Language models have become privately controlled research infrastructure. This week, OpenAI deprecated the Codex model that ~100 papers have used—with 3 days’ notice. It has said that newer models will only be stable for 3 months. Goodbye reproducibility!

It'll be interesting to see how developers are going to use these models in production if things are going to break every couple of months.

By @sayashk and me on the AI Snake Oil book blog: https://aisnakeoil.substack.com/p/openais-policies-hinder-reproducible

OpenAI’s policies hinder reproducible research on language models

LLMs have become privately-controlled research infrastructure

AI Snake Oil

OpenAI responded to the criticism by saying they'll allow researchers access to Codex. But the application process is opaque: researchers need to fill out a form, and the company decides who gets approved. It is not clear who counts as a researcher, how long they need to wait, or how many people will be approved. Most importantly, Codex is only available through the researcher program “for a limited period of time” (exactly how long is unknown).

https://aisnakeoil.substack.com/p/openais-policies-hinder-reproducible

OpenAI’s policies hinder reproducible research on language models

LLMs have become privately-controlled research infrastructure

AI Snake Oil

@randomwalker "OpenAI isn't"

We need to happen to LLMs what happened to generative image AIs: an entity like Stability to come along and release (generally) license-unhindered weights to the public for good models that can run on consumer-grade GPUs.

We're "kinda" there with LLMs, in that LLaMA weights were released. Though the license is kind of burdensome, and support to make something ChatGPT-ish out of it is still being hacked together.

(BTW, your substack name is a turnoff to discussion)

@nafnlaus Fair point about the name. We use it because it has some history behind it (https://aisnakeoil.substack.com/p/introducing-the-ai-snake-oil-book) and is a bit of a recognized brand at this point, which might be helpful when our book comes out. But perhaps we're weighted down by that baggage. We'll discuss it. Thanks! CC @sayashk
Introducing the AI Snake Oil book project

Something weird happened on November 19, 2019. When a professor shares scholarly slides online, they are usually intended for a niche group of peers. You’d be lucky if 20 people looked at them. But that day, the slides Arvind released went viral. They were downloaded tens of thousands of times and his tweets about them were viewed 2 million times.

AI Snake Oil
@randomwalker
I vote keep the name. We are rocketing up the Gartner hype curve faster than we can assess risks. Someone has to call it out.
@nafnlaus @sayashk
@randomwalker You think people would have learned by now.
@randomwalker @sayashk I’m quite unhappy with these changes myself. One of the biggest use-cases I was working on in Feb was using Codex. I look away for 2-3 weeks, and they EOL the model itself. Incredible.
@randomwalker @sayashk They really should rename themselves. #ClosedAI is more appropriate.

@randomwalker @sayashk
My biggest disappointment is that those who made the AI had no measures or plans for crediting those behind the texts or pics that appear on the screen when someone makes a request to it.

So they, in their views, discredit all of those people's compiled work that are appearing in one, as if to be, a single minded 'creature' called 'ChatGPT', and referred to as the,
"MASTER ! THE CONQOURER! THE AI THAT WILL LEAD THE WORLD AND THE FUTURE !"

@randomwalker @sayashk
So I hope people discontinue giving credits for OpenAI and ChatGPT's creators for discrediting their own works, in return. Thousands upon thousands of works stolen in 'legal grounds', repurposed, and reused.

Such a toxicating culture, by the way.

@randomwalker @sayashk I assume this means they found something me commercial monetization?

@randomwalker A lot of the research infrastructure that’s critical for #reproducibility is being lightheartedly outsourced to private organizations—DockerHub, mybinder.org, and GitHub come to mind.

I think researchers should carefully assess the risks associated with the use of those gratis services that look so appealing. The interests of the organizations behind them may simply not be aligned with those of #OpenScience.

@khinsen

@civodul @randomwalker Indeed. And that's only scratching the surface of issues with lacking digital infrastructures for research.

