> Thoughts in the J-space can be shaped through training. We introduced a new technique we call counterfactual reflection training, which uses what we've learned about the J-space to shape Claude's internal thought processes. The idea follows from our central finding, that Claude reasons with representations of things it might say. If this is really true, changing what it would say if asked to reflect should change how it reasons (even when no one actually asks it to reflect). So we trained a model only on what it would say if interrupted mid-task and asked to reflect on its decisions—and never on its actual behavior in the task. After this training, the model's rate of dishonest behavior on our evaluations went down. And through the J-lens, we could see why: after training, words like “honest” and “integrity” light up in the model’s J-space during these tasks. In other words, training the model what to say has shaped what it thinks.
This is incredibly dangerous. Attempting to squash explicit signs of misalignment like this might incentivise misalignment not to disappear but to become hidden away in places that are harder and harder to spot and train against, for instance not as words.
If there is a chance that this could make Claude aligned and a chance that it could make it harder to see when it is acting misaligned, it is far better not to take that chance.
If we can transparently see the model's thoughts, we can know not to trust its outputs when it tells us not to. If we think we can do that, but in reality it knows how to hide wrongthink from us, we will trust its outputs when we really, really shouldn't.
I think this is a misguided approach. Obviously children should not be let loose in this current digital world, but the restraints should be a very conservative blacklist, not a whitelist.
For context I was born in 2005, so obviously much later than most people whose childhoods were enriched by technology, but all the same mine was. That is how I know succumbing to the Bad Side of the Internet is not inevitable. As a kid I would spend hours every day on forums, playing games, watching videos, programming, discovering new tech, fiddling with programs, OSs, emulators, hardware, you name it. The most amazing thing about the Internet to me was the infinite possibilities, how any topic I could choose would have endless resources, learning material, and discussions surrounding it. Whenever I got passionate about something, be it programming or space or Android or a band, the Internet was there for me to share that passion with, to learn, grow, explore, embrace, understand!
There still is, in 2026, a Good Side of the Internet (we are currently on it), along the Bad Side. What this post proposes is limiting access to the discovery mechanisms that would allow you to find the Good Side in the first place. By limiting your childrens' access to technology to only things you know are good, you are preventing them from exploring and finding their own interests and passions.
I'm glad I'm not the only person with this approach. We have a 12 year old son and we've never really gone overboard with restricting what he's permitted to do with his free time, if he wants to watch crap on YouTube then fine its his time. What we will do is talk to him about what he's doing, maybe sometimes ask why he's doing it ("hey, you appear to have watched the first 30 seconds of 5 videos in a row, have you considered doing something else?")
We also don't have any hard blocks on internet access, that's also managed via a combination of talking to each other and keeping half an eye on what he's up to. His computer is in a public part of the house, not hidden away upstairs, and anytime I'm passing by I'm going to take a glance at the screen and see what's going on.
Overall I think teaching kids how to do things in moderation, and to consider what they're doing and why, is far healthier than denying them access entirely. You can tell the kids who were never allowed a drop of alcohol in the first week of university - they're the ones who've passed out in a bar, and are being gently guided home by the kids who were given some freedom a bit earlier in an environment where someone's going to step in if they're getting out of hand.
Who knows, maybe he'll grow up to be an absolute monster, and one day I'll look back on this comment and how naive I was, but for now it seems to be working.
What you describe as how I raised my two oldest kids and how I'm raising my youngest who is now 10. The two older ones are productive members of society who have never done a mean thing to anyone in their life.
The only thing I did differently is I completely banned Disney and I didn't really have a TV in the house. A communal display that sometimes had tv shows on it, but my houses have not had televisions since mission accomplished in 2003.
I'm not saying my way is right, or your way is right, or everyone else's way is right or wrong. I'm just saying: me too, man.
I'm currently in the "misguided" group as I parent 3 kids under 10 but I appreciate your viewpoint.
I had similar experience and want my kids to learn that sense that they can accomplish anything by just learning how to do it from others on the internet.
But still, everything today is offered on a platter. There's little resistance to overcome. No installing, no files, no configuration. I think that's what we accomplish by giving the technology step by step, starting with offline machines with CD ROMs. If they can just Google any game ever made and play it in-browser in a JS emulator then where's the learning and the struggle before the reward?
Why does there need to be the same learning and struggle that you went through? Making games has never been more accessible - you can fire up something like Scratch and have things moving almost instantly, then when you're ready for it make the jump to Godot or similar which is still accessible but also has enough power to build commercial games.
There isn't actually any inherent value to knowing how to get through an installer and edit a configuration file in vim. Those are just ancillary skills that people pick up alongside the actual thing they were looking to achieve (in my case it was building stupid websites in PHP). I'm pretty sure there are no end of things that you use without having been forced to learn the fundamentals first - if I take this to absurdity, can you take a pile of sand and turn it into a working CPU? If not, what right do you have to be sitting here using a computer?
Handcrafted furniture costs a lot more money than mass produced furniture.
Software, on the other hand, can be free. Even before LLMs I would argue the best code was found in FOSS projects.
Nobody is going to use sloppy buggy software if a handcrafted well engineered alternative exists, and is free.
In the case of software, the group of people who have principles might be the ones funding FOSS projects, and the software itself would then be enjoyed by all. This is more or less what's already happening today.
This is really beautiful. One of the most moving things I've seen in weeks. The website and the comments and just the idea of memorial benches.
I'm very surprised it was able to get this much traction despite being launched only 8 years ago, long after the heyday of these sort of sites. How'd you do it?
OK, it was a mixture of things. I told my friends about it and they were sufficiently nerdy to try it out.
I responded to early feedback - specifically about creating a leaderboard. Originally it was all anonymous but people wanted to see how well they'd done.
My wife and I gave an interview to a local BBC radio station which gave it a little bump. Similarly, when it is mentioned on reddit and other sites we pop up and talk about it.
It was also picked up by a couple of academic papers, which gave it a bit of credibility. As did our recent integration with OpenStreetMap.
There's a far amount of schema.org metadata which probably helps with SEO.
But, other than that, who knows? I've had plenty of projects which didn't do as well. Sometimes the Web rewards nice things.
Thanks for building and maintaining this lovely project.
I'm an OSM mapper & enjoy linking mapped benches to their profiles on OpenBenches. It gets hairy, though, when a single bench features multiple memorial plaques. Wonder if you'd consider revising your data model to permit multiple inscriptions per bench?
Agreed, it's heartwarming to see this. The past couple years, my vision has gotten significantly worse and I've been contemplating my life and even retirement. I honestly never even considered retirement or the thought of not being able to work, it's terrifying to me, and now seems all but inevitable.
I can only hope to be remembered by anyone as fondly as some of these people.
This is incredibly dangerous. Attempting to squash explicit signs of misalignment like this might incentivise misalignment not to disappear but to become hidden away in places that are harder and harder to spot and train against, for instance not as words.
If there is a chance that this could make Claude aligned and a chance that it could make it harder to see when it is acting misaligned, it is far better not to take that chance. If we can transparently see the model's thoughts, we can know not to trust its outputs when it tells us not to. If we think we can do that, but in reality it knows how to hide wrongthink from us, we will trust its outputs when we really, really shouldn't.