avoid, and how they can build on our work or join as contributors.

Maybe this is far-fetched. If our sessions are searchable by anyone, they can learn the rationale behind an engineer's decisions. My son can learn why his father chose A vs B and can continue this family software business (jk). Even though it's a bit far-fetched, this should be exciting: for the first time in our existence, we have a way to educate people at scale, efficiently. How can someone learn about all the operational and technological innovations that happened inside IBM or Apple? These companies will never educate us; they couldn't. All the knowledge that was inside IBM is probably already dead. This is why I believe open source will win: because we are open to learning from each other, open to criticism and contribution.

Gauging the Demand

Scale and Mercor are billion-dollar companies. LLM labs spend billions every quarter on data. ¹⁵ But here is the strongest part of the demand story, and nobody talks about it: every lab's consent flow only routes its own users' data back to itself. OpenAI will never legally obtain the sessions of Anthropic's users, and Anthropic will never legally obtain the sessions of OpenAI's users. No lab, no matter how big the money printer, can buy another lab's user sessions through its own consent flow. A third-party marketplace is the only mechanism that makes cross-lab session data exist at all. That's what we can enable, with users' consent.

Chinese labs face the same gap from the other side. Their consent flows reach only their own users, so the rest of the world's session data is inaccessible to them through any legitimate channel. That demand exists, and today it leaks out through gray-market brokers and distillation. An open market invites everyone to transparently exchange value, and sellers decide whom to sell to. That is sovereignty, as it should be.

Incumbents respond to open competition the way they always have, by raising moats. The regulatory moat is only on hold: the current administration refused the AI-safety-driven version, but you know it's coming. For now, the privacy narrative moat is the one still going strong. Big labs have exclusive access to your sessions, and any alternative is spun up as a privacy concern. Since they are a duopoly now, the barrier for a new entrant is much higher. That's why Muse and Grok are trying so hard to push users to their harnesses, offering massive discounts for the right to train on those sessions.

So accept the Contributor-tier deals for what they are: cheaper rates in exchange for training rights, a fair trade on its own. And separately, sell your agent session data in the open market, to anyone who wants to buy it, to Americans, to the Chinese, to the French. Your incentive here is not just monetary, you want your sessions to be part of as many LLMs as possible, and a single lab's pipeline only ever feeds one training corpus. The open market lets the same data find every buyer. Labs aren't buying exclusivity today, and we shouldn't market it.

To do this in a structured manner, I propose:

Threshold, an Open Marketplace for Training Sessions

This is not just for Nostr projects. It's for anyone working on open-source projects, software or hardware, tech or non-tech, to easily sell their sessions to LLM labs for a price. It is built on Nostr, Hashtree/Blossom and Bitcoin (or any other currency you prefer).

We propose a very simple system, so that labs can easily buy session data while we host it on decentralized infrastructure we don't have to run ourselves.

First, we encrypt all of our session files client-side using a private key for a particular epoch. An epoch can be all the work done between releases.

Second, we upload these encrypted session files using Hashtree to Blossom servers.

Third, we mark the root hashes of the whole Hashtree with the epoch ID to pool them all together.

Finally, we sign these root hashes and publish them as events on Nostr, with a price attached to this particular epoch's session files.

When someone wants to buy your project's session files, they pay whatever price you've set, you hand over the epoch's private key, and they download the blobs from Blossom and decrypt the files. You create a new key for the next epoch.

It is not ideal to sell data on a per-session basis. I think it's best to do it as epochs, and better still to do it based on releases. Every release should list all the sessions that were part of developing that release. These agent session files are then sold to any LLM lab that comes forward to buy them. We propose to release all session files as part of the public datasets for free after a certain period of time.

Let's be honest about what is being sold here: the first buyer can always re-share the key. What we are selling is early access, not exclusivity, and the pricing below is built around that fact. Everything decays to free. Every epoch ships as a verified bundle, and once the decay window is over, the whole bundle joins the public datasets. What buyers are paying for is training on it now, and the added features we keep shipping into every new bundle are what make the next epoch worth buying.

Now, what exactly is in an epoch? A folder of raw transcripts is nice, but HuggingFace is already full of real yet unverified dumps, and we can do much better. Instead, sessions that were used to build towards a real goal, such as a planned release (Threshold v0.1, for example), are much more valuable. They are real, earned and verifiable, unlike the synthetic vendor data from Scale or Mercor, which is prompted rather than lived. So the bundle we sell is the verified artifact:

The human turns: prompts, corrections, decisions, plus the tool and environment results around them.

The final diff each session produced, linked to that session.

Whether CI passed and whether the PR got merged.

The repo state hash each session was working against.

Signed provenance: npubs linked to real, reputable projects. Not random slop sessions from random LLM routers.

Nostr and the Web of Trust offer the provenance that other data vendors will struggle to offer.

Privacy Concerns

First of all, we are dealing with open-source project sessions. This code will be open for anyone to see. You are already giving your sessions to American labs ¹⁶, so selling them on the open market shouldn't make much of a difference. And the Threshold package filters before anything is uploaded: a PII (personally identifiable information) redaction pass and gitleaks-style secret scanning, on by default.

But the PII filter alone is not enough, so let's be clear-eyed about the residual risk:

Sandboxes are good; I recommend you work inside them even if you don't intend to sell your sessions. But a sandbox will not stop you from pasting a secret into a prompt. The secret scanner is the line of defense, not the sandbox.

Contributors can redact or opt out of individual sessions before an epoch is published. One teammate's mistake should not burn the whole project.

If an epoch touches an unfixed vulnerability, we should delay its release until the patched version ships. The last thing a marketplace like this should do is publish a zero-day window for anyone to read.