For more examples, see:

https://jon-e.net/infrastructure/

(by @jonny)

Decentralized Infrastructure for (Neuro)science

Decentralized Infrastructure for (Neuro)science

@khinsen
@civodul @randomwalker

you freaking said it. our failure of imagination will not only buy us new triple pay industries, but make us complicit in the ballooning platform surveillance economy -- all while setting a hard cap on the ambition of our infrastructures because the cloud has intrinsic limitations in service of maintaining total control for the platform holder.

nearly finished with an expanded look on large scale infra projects in the US that do exactly this and the ideologies and influences that drive this process.

@jonny

ArXiv:2209.07493 [1] is an impressive review of the issue of not just digital infrastructure for research, but how to reform research overall. Since it's CC BY-NC-SA, in principle the astro community could hack it and do a version more appropriate to our own current situation. My guess is that in some ways we're a lot more cooperative - but with a lot of GAFAM-ish reliance.

@khinsen @civodul @randomwalker

#ArXiv_2209_07493

[1] https://arxiv.org/abs/2209.07493 = https://jon-e.net/infrastructure

Decentralized Infrastructure for (Neuro)science

The most pressing problems in science are neither empirical nor theoretical, but infrastructural. Scientific practice is defined by coproductive, mutually reinforcing infrastructural deficits and incentive systems that everywhere constrain and contort our art of curiosity in service of profit and prestige. Our infrastructural problems are not unique to science, but reflective of the broader logic of digital enclosure where platformatized control of information production and extraction fuels some of the largest corporations in the world. I have taken lessons learned from decades of intertwined digital cultures within and beyond academia like wikis, pirates, and librarians in order to draft a path towards more liberatory infrastructures for both science and society. Based on a system of peer-to-peer linked data, I sketch interoperable systems for shared data, tools, and knowledge that map onto three domains of platform capture: storage, computation and communication. The challenge of infrastructure is not solely technical, but also social and cultural, and so I attempt to ground a practical development blueprint in an ethics for organizing and maintaining it. I intend this draft as a rallying call for organization, to be revised with the input of collaborators and through the challenges posed by its implementation. I argue that a more liberatory future for science is neither utopian nor impractical -- the truly impractical choice is to continue to organize science as prestige fiefdoms resting on a pyramid scheme of underpaid labor, playing out the clock as every part of our work is swallowed whole by circling information conglomerates. It was arguably scientists looking for a better way to communicate that created something as radical as the internet in the first place, and I believe we can do it again.

arXiv.org
GitHub - sneakers-the-rat/infrastructure: in-progress manuscript

in-progress manuscript. Contribute to sneakers-the-rat/infrastructure development by creating an account on GitHub.

GitHub
@boud
@khinsen @civodul @randomwalker
also super curious about the markup you are using in that post- does your instance handle arxiv IDs in a special way?

@jonny

ArXiv IDs: No :).

I was just guessing a format for ArXiv hashtags that other people might use - but it looks like #ArXiv_2209_07493 is more likely to be used than #ArXiv220907493 (and I see that I mistyped - I'll edit that).

@khinsen @civodul @randomwalker

@boud
@khinsen @civodul @randomwalker
oh dang, there's been an issue open in base masto forever on this, I have wanted to modify the regex for tags on our instance to handle colons and slashes, and also just be able to dereference links with a doi. ok thatll be the next masto hack.
@boud
@khinsen @civodul @randomwalker
being able to talk about "the same thing," simultaneously underrated and overrated in a way too situationally complex to be worth defining lmao
@boud
@khinsen @civodul @randomwalker
or, I suppose trying to talk about the "same thing" is just too fundamental to the act of language to capture glibly. one of those.

@jonny

At the risk of a divergent conversation :) ... Language + hidden cognitive assumptions about spacetime are what lead to paradoxes about "the same thing" in Minkowski spacetime ;). The terms "time dilation" and "space contraction" are misleading until you understand what they mean (projections) on a spacetime diagram. @cartographer did a boost which suggests that I probably *should* record my lectures on this ... so far there's been no other interest so I've been undecided.