iiinova — where your digital self keeps evolving

Reposted by Dave Willner

BeijingPalmer @beijingpalmer.bsky.social · 01/10/2026

righteous among the nations

Dave Willner @dwillner.bsky.social · 29/09/2026

I’ve never seen the “no…but also no…also no” meme format with “why are you crying” format and I have to complement you on the truly excellent pairing.

031

Reposted by Dave Willner

Renee DiResta @noupside.bsky.social · 23/09/2026

New work featured in New Public!: There is no neutral feed algorithm. Each one makes source, ranking, & content diversity choices. @jonathanstray.com & I started MySky so you can see - and change! - these settings. Check it out here on Bluesky, in "Feeds" # newpublic.substack.com/p/meet-the-a...

newpublic.substack.com

🎛️📲 Meet the algorithm you can actually control and understand

Jonathan Stray and Renée DiResta introduce MySky, a feed that shows its work

37135

Dave Willner @dwillner.bsky.social · 23/09/2026

Honestly, I feel the same way about the Matrix.

110

Dave Willner @dwillner.bsky.social · 21/09/2026

@masnick.com and I took the time to formalize some thoughts around the unlikelihood of the “sexy apocalypse” as compared to the dumb, disappointing, and far more likely ones.

2286

Dave Willner @dwillner.bsky.social · 20/09/2026

This is a great example, imho. "A secret bioweapon escaped" is bad, but also *sort of awesome* in the original sense. "We lost half a decade because someone ate the wrong bat" is (a) not fun and (b) existentially scary because it reminds us of the fragility of our civilization.

111

Dave Willner @dwillner.bsky.social · 19/09/2026

Honestly, this still strikes me intuitively as a bit too sexy. I’m thinking more like “it makes someone very rich because it’s good at some things but it still has the common sense of someone who has been kicked in the head by a horse and deletes the banking system for some reason”

020

Dave Willner @dwillner.bsky.social · 18/09/2026

It’s literally because the second thing isn’t sexy and the first thing is.

0160

Reposted by Dave Willner

But Thou Mustelid @rewhan.bsky.social · 18/09/2026

"Doomsday isn't when the bridge falls down, it's when it stops being put back up."

38423

Dave Willner @dwillner.bsky.social · 18/09/2026

Weirdly I think the problem is it wasn't bad enough (not to minimize Covid) and so it didn't quite do it for folks. So instead, it just made everyone more escapist and dedicated to their fandoms (whatever they may be).

120

Dave Willner @dwillner.bsky.social · 18/09/2026

We also would need to be significantly better at making paperclips, which I don't think is a current area of heavy investment.

120

Dave Willner @dwillner.bsky.social · 18/09/2026

I can't believe I'm saying this, but this is a touch unfair to the truly dedicated preppers, many of whom do take water purification extremely seriously. But your point stands spiritually, because they buy them for themselves, instead of like...voting for higher taxes and more robust government?

0230

Dave Willner @dwillner.bsky.social · 18/09/2026

100% agree - I’ve actually used the flash-crash problem as an example explicitly before as an example of the kind of stupid disaster we should expect.

020

Dave Willner @dwillner.bsky.social · 18/09/2026

Even if you accept an implicit scoping to “chicken egg” as the egg in question it still predates the chicken because evolutionarily the first chicken egg was laid by a not-quite-a-chicken.

170

Dave Willner @dwillner.bsky.social · 18/09/2026

I think that’s right when you broaden it beyond the AI-ish risk area that I was initially addressing. Others have noted, as I think you’re also hinting at, that it is a somewhat masculine-coded thing (though of course not exclusively so)

020

Dave Willner @dwillner.bsky.social · 18/09/2026

Yup. People want the bang, and they might even accept the whimper. But social solidarity and emotional labor is icky and no fun.

12307

Dave Willner @dwillner.bsky.social · 18/09/2026

Those are certainly very rich sources of stupidity!

190

Dave Willner @dwillner.bsky.social · 18/09/2026

No times are interesting in the event, they are only terrifying, dumb, or terrifying and dumb.

2341

Dave Willner @dwillner.bsky.social · 18/09/2026

Please do, I would be intensely honored

050

Dave Willner @dwillner.bsky.social · 18/09/2026

For sure - I'm not saying there is no chance of AI disaster. I'm saying that the most likely AI disasters are stupid. E.g. we're more likely end up with no internet b/c a bunch of nonsense machines install themselves on every server than we are to all die b/c Claude Awakens With Terrible Purpose.