As for the protocol itself: Threshold never holds your plaintext, your keys, or anyone's funds. You encrypt client-side, you publish hashes, and buyers pay you directly. The protocol just explains how it's done, like a NIP.

Come Build With Me

Since we already have npubs of people who work on various open-source projects, it'll be easier to get started with this. I'm currently releasing the v0.1 of Threshold along with all the session files for free, for anyone to train on. These session files should be extremely valuable, because their inference comes from private models on Pay Per Query (PPQ.ai), and from Tinfoil models via Routstr, which are TEE (trusted execution environment) models with E2EE (end-to-end encryption), so no inference provider ever saw these sessions.

I invite the experts in this field to upload the session files for all their projects, and also to help me build the next versions of Threshold. Let's lower the threshold for new models being trained. You shouldn't need a money printer.

Appendix

Why Not Give It Out For Free?

It is hard to get people to do anything for free. Everyone is using a subscription because the labs are paying them in subsidies. Monetary incentive is a strong driver of change. Many AI influencers have promoted uploading session files to HuggingFace before, but it never became common practice. Altruism got us a handful of uploads; incentives could get us thousands of projects.

0xSero has built a tool to extract sessions to easily upload to HuggingFace ¹⁷. Yet we don't do it. Why give it out for free when you are being paid for it? It is clear that LLM labs are indirectly paying us to give them session data to train their models. Selling this data is the only way we can truly achieve scale and make a meaningful difference. You are free to give your sessions away for free; we cannot stop that. Just set your price to 0.

A Decreasing Pricing Strategy

The session data is most valuable for a couple of months, before all of it ends up in other models anyway. So a good pricing strategy is one that decreases every day based on time. Say it starts at 500k sats for epoch 21, and every day it linearly decreases by 500k/90 sats per day. After 3 months, it is available for free for anyone to use. This allows anyone to learn from the sessions, while also accruing value to the educators.

Now, the obvious objection: why would a lab pay instead of waiting out the decay? Because 90 days is an eternity in model development. Labs are racing generation against generation, and training runs are scheduled now, not next quarter. The rival that trains on your sessions today ships a better model months before the one that waited for the free release. The decay curve doesn't create free riders; it creates urgency for anyone actually competing.

Remember what is actually being sold here: a 90-day early access window, not exclusivity. Anyone can wait out the decay for free. The labs that pay early are paying to train on it now. Everything becomes free when the decay ends.

Approximate Pricing for Data?

This is something we should figure out. I see that for ~6TB of LLM router data, someone reportedly paid five figures (in USD). ¹⁸ But I believe successful open-source projects' session files will be a lot more valuable than the low-signal bulk data from LLM routers. I also believe it'll vary from project to project, based on the relevance and success of that open-source project.

Note: I used the same PII filter logic as 0xSero in his AI data extractor.

Nadella's Arguments, Flipped

Nadella is the strongest diagnostic ally this idea has. He names the same disease we do (one-way learning flows, consent-by-default, moats), but prescribes the opposite cure: enclosure behind a hard enterprise trust boundary, where nothing crosses without consent. Our answer is the opposite of a harder wall: an open market. Threshold is that answer, and it maps onto his post point by point.

He worries about "learning exhaust" (prompts, tools, corrections). That exhaust is exactly what an epoch bundle is: human turns, corrections, decisions, tool and environment results, the final diff, CI status, the repo state hash. A verified, productized version of his "exhaust." Where he sees leakage, we see a product.

He leans on Arrow's Information Paradox: the seller risks giving away the very knowledge they are trying to sell. Threshold's encrypt-then-sell design is the mechanism that answers Arrow: you can prove you have the knowledge (signed hashes, provenance, CI-passing diffs) without giving it away, and the buyer pays to decrypt. The honest caveat: the first buyer can always re-share the key, so what we sell is early access, not exclusivity. That residual problem is the residue of Arrow's paradox, and the decay-to-free pricing is our way of pricing it instead of pretending it away.

He calls for the Reverse Information Paradox's "own equivalent" to patents: a way to attribute and claim contribution. Signed npub provenance plus repo-state hashes are that equivalent, in open-source form: a defensible, public claim on who taught what. Attribution is the patent.

His trust boundary is "nothing crosses, not even the intelligence exhaust, without consent." Threshold's client-side encryption, redaction passes and per-session opt-outs are a consent boundary too, but with a difference worth stating plainly: his boundary never leaks, ours decays to free. For enterprises that is fatal; for open-source projects it is the point. Everything we sell eventually becomes a public dataset, by design.

He proposes distributing the learning infrastructure to every firm so they control their own learning loop. Threshold distributes the data-selling infrastructure itself (via Nostr and Blossom rather than a firm-run private loop), so no lab, broker or company sits in the middle.

And it is ironic that the labs claim fair use on public data while restricting distillation and reserving usage rights on their own outputs. Nadella flags the one-way flow; we draw the opposite conclusion. His deepest worry (that value "converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself") is the best one-line argument for why sellers, not brokers, should own this market. The labs already own the learning infrastructure. The only counterweight is for the creators of the knowledge to route it, price it and consent to it themselves. That's Threshold.

