The manifesto argues in plain HTML that if you cannot run, audit, or fork the model your product depends on, you are a tenant of the vendor rather than the owner of your own stack. It frames open source AI not as an ideological preference but as the only path to genuine product control, echoing the structural arguments the FSF has made for decades.
By submitting the manifesto and watching it climb from 477 to 1,363 points over ten hours, vednig surfaced an argument that resonated with an HN audience already primed by parallel threads about proactive agents, reverse-centaur dynamics, and Kimi K2.7's tokens-per-task pitch. The sustained climb past the first curiosity wave signals the position has real operational backing.
The editorial argues the manifesto's persistence isn't about FOSS purism but about a concrete pain point: when your product depends on a proprietary model, the vendor's product decisions — like Claude Fable's 'relentlessly proactive' default — become your product decisions, whether you want them or not. The parallel trending of Willison's Fable piece, Grinberg's reverse-centaur essay, and Moonshot's Kimi K2.7-Code release is offered as evidence that the audience is reading these as one connected
A single-page manifesto at opensourceaimustwin.com went from 477 points on Hacker News to 1,363 in roughly ten hours. No product. No team page. No screenshots, no benchmarks, no founder bio, no commercial pitch. Just one argument rendered in plain HTML: if you cannot run, audit, or fork the model your product depends on, you are a tenant, not an owner.
Manifestos that get a flash spike on HN are common. Manifestos that nearly triple their score after the first wave of curious clicks have already moved on are not. The site's persistence over a half-day window is the signal worth reading. The argument itself isn't new — variations of it have been in circulation since the first Llama weights leaked in 2023, and the FOSS movement has been making structurally identical arguments since the Free Software Foundation was a one-room operation. The question is why this particular framing, in this particular week, is still climbing.
The answer isn't on the page. It's in the HN comments, and it's in the parallel threads the same audience has been clicking on for the last forty-eight hours.
Look at what else has been trending in roughly the same window. Simon Willison's piece on Claude Fable's "relentlessly proactive" agent behavior pulled 680 points and a comment section split clean between hobbyists who love the new default and senior devs who dread it in production. Miguel Grinberg's "I Am Not a Reverse Centaur" essay hit 178 points by naming what happens when the human stops leading the agent loop. Tom Bedor's piece on the collapsing cost of asking for human attention cleared 1,306. Moonshot's Kimi K2.7-Code release — an open-weight model pitched explicitly on tokens-per-task rather than benchmark wins — hit 287 in the same audience.
The pattern is operational, not ideological. Every one of these stories is downstream of the same uncomfortable fact: when your product depends on a proprietary model, the vendor's product decisions become your product decisions, on the vendor's schedule, with no veto. Claude Fable shipped a behavior change that broke the implicit contract some teams had built their UX around. Anthropic, OpenAI, and Google have all deprecated production-grade models on six-to-twelve-month timelines, leaving downstream teams to redo prompt engineering, regression suites, and calibration data against a new black box. Pricing has whiplashed. Rate limits have shifted under load. Safety filters have tightened mid-quarter and broken integrations that shipped clean.
The manifesto isn't winning because more developers read the manifesto. It is winning because more developers are paying the operational tax it predicts, and they keep arriving on the page because someone finally wrote the abstract version of the bill they just paid. Every front-page incident involving a closed model is doing the manifesto's marketing for it.
That said, the manifesto has a definitional hole worth naming, and the HN comments are starting to. It uses "open source AI" as a single phrase, but the open AI stack has at least three distinct layers — open weights, open training data, and open inference — and almost none of the popular "open" models are open on all three. Llama 3, DeepSeek-V3, Qwen 2.5, Kimi K2.7 — all open-weight, mostly unauditable on training data, and most teams run them through whatever inference stack their cloud vendor ships rather than vLLM or llama.cpp on bare metal. "Open weight only" is closer to "tenant with portable luggage" than "owner." You can leave when the vendor goes hostile. You cannot answer the question "what did this model see during training" with anything other than the vendor's word.
Even with that caveat, open-weight buys you the one capability the manifesto actually cares about in practice: deployment sovereignty. You can run the model on your own hardware, in your own jurisdiction, against your own logs, with your own fine-tunes, indefinitely. That is a much weaker claim than "open source" in the FOSS tradition, but it is the claim that maps to the operational pain driving the HN signal.
The honest practitioner question is not "should I switch to open models." It is "which of my model dependencies survive a vendor going hostile, going down, or simply changing the default behavior of a model name I shipped against." Run that audit. For anything where a silent behavior change, a deprecation, or a sudden pricing move would force a quarter of unplanned engineering work, the answer needs to be an open-weight model you can pin and run.
The expensive trap is calibration lock-in: every eval suite, every prompt template, every guardrail you write against a specific proprietary endpoint is capital you have to write off when that endpoint changes. If your team has spent six months tuning a Claude or GPT pipeline and the next model release breaks the behavior — Fable-style — you are buying that engineering again, and the second time without the budget runway you had the first time. Pin the prompts and evals to open-weight checkpoints you control for anything mission-critical. Use the closed APIs as a commodity quality layer on top — easy to swap, cheap to lose.
Concrete moves worth doing this quarter: keep a same-task fallback open-weight model warm even when closed APIs are primary. Version your prompts against a specific checkpoint hash, not "Claude" or "GPT-4." Treat your fine-tunes and embedding spaces as production assets, with the same exit-cost discipline you would apply to a database migration. Stand up at least one inference path on hardware you own or rent on instance-hour terms, even if it costs more per token than the closed API today — the option value is the point.
The manifesto will keep climbing as long as proprietary model drift keeps producing front-page incidents, which is to say, indefinitely. The next move in this debate is going to be definitional: "open weights" becomes the procurement minimum, "open training data" becomes the next frontier, and the vendors fighting over the word "open" will look like the early-2010s "open core" battles all over again. Expect enterprise RFPs to start asking for weight access explicitly, and expect EU AI Act compliance work to start citing it as a risk-mitigation primitive. The manifesto is a weak technical argument and a strong operational one. That is why it is winning.
This, and distributed LLM inference. We are at a point where no single person can setup a rig to run a SOTA model, it is just too expensive.So we must build and adopt frameworks that allow individuals to share resources to run SOTA models in a distributed manner. That way they will also be non-censo
It won in my house/my business right from the start. (Well, open weights, at least — which is an uncomfortable nuance.)I have never understood the willingness to make the functioning of or development of a product so completely dependent on the secret sauce of one of two big unprofitable, inscr
When "open source" means freeware, it's like saying "we want free copies".What we should be saying is: We want a public, community-ran project that does pretraining and training collectively. This means working on a training corpus in public and somehow coordinating the trai
I would be totally willing to pay $50 per month to support an open source AI lab, rather to get open source models as byproducts of corporations.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I've been contemplating a decentralized model training system for some time using volunteer machines that we all contribute. But, it is astronomically difficult. The communication speeds are untenable.And, there is the issue of data poisoning from untrusted nodes. I've almost cracked that