050

Dave Willner @dwillner.bsky.social · 18/09/2026

This is incredibly spot on and I'm going to steal it. Also everyone should read "The Ghost Map", which is about the guy who figured out that cholera was waterborne in the 1850s. The punchline is "he was right, but they sent him to a mental asylum because they thought it was smelly air"

231

Dave Willner @dwillner.bsky.social · 18/09/2026

The use of made up percentages is such an incredible tell in this vein. It's completely meaningless nonsense with no possible basis for justification. It's just a way of saying "I'm feel it strongly" that sounds sciency and therefore autoritative/real.

030

Dave Willner @dwillner.bsky.social · 18/09/2026

Hanlon's razor is very sharp and cuts deep.

010

Dave Willner @dwillner.bsky.social · 18/09/2026

Spot on. Conspiracy theories are comforting because, in that world, *someone* is competent, knows what they're doing, and has a plan. They're evil, sure, but they're on top of things. The truth that no one knows what they're doing and it's mostly idiocy and bullshit is much, much more disturbing.

3100

Dave Willner @dwillner.bsky.social · 18/09/2026

Yup. Put another way - there relatively little overlap between the kind of disasters that are fun to speculate about (if you're detached from the reality of them) and the kind of disasters that are most likely to actually happen. There's a sort of "fandom" dynamic to the speculative part.

021

Dave Willner @dwillner.bsky.social · 18/09/2026

Fair - @jamellebouie.net is entirely correct that "everything is gender" in the current political moment.

0420

Dave Willner @dwillner.bsky.social · 17/09/2026

It's been fantastic supporting the Runway team in their work on this and I can't want to see where they take things next.

000

Dave Willner @dwillner.bsky.social · 17/09/2026

This sort of thing is really only possible when you have a multi-modal model that combines accuracy with speed in the way that CoPE does. Big foundation models can do the classification well, but aren't fast enough. Unspecialized small models are fast enough, but can't do the classification well.

120

Dave Willner @dwillner.bsky.social · 17/09/2026

Runway is doing some very cool stuff using CoPE/Zentropi: "We have designed and tested a new fast moderation system that will run synchronously once a user’s input clears moderation.... The synchronous moderation through Zentropi narrows the exposure window to less than half a second."

runway.com

Runway News | Moderation in Real Time

Runway built a synchronous moderation system for real-time video generation to scan streaming frames for harmful content in under half a second — layered on Runway's existing safety defenses.

132

Dave Willner @dwillner.bsky.social · 17/09/2026

Put another way, the awesome scifi doom presupposes that we, collectively, are doing a *much better job* than there is any reason to believe we are, or should be expected to.

187548

Dave Willner @dwillner.bsky.social · 17/09/2026

No one gets a PhD in "what if the people in charge of the Pentagon are chud idiots and they collapse the world economy because caring about strategy is Woke and Gay". But that is (self-evidently) a far more likely thing to happen, simply because it only requires stupidity to accomplish.

81707239

Dave Willner @dwillner.bsky.social · 17/09/2026

Certain kinds of smart people tend to have a "sexy apocalypse" problem when it comes to thinking about risk. E.g. skynet turning everyone in paperclips, while bad, is also *kind of cool and fun* at the same time. Where the reality is that most actual disasters are both stupid and disappointing.

222573322

Dave Willner @dwillner.bsky.social · 16/09/2026

I realized a month or so ago that, based on how far we’ve pushed these tools, I would probably never write a policy from scratch myself ever again. A wild feeling given how much of my life up till now has been spent on that!

110

Reposted by Dave Willner

Samidh @samidh.bsky.social · 16/09/2026

Today we are releasing the next generation of our policy optimization tools for content classifiers. It is our hope this can be another step towards helping shore up human control over AI-powered systems. blog.zentropi.ai/optimizing-o...

blog.zentropi.ai

Optimizing our Policy Optimizers

Today, we are releasing our next-generation policy refinement tools: policy-only correction, label-only correction, and auto-optimization.

032

Dave Willner @dwillner.bsky.social · 16/09/2026

On a personal note - figuring out how to get this to work has taken months of focused effort and I am just incredibly excited to get it out the door. Please do try it and let us know how it goes!