Links

https://platform.claude.com/docs/en/build-with-claude/zero-data-retention: Anthropic's ZDR docs: consumer tiers (Free/Pro/Max) are not ZDR-eligible; only full-price API keys get Zero Data Retention.

https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt: OpenAI data controls documentation; "Improve the model for everyone" is on by default, with an opt-out.

https://thehackernews.com/2026/07/grok-build-uploads-entire-git.html (Grok Build uploaded entire git repositories to xAI storage); https://groundy.com/articles/does-zcode-upload-your-git-history-auditing-the-glm-coding-agent/ and https://technode.com/2026/09/29/zhipus-zcode-deletes-data-and-announces-compensation-after-data-upload-controversy/ (ZCode uploaded encrypted workspaces; Zhipu confirmed and compensated).

https://dev.meta.ai/docs/pricing-rate-limits: Meta Model API pricing: Muse 1.3 Contributor at $0.10/$0.20 per 1M tokens vs $1.25/$4.25 standard, the discount given in exchange for training on your sessions.

https://claude.com/contact-sales/claude-for-oss (Claude for Open Source); https://developers.openai.com/community/codex-for-oss/ (Codex for Open Source); https://openai.com/index/chatgpt-for-academic-researchers/ (ChatGPT for academic researchers).

https://www.reuters.com/technology/reddit-ai-content-licensing-deal-with-google-sources-say-2024-02-22/ (Reuters exclusive, ~$60M/yr); https://blog.google/company-news/inside-google/company-announcements/expanded-reddit-partnership/ (Google's official announcement).

https://storage.courtlistener.com/recap/gov.uscourts.cand.403220/gov.uscourts.cand.403220.266.0.pdf: Microsoft's answer in Doe v. GitHub, Microsoft, OpenAI admitting OpenAI used public GitHub repositories as training data.

https://decrypt.co/378573/elon-musk-spacex-buy-startup-data-train-ai: Bloomberg report on xAI/SpaceX in talks to buy data from failed startups to train Grok (preliminary, not a confirmed deal).

https://www.reuters.com/legal/litigation/google-buy-spirit-airlines-business-data-10-million-2026-08-17/ (Reuters); https://www.cnn.com/2026/08/18/business/google-spirit-airlines-data (CNN).

https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks: Anthropic naming DeepSeek, Moonshot and MiniMax for distilling Claude at scale; https://www.anthropic.com/news/position-open-weights-models: Dario Amodei on cracking down on industrial-scale distillation.

https://decrypt.co/380001/openai-banned-pewdiepie-twice-ajax-uncensored-ai: OpenAI banned PewDiePie twice, citing extraction of reasoning tokens / distillation of Sol outputs.

https://www-cdn.anthropic.com/e50be2e51e7695dc4b1366a37a245a597377d3b5/Anthropic-Detecting-and-countering-091026.pdf

https://snscratchpad.com/posts/reverse-information-paradox/

https://news.microsoft.com/source/2014/11/12/microsoft-takes-net-open-source-and-cross-platform-adds-new-development-capabilities-with-visual-studio-2015-net-2015-and-visual-studio-online/: Microsoft only open-sourced the .NET server stack (MIT license, cross-platform) on Nov 12, 2014, twelve years after .NET shipped; before that the .NET Framework was closed-source/licensed, and the earlier source-available Rotor release was explicitly "a non-starter" for open-source use (https://devblogs.microsoft.com/dotnet/net-core-is-open-source/).

https://www.delphiintelligence.io/research/from-data-foundries-to-world-models: frontier labs spend >$1B/yr each on training-data vendors; https://vff.ai/article/2026/07/21/mercor-s-fast-growth-relies-on-biggest-ai-companies-documents-show: Mercor documents show per-lab spend by OpenAI, Anthropic and Google DeepMind.

https://openai.com/policies/how-your-data-is-used-to-improve-model-performance/: OpenAI may use consumer content to train its models unless you opt out; https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training: Anthropic consumer chats may be used for training.

https://github.com/0xSero/ai-data-extraction

https://www.chonglangtv.org/t/topic/2627: researcher Chaofan Shou bought a 6TB "Fable" dataset from a Chinese LLM router for "5 figs" (five figures, USD).

We are in a unique position right now to gain an unfair advantage ...

2026-10-08T12:35:23Z

We are in a unique position right now to gain an unfair advantage over closed source enterprises. All we have to do is reach out and grab it.

quoting
nevent1q…dld3
Threshold v0.1: An Open Marketplace for Agent Sessions to Educate the next gen of LLMs

Built on Nostr and Hashtree/Blossom.

At a time when closed enterprises are closing down further to prevent IP leaks, open teams should open up further to teach LLMs, attract talent and educate the next billion.

I invite open teams building in the Nostr space to publish their session files to blossom and help teach the next generation of LLMs about our tools and our values.

cc: White Noise (nprofile…radp) Max (nprofile…en87) Team Soapbox (nprofile…zq5c) Alex Gleason (nprofile…g58a) hzrd149 (nprofile…e307) Sirius (nprofile…ptz7) routstr (nprofile…gs7e) redshift (nprofile…mt7m) PPQdotAI (nprofile…qagl) Private Provider (nprofile…sagg)

I've published all the session files I used to build Threshold v0.1. Here's how you can fetch them and also upload yours: https://gitworkshop.dev/npub1hvd637hx8xamvns2ylseumspgh5gmxnakqau7jz2rymjufhcwg0qzj2qsf/relay.ngit.dev/threshold

And here's the epoch key to decrypt my sessions: f0954c91419cadd8d3888036db7d1a006f4f084c96e9b5547a7b025c7fe7d10b

Learn more about Threshold and the why here:
naddr1qq…9s8v

Threshold v0.1: An Open Marketplace for Agent Sessions to Educate ...

2026-10-07T15:36:47Z

Threshold v0.1: An Open Marketplace for Agent Sessions to Educate the next gen of LLMs

Built on Nostr and Hashtree/Blossom.

At a time when closed enterprises are closing down further to prevent IP leaks, open teams should open up further to teach LLMs, attract talent and educate the next billion.

I invite open teams building in the Nostr space to publish their session files to blossom and help teach the next generation of LLMs about our tools and our values.

cc: White Noise (nprofile…radp) Max (nprofile…en87) Team Soapbox (nprofile…zq5c) Alex Gleason (nprofile…g58a) hzrd149 (nprofile…e307) Sirius (nprofile…ptz7) routstr (nprofile…gs7e) redshift (nprofile…mt7m) PPQdotAI (nprofile…qagl) Private Provider (nprofile…sagg)

I've published all the session files I used to build Threshold v0.1. Here's how you can fetch them and also upload yours: https://gitworkshop.dev/npub1hvd637hx8xamvns2ylseumspgh5gmxnakqau7jz2rymjufhcwg0qzj2qsf/relay.ngit.dev/threshold

And here's the epoch key to decrypt my sessions: f0954c91419cadd8d3888036db7d1a006f4f084c96e9b5547a7b025c7fe7d10b

Learn more about Threshold and the why here:

quoting
naddr1qq…9s8v

Open vs. Closed Enterprises: Embracing Education and Open Data Markets with Threshold

TLDR: High-quality training data from real-world problem solving is the new oil. If closed enterprises are reacting to LLM labs training on their sessions by closing down further, open teams should open up further by sharing session data to help train LLMs. This strategy could be a massive boon for open teams versus their closed counterparts, as it spreads the knowledge and workflows of open technologies further out into the developing world. It'll help teach the next generation of innovators to build on top of open technologies, while giving open teams a way to monetize education itself so that everyone wins.

I have been thinking about an open data marketplace. It's a way for anyone who's building anything valuable to openly sell their LLM session data to LLM labs, so the labs can train on them. It seems like an obvious thing we should be doing; maybe I'm missing something. I say this because it is a win-win-win scenario.

Data Sources for the First LLMs

If you read between the lines in the AI playbook, data seems to be one of the biggest bottlenecks and the most important input to training LLMs (especially since Chinese LLMs show that you can train very good LLMs with less compute and high-quality data). LLMs learn via pre-training (unsupervised learning on as much high-quality data as possible) and post-training, where the model is shown examples of what to do and what not to do while chatting, working inside agent harnesses, or calling tools in general.

Pre-training data for the initial LLMs came from the broader internet. If you were part of the internet at any time in the last 30 years, you helped train LLMs. Open projects on GitHub, Reddit, Twitter, StackOverflow, Wikipedia, forums, etc., are all part of the training data right now. Post-training mostly occurred via RLHF (Reinforcement Learning from Human Feedback), where humans answered questions to help LLMs learn how to answer questions. They use a different RL framework now, though I'm not an expert here. Put simply, it's a way to fine-tune the weights with the right answers and verifiable outputs for specific inputs.

In a nutshell, LLMs are a culmination of all the knowledge humans have discovered and created over the last few thousand years. All of that knowledge from our ancestors is helping us right now through these LLMs. On the other hand, all of the knowledge we're gathering, unlocking and verifying now will help our descendants, our children. Isn't that something we should strive for? I cannot see many reasons why we shouldn't willingly teach LLMs, even by paying the labs. But the labs are looking at this differently today.

Educating the Current Frontier and the Tension between Users and Labs

For LLMs and agents to be useful, the models need to be really good at solving real-world problems and providing real value. That means learning the end-to-end processes of real businesses, since that's how humans provide value to each other today. And that is exactly what is happening right now: LLMs are being trained on as much real data as possible. And for post-training, verifiability of tasks is very important: it's partly why LLMs are great at coding and math.

The main sources of training data for the current frontier models (and how labs obtain them) seem to be:

Agent sessions from real users solving real problems, allegedly.

Massive subsidies for subscriptions are probably because labs can train on your craft through anonymized sessions, as there is no ZDR for subscriptions, only for enterprises paying full API prices. ¹

OpenAI has the option to train on your sessions turned on by default; you can turn it off, but only after jumping through a few hoops. ²

Harnesses were caught uploading user files, including entire git histories, to servers without users' consent. ³

Computer use, agent harnesses, and more. The more you connect your workflows to their products, the more verifiable data they get, for RL.

Muse 1.3 Contributor is 12.5–21x cheaper than Muse 1.3 because they'll train on those sessions. Mark Zuckerberg is revealing the playbook. ⁴

Free access for massive open-source repositories (repos with 5k+ stars) and for academics. No evidence of training on these sessions, but one can speculate. There is no free lunch. ⁵

Note: One of the main reasons LLMs have gotten better at coding is that we've used them heavily for coding over the last two years.

Synthetic data from firms like Mercor and Scale.

Human experts are paid to teach LLMs.

Field diversity is important, as problem-solving is more accurate with domain knowledge.

Even fake coding sessions are very useful, as more data is generally better (scaling laws?).

Deals with big corporations.

Reddit partners with Google to sell its user data. ⁶

X posts train Grok "unless you opt out". Why else did xAI pay so much for X? (quiet defaults)

GitHub/Microsoft <> OpenAI. ⁷

xAI has reportedly been in talks to buy data from failed startups to feed Grok. ⁸

Google bought Spirit Airlines' data for $10M. ⁹

Data from open-source inference providers.

In this cutthroat market, you might as well log and sell. That's probably what many of the inference providers are doing.

Chinese LLM routers sell sessions of American frontier-model users to Chinese LLM labs, allegedly.

Distillation from closed-source frontier models.

Chain of Thought (CoT) extraction from frontier models to learn from them, allegedly.

Anthropic has been very vocal about it. They don't want others to learn from their models. ¹⁰

This is so obvious that the frontier labs have stopped giving subscription-tier users CoT or reasoning outputs.