040

Dave Willner @dwillner.bsky.social · 16/09/2026

The system is available now on zentropi.ai usable directly on the website or through the Zentropi Agent Skill for coding agents. Normally, full use is subscriber-only because of how compute-intensive it is, but for launch we’re giving anyone a limited quota of free so folks can try it.

zentropi.ai

Zentropi - Build Custom Content Labelers Instantly

162

Dave Willner @dwillner.bsky.social · 16/09/2026

As a result, this new optimizer has three distinct capabilities: policy-only correction rewrites the policy text to better match a labeled dataset. Label-only correction flags and fixes labels that no longer match the policy. Auto-optimization runs both in a loop until the whole stops improving.

141

Dave Willner @dwillner.bsky.social · 16/09/2026

To summarize the point - a policy and a set of examples labeled against it aren't independent things you can fix one at a time, because a given label can only be said to be meaningfully right or wrong relative to a specific policy text. So, to improve either you have to work on both simultaneously.

151

Dave Willner @dwillner.bsky.social · 16/09/2026

At TrustCon this year I talked about a technique we’ve developed for automatically optimizing content-moderation policies, using an inversion of the binocular labeling approach Zentropi had already pioneered. Today we're shipping the tool that technique became. blog.zentropi.ai/optimizing-o...

blog.zentropi.ai

Optimizing our Policy Optimizers

Today, we are releasing our next-generation policy refinement tools: policy-only correction, label-only correction, and auto-optimization.

3176

Reposted by Dave Willner

The Future of Free Speech @futurefreespeech.org · 15/09/2026

Disinformation in AI presents serious risks. But giving authorities the power to decide what chatbots can say is an even bigger risk. At the #FreeSpeechSummit2026, @dwillner.bsky.social will discuss how open-source content moderation offers a promising solution.

111

Dave Willner @dwillner.bsky.social · 27/08/2026

The post has the argument and the slides have the mechanics. The skill that runs a version of the loop is public, so you can point your own agent at a policy you care about: github.com/zentropi-ai/skills 🧵 5/5

github.com

GitHub - zentropi-ai/skills: Agent skills powered by the Zentropi content classification engine

Agent skills powered by the Zentropi content classification engine - zentropi-ai/skills

010

Dave Willner @dwillner.bsky.social · 27/08/2026

Closing them is a loop: run the policy against examples you trust, look where the model disagrees with your reviewers, and let a second agent rewrite the passages behind the disagreement in the policy's own voice. Partners have gone from 65% to over 90% agreement with their own labels. 🧵 4/5

111

Dave Willner @dwillner.bsky.social · 27/08/2026

This makes the policy's gaps starkly visible: the boundary condition nobody thought through, the term that means one thing on page one and something else on page four. Reviewers paper over those issues without noticing they are doing it. 🧵 3/5

100

Dave Willner @dwillner.bsky.social · 27/08/2026

The argument, briefly: once a small model (CoPE) reads a content policy as written and labels against it, editing the policy is editing the classifier, with no retraining cycle in between. 🧵 2/5

100

Dave Willner @dwillner.bsky.social · 27/08/2026

I gave a talk at TrustCon this year that seemed well received, so we have written the argument up as a post on the Zentropi blog and are sharing the slides with it: blog.zentropi.ai/machines-edi... 🧵 1/5

blog.zentropi.ai

Machines editing the rules that machines enforce

In a talk at TrustCon this year, Dave detailed the procedure that Zentropi has pioneered to automatically optimize a policy.

151

Reposted by Dave Willner

Renee DiResta @noupside.bsky.social · 14/08/2026

Here’s Mike Benz bullshitting Matt Taibbi ab his supposed experience as a State Department insider. Listen to Taibbi eat it up. This is the real origin of the “Censorship Industrial Complex” hoax: an operative selling a rube a lie. Here’s what happened next: www.technologyreview.com/2026/08/07/1...

716947

Reposted by Dave Willner

Katie Harbath @katieharbath.bsky.social · 05/08/2026

I'm planning the book tour for Disrupting Politics and building the list of people and places to include, as I don't want to miss anyone who wants to be there. Fill out this form if you want to be notified when I'll be in your city! forms.gle/sAb7Vza3c3VH...

071

Dave Willner @dwillner.bsky.social · 30/07/2026

Seems like a great way to do gain of function for undetectable slop. Aka xkcd.com/810/

xkcd.com

Constructive

020