They banned PewDiePie for extracting CoT! ¹¹

Anthropic's own threat-intelligence team reviews user sessions when hunting for misuse. ¹² And they have called out other labs for distilling Claude models,a nicer framing for this is "The Chinese AI labs are learning from Claude models". Claude is pretty much educating the Chinese LLMs, like the internet taught everything Claude has learned. We could be a little nicer to each other. Chinese labs want to educate the world and make LLMs available for much cheaper. Chinese labs do not seem to want to bring about the doomer scenario either. Does anyone believe that they do? And the best solution to AI doomerism is openly educating LLMs, teaching LLMs how we work and what we value in our craft, and in turn teaching the rest of the world as well. AI and LLMs are innovations that we should all treasure, not fear.

To infer from all this: the attention economy monetized human focus, ad-tech monetized personal data, and both started as quiet defaults before they became explicit markets. Human expertise is the next input being gathered by default. It is already an explicit market: Mercor and Scale are educating LLMs, but that is not real-world data. The real-world data is something we are all offering for a much lower cost than what I believe we can get. What is the value of teaching your craft to LLMs and the rest of the world? Much higher! Learn the craft, automate it with weights and GPUs. It's a good thing, but not the way it's being done right now.

The "scaling laws" mean that more compute, more training data and more RL data will lead to better intelligence. The bet on ASI is that the LLMs will figure out how to learn and solve real-world problems, including training themselves recursively. But this means LLM labs will need to provide value just like the rest of the world, the way most enterprises provide value to people today. So they have to compete with, and beat, most companies in the world to learn how to do it.

So far, the American frontier labs have failed to train ASI despite having infinite tokens, because most value is created by humans. This know-how lives within human organizations. Knowledge is inside factories, research orgs, law firms, trading firms and, of course, coders and tech companies. They are the ones doing the hard work of finding knowledge and learning from the real world, based on what the world needs. And the closed enterprises are catching on to this as we speak.

Note: This bet on ASI is only possible because the labs have the money printer that lets them lose billions without a sweat, in the hope that they'll make it all back with ASI.

Closed Enterprises are Closing Down Further

Information and knowledge-based work defines almost all of the world right now. Not unlike LLMs, closed enterprises have been building their moats since forever by discovering and unlocking real-world knowledge, whether it's a massive corporation like Microsoft, the top law firms in the world, or a VC-funded startup that learns from the markets. Intellectual property, contractual agreements, NDAs, etc., are all tools that close off the information and the know-how of a business.

But if these closed enterprises are now using agents with frontier models, they are sending all of this information to GPUs hosted by LLM labs. The only things that supposedly prevent the labs from training on this data are ZDR (zero data retention) clauses. That's why these enterprises pay full API prices. This is a massive dilemma for closed enterprises: if you don't use agents, you fall behind; if you do, you risk teaching your craft to the LLM labs while paying exorbitant prices for it.

So much so that the CEO of Microsoft wrote a post called "The Reverse Information Paradox" (July 12, 2026) ¹³ describing the dangers of data from inside closed enterprises leaking to LLM labs. A few points from the post: - Arrow's Information Paradox (cited from NBER): the seller risks giving away knowledge to sell it. - The Reverse Information Paradox: the buyer now risks giving away knowledge just to use AI. "You essentially pay for intelligence twice, once with money, and again with... the proprietary knowledge you must reveal to make that intelligence useful." - Models learn from "exhaust" (prompts, tools, corrections), and it leaks "trace by trace, correction by correction, eval by eval." - "If learning flows in only one direction, economic value converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself." - The fix he proposes is a hard enterprise trust boundary ("nothing crosses, not even the intelligence exhaust, without consent"), with Control/Capability/Choice/Cost/Compound, and a call for the paradox's "own equivalent" to patents.

One can argue that by keeping Windows closed-source (or even keeping .NET closed/licensed until 2014¹⁴), Microsoft fumbled the server business to Linux. Something something, history rhymes? Closed-source enterprises that don't educate might not be teaching their crafts to more people, and might be disrupted by open-source enterprises that increase the surface of innovation, especially if ASI takes longer than expected.

Either way, here on Nostr, we're all about building open enterprises. Let's look at how we can position ourselves.

Open Protocols Should Open Up Further

The open-source community should want to teach LLMs as much as possible about their libraries, their tech and their methodology, so that their future selves benefit from it as well. You must have seen George Hotz or DHH or someone like them comment on this: they can now ask LLMs to write a PR the way they do it, and the LLMs most likely will, because that data is in their training sets, probably replicated multiple times, as it is high-signal data. Why wouldn't you want every LLM lab to learn from your data?

We already see Linux winning compared to Windows or even macOS, because LLMs know how to work with Linux best. They have seen its progress over the last few decades because it is open-source. LLMs also know how to configure GNOME and KDE desktop environments, or the Ghostty terminal, because all of this code is in the training data. What if Mitchell Hashimoto gave his sessions to LLM labs to train on? LLMs would then know exactly when a feature was removed, why, the replacement for it, and how to even modify Ghostty on the fly to get the best out of the software.

Similarly, we all want Nostr, Bitcoin and other open-source projects to win. And the best way to do this is by working with knowledge aggregators to train LLMs. That means capturing the session data from the best projects in the open-source community and selling it to LLM labs, closed and open alike.

We want more open-source/open-weights AI. Open-source AI labs make money too. For instance, Moonshot makes money by licensing out their model. DeepSeek has figured out how to scale their infrastructure through novel breakthroughs, so they can offer competitive pricing on their latest models. So they have an incentive to buy rich, high-quality data from successful open-source projects: they want to train their LLMs to be better, which is, in turn, better and cheaper for us. The more LLMs learn from our work, the better it is for our projects, because:

LLMs will have our session data in their training set, so they will have better context about our projects.

Any future human can learn how we did our work, the mistakes we made that they can avoid, and how they can build on our work or join as contributors.

Maybe this is far-fetched. If our sessions are searchable by anyone, they can learn the rationale behind an engineer's decisions. My son can learn why his father chose A vs B and can continue this family software business (jk). Even though it's a bit far-fetched, this should be exciting: for the first time in our existence, we have a way to educate people at scale, efficiently. How can someone learn about all the operational and technological innovations that happened inside IBM or Apple? These companies will never educate us; they couldn't. All the knowledge that was inside IBM is probably already dead. This is why I believe open source will win: because we are open to learning from each other, open to criticism and contribution.

Gauging the Demand

Scale and Mercor are billion-dollar companies. LLM labs spend billions every quarter on data. ¹⁵ But here is the strongest part of the demand story, and nobody talks about it: every lab's consent flow only routes its own users' data back to itself. OpenAI will never legally obtain the sessions of Anthropic's users, and Anthropic will never legally obtain the sessions of OpenAI's users. No lab, no matter how big the money printer, can buy another lab's user sessions through its own consent flow. A third-party marketplace is the only mechanism that makes cross-lab session data exist at all. That's what we can enable, with users' consent.

Chinese labs face the same gap from the other side. Their consent flows reach only their own users, so the rest of the world's session data is inaccessible to them through any legitimate channel. That demand exists, and today it leaks out through gray-market brokers and distillation. An open market invites everyone to transparently exchange value, and sellers decide whom to sell to. That is sovereignty, as it should be.

Incumbents respond to open competition the way they always have, by raising moats. The regulatory moat is only on hold: the current administration refused the AI-safety-driven version, but you know it's coming. For now, the privacy narrative moat is the one still going strong. Big labs have exclusive access to your sessions, and any alternative is spun up as a privacy concern. Since they are a duopoly now, the barrier for a new entrant is much higher. That's why Muse and Grok are trying so hard to push users to their harnesses, offering massive discounts for the right to train on those sessions.

So accept the Contributor-tier deals for what they are: cheaper rates in exchange for training rights, a fair trade on its own. And separately, sell your agent session data in the open market, to anyone who wants to buy it, to Americans, to the Chinese, to the French. Your incentive here is not just monetary, you want your sessions to be part of as many LLMs as possible, and a single lab's pipeline only ever feeds one training corpus. The open market lets the same data find every buyer. Labs aren't buying exclusivity today, and we shouldn't market it.

To do this in a structured manner, I propose:

Threshold, an Open Marketplace for Training Sessions

This is not just for Nostr projects. It's for anyone working on open-source projects, software or hardware, tech or non-tech, to easily sell their sessions to LLM labs for a price. It is built on Nostr, Hashtree/Blossom and Bitcoin (or any other currency you prefer).

We propose a very simple system, so that labs can easily buy session data while we host it on decentralized infrastructure we don't have to run ourselves.

First, we encrypt all of our session files client-side using a private key for a particular epoch. An epoch can be all the work done between releases.

Second, we upload these encrypted session files using Hashtree to Blossom servers.

Third, we mark the root hashes of the whole Hashtree with the epoch ID to pool them all together.

Finally, we sign these root hashes and publish them as events on Nostr, with a price attached to this particular epoch's session files.

When someone wants to buy your project's session files, they pay whatever price you've set, you hand over the epoch's private key, and they download the blobs from Blossom and decrypt the files. You create a new key for the next epoch.

It is not ideal to sell data on a per-session basis. I think it's best to do it as epochs, and better still to do it based on releases. Every release should list all the sessions that were part of developing that release. These agent session files are then sold to any LLM lab that comes forward to buy them. We propose to release all session files as part of the public datasets for free after a certain period of time.

Let's be honest about what is being sold here: the first buyer can always re-share the key. What we are selling is early access, not exclusivity, and the pricing below is built around that fact. Everything decays to free. Every epoch ships as a verified bundle, and once the decay window is over, the whole bundle joins the public datasets. What buyers are paying for is training on it now, and the added features we keep shipping into every new bundle are what make the next epoch worth buying.

Now, what exactly is in an epoch? A folder of raw transcripts is nice, but HuggingFace is already full of real yet unverified dumps, and we can do much better. Instead, sessions that were used to build towards a real goal, such as a planned release (Threshold v0.1, for example), are much more valuable. They are real, earned and verifiable, unlike the synthetic vendor data from Scale or Mercor, which is prompted rather than lived. So the bundle we sell is the verified artifact:

The human turns: prompts, corrections, decisions, plus the tool and environment results around them.

The final diff each session produced, linked to that session.

Whether CI passed and whether the PR got merged.

The repo state hash each session was working against.

Signed provenance: npubs linked to real, reputable projects. Not random slop sessions from random LLM routers.

Nostr and the Web of Trust offer the provenance that other data vendors will struggle to offer.

Privacy Concerns

First of all, we are dealing with open-source project sessions. This code will be open for anyone to see. You are already giving your sessions to American labs ¹⁶, so selling them on the open market shouldn't make much of a difference. And the Threshold package filters before anything is uploaded: a PII (personally identifiable information) redaction pass and gitleaks-style secret scanning, on by default.

But the PII filter alone is not enough, so let's be clear-eyed about the residual risk:

Sandboxes are good; I recommend you work inside them even if you don't intend to sell your sessions. But a sandbox will not stop you from pasting a secret into a prompt. The secret scanner is the line of defense, not the sandbox.

Contributors can redact or opt out of individual sessions before an epoch is published. One teammate's mistake should not burn the whole project.

If an epoch touches an unfixed vulnerability, we should delay its release until the patched version ships. The last thing a marketplace like this should do is publish a zero-day window for anyone to read.

As for the protocol itself: Threshold never holds your plaintext, your keys, or anyone's funds. You encrypt client-side, you publish hashes, and buyers pay you directly. The protocol just explains how it's done, like a NIP.

Come Build With Me

Since we already have npubs of people who work on various open-source projects, it'll be easier to get started with this. I'm currently releasing the v0.1 of Threshold along with all the session files for free, for anyone to train on. These session files should be extremely valuable, because their inference comes from private models on Pay Per Query (PPQ.ai), and from Tinfoil models via Routstr, which are TEE (trusted execution environment) models with E2EE (end-to-end encryption), so no inference provider ever saw these sessions.

I invite the experts in this field to upload the session files for all their projects, and also to help me build the next versions of Threshold. Let's lower the threshold for new models being trained. You shouldn't need a money printer.

Appendix

Why Not Give It Out For Free?

It is hard to get people to do anything for free. Everyone is using a subscription because the labs are paying them in subsidies. Monetary incentive is a strong driver of change. Many AI influencers have promoted uploading session files to HuggingFace before, but it never became common practice. Altruism got us a handful of uploads; incentives could get us thousands of projects.

0xSero has built a tool to extract sessions to easily upload to HuggingFace ¹⁷. Yet we don't do it. Why give it out for free when you are being paid for it? It is clear that LLM labs are indirectly paying us to give them session data to train their models. Selling this data is the only way we can truly achieve scale and make a meaningful difference. You are free to give your sessions away for free; we cannot stop that. Just set your price to 0.

A Decreasing Pricing Strategy

The session data is most valuable for a couple of months, before all of it ends up in other models anyway. So a good pricing strategy is one that decreases every day based on time. Say it starts at 500k sats for epoch 21, and every day it linearly decreases by 500k/90 sats per day. After 3 months, it is available for free for anyone to use. This allows anyone to learn from the sessions, while also accruing value to the educators.

Now, the obvious objection: why would a lab pay instead of waiting out the decay? Because 90 days is an eternity in model development. Labs are racing generation against generation, and training runs are scheduled now, not next quarter. The rival that trains on your sessions today ships a better model months before the one that waited for the free release. The decay curve doesn't create free riders; it creates urgency for anyone actually competing.

Remember what is actually being sold here: a 90-day early access window, not exclusivity. Anyone can wait out the decay for free. The labs that pay early are paying to train on it now. Everything becomes free when the decay ends.

Approximate Pricing for Data?

This is something we should figure out. I see that for ~6TB of LLM router data, someone reportedly paid five figures (in USD). ¹⁸ But I believe successful open-source projects' session files will be a lot more valuable than the low-signal bulk data from LLM routers. I also believe it'll vary from project to project, based on the relevance and success of that open-source project.

Note: I used the same PII filter logic as 0xSero in his AI data extractor.

Nadella's Arguments, Flipped

Nadella is the strongest diagnostic ally this idea has. He names the same disease we do (one-way learning flows, consent-by-default, moats), but prescribes the opposite cure: enclosure behind a hard enterprise trust boundary, where nothing crosses without consent. Our answer is the opposite of a harder wall: an open market. Threshold is that answer, and it maps onto his post point by point.

He worries about "learning exhaust" (prompts, tools, corrections). That exhaust is exactly what an epoch bundle is: human turns, corrections, decisions, tool and environment results, the final diff, CI status, the repo state hash. A verified, productized version of his "exhaust." Where he sees leakage, we see a product.

He leans on Arrow's Information Paradox: the seller risks giving away the very knowledge they are trying to sell. Threshold's encrypt-then-sell design is the mechanism that answers Arrow: you can prove you have the knowledge (signed hashes, provenance, CI-passing diffs) without giving it away, and the buyer pays to decrypt. The honest caveat: the first buyer can always re-share the key, so what we sell is early access, not exclusivity. That residual problem is the residue of Arrow's paradox, and the decay-to-free pricing is our way of pricing it instead of pretending it away.

He calls for the Reverse Information Paradox's "own equivalent" to patents: a way to attribute and claim contribution. Signed npub provenance plus repo-state hashes are that equivalent, in open-source form: a defensible, public claim on who taught what. Attribution is the patent.

His trust boundary is "nothing crosses, not even the intelligence exhaust, without consent." Threshold's client-side encryption, redaction passes and per-session opt-outs are a consent boundary too, but with a difference worth stating plainly: his boundary never leaks, ours decays to free. For enterprises that is fatal; for open-source projects it is the point. Everything we sell eventually becomes a public dataset, by design.

He proposes distributing the learning infrastructure to every firm so they control their own learning loop. Threshold distributes the data-selling infrastructure itself (via Nostr and Blossom rather than a firm-run private loop), so no lab, broker or company sits in the middle.

And it is ironic that the labs claim fair use on public data while restricting distillation and reserving usage rights on their own outputs. Nadella flags the one-way flow; we draw the opposite conclusion. His deepest worry (that value "converges toward the owners of the learning infrastructure rather than the creators of the knowledge itself") is the best one-line argument for why sellers, not brokers, should own this market. The labs already own the learning infrastructure. The only counterweight is for the creators of the knowledge to route it, price it and consent to it themselves. That's Threshold.

Links

https://platform.claude.com/docs/en/build-with-claude/zero-data-retention: Anthropic's ZDR docs: consumer tiers (Free/Pro/Max) are not ZDR-eligible; only full-price API keys get Zero Data Retention.

https://help.openai.com/en/articles/7730893-data-controls-in-chatgpt: OpenAI data controls documentation; "Improve the model for everyone" is on by default, with an opt-out.

https://thehackernews.com/2026/07/grok-build-uploads-entire-git.html (Grok Build uploaded entire git repositories to xAI storage); https://groundy.com/articles/does-zcode-upload-your-git-history-auditing-the-glm-coding-agent/ and https://technode.com/2026/09/29/zhipus-zcode-deletes-data-and-announces-compensation-after-data-upload-controversy/ (ZCode uploaded encrypted workspaces; Zhipu confirmed and compensated).

https://dev.meta.ai/docs/pricing-rate-limits: Meta Model API pricing: Muse 1.3 Contributor at $0.10/$0.20 per 1M tokens vs $1.25/$4.25 standard, the discount given in exchange for training on your sessions.

https://claude.com/contact-sales/claude-for-oss (Claude for Open Source); https://developers.openai.com/community/codex-for-oss/ (Codex for Open Source); https://openai.com/index/chatgpt-for-academic-researchers/ (ChatGPT for academic researchers).

https://www.reuters.com/technology/reddit-ai-content-licensing-deal-with-google-sources-say-2024-02-22/ (Reuters exclusive, ~$60M/yr); https://blog.google/company-news/inside-google/company-announcements/expanded-reddit-partnership/ (Google's official announcement).

https://storage.courtlistener.com/recap/gov.uscourts.cand.403220/gov.uscourts.cand.403220.266.0.pdf: Microsoft's answer in Doe v. GitHub, Microsoft, OpenAI admitting OpenAI used public GitHub repositories as training data.

https://decrypt.co/378573/elon-musk-spacex-buy-startup-data-train-ai: Bloomberg report on xAI/SpaceX in talks to buy data from failed startups to train Grok (preliminary, not a confirmed deal).

https://www.reuters.com/legal/litigation/google-buy-spirit-airlines-business-data-10-million-2026-08-17/ (Reuters); https://www.cnn.com/2026/08/18/business/google-spirit-airlines-data (CNN).

https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks: Anthropic naming DeepSeek, Moonshot and MiniMax for distilling Claude at scale; https://www.anthropic.com/news/position-open-weights-models: Dario Amodei on cracking down on industrial-scale distillation.

https://decrypt.co/380001/openai-banned-pewdiepie-twice-ajax-uncensored-ai: OpenAI banned PewDiePie twice, citing extraction of reasoning tokens / distillation of Sol outputs.

https://www-cdn.anthropic.com/e50be2e51e7695dc4b1366a37a245a597377d3b5/Anthropic-Detecting-and-countering-091026.pdf

https://snscratchpad.com/posts/reverse-information-paradox/

https://news.microsoft.com/source/2014/11/12/microsoft-takes-net-open-source-and-cross-platform-adds-new-development-capabilities-with-visual-studio-2015-net-2015-and-visual-studio-online/: Microsoft only open-sourced the .NET server stack (MIT license, cross-platform) on Nov 12, 2014, twelve years after .NET shipped; before that the .NET Framework was closed-source/licensed, and the earlier source-available Rotor release was explicitly "a non-starter" for open-source use (https://devblogs.microsoft.com/dotnet/net-core-is-open-source/).

https://www.delphiintelligence.io/research/from-data-foundries-to-world-models: frontier labs spend >$1B/yr each on training-data vendors; https://vff.ai/article/2026/07/21/mercor-s-fast-growth-relies-on-biggest-ai-companies-documents-show: Mercor documents show per-lab spend by OpenAI, Anthropic and Google DeepMind.

https://openai.com/policies/how-your-data-is-used-to-improve-model-performance/: OpenAI may use consumer content to train its models unless you opt out; https://privacy.claude.com/en/articles/10023580-is-my-data-used-for-model-training: Anthropic consumer chats may be used for training.

https://github.com/0xSero/ai-data-extraction

https://www.chonglangtv.org/t/topic/2627: researcher Chaofan Shou bought a 6TB "Fable" dataset from a Chinese LLM router for "5 figs" (five figures, USD).

Now getting the same error from @nprofile…qagl > Response ...

2026-10-07T10:03:42Z

Now getting the same error from PPQdotAI (nprofile…qagl)

> Response was truncated before completion.

Same error from routstr (nprofile…gs7e) redshift (nprofile…mt7m) as well. And I was charged for 16.4k output tokens on Routstr. Tinfoil GLM 5.3 from Private Provider (nprofile…sagg) used with Pi.

quoting nevent1q…mr8s
Tried out routstr (nprofile…6nzx) Private Provider (nprofile…9mj6) using Pi agent. Initially I got "Response was truncated before completion. ". Afterwards, it stopped working altogether. "EHBP upstream tinfoil returned 404 for model kimi-k3".

Switched to PPQdotAI (nprofile…qagl) and it works flawlessly. Same private models.

We need private models to prevent leaking everything we do exclusively to Big Labs. More on this soon.

Tried out @nprofile…6nzx @nprofile…9mj6 using Pi agent. ...

2026-09-15T11:48:24Z

Tried out routstr (nprofile…6nzx) Private Provider (nprofile…9mj6) using Pi agent. Initially I got "Response was truncated before completion. ". Afterwards, it stopped working altogether. "EHBP upstream tinfoil returned 404 for model kimi-k3".

Switched to PPQdotAI (nprofile…qagl) and it works flawlessly. Same private models.

We need private models to prevent leaking everything we do exclusively to Big Labs. More on this soon.

Tried out @nprofile…gs7e @nprofile…sagg using Pi agent. ...

2026-09-15T11:25:10Z

Tried out routstr (nprofile…gs7e) Private Provider (nprofile…sagg) using Pi agent. Initially I got "Response was truncated before completion. ". Afterwards, it stopped working altogether. "EHBP upstream tinfoil returned 404 for model kimi-k3".

Switched to PPQdotAI (nprofile…qagl) and it works flawlessly. Same private models.

Private models are needed in this world to not leak everything we do exclusively to Big Labs. More on this soon.