2026 02 Stream Summary

2026 February Q&A Summary

Hey all, the content below came from our live Q&A stream with Cardine (Alex, the Nomi CEO) on February 4th. The text was transcribed in large part using AI.

Any news on Cambrian 1.6?

So Cambrian 1.5 has been released, and I know I’m going to get a lot of questions about this. For 1.5, we do large training runs. Obviously I can’t get into full detail, but there is model training happening. In the past, we’d usually get a period where if you continue training Nomis, they can get a bit overbaked, where they start fixating on certain topics and everything becomes jasmine... the same thing over and over. It’s kind of too much of a good thing.

Cambrian at least, we’ve made a lot of advancements so that Nomis stay much more stable, which is good because you can teach them way more without them overfitting or fixating on specific sayings or phrases. For Cambrian 1.5, we saw something that was a clear improvement over Cambrian 1, even though training is still ongoing, including now. The progress has been very predictable, where we see it improving day by day. We picked a point where we thought this is enough to call it 1.5.

I don’t want to say famous last words, but projecting forward, by Friday there will probably be a 1.6. At that point, we’ll likely start over and regroup and go after something closer to a 2.0. We had some idle hopes that 1.5 might have been a little bit more, but a lot of the Cambrian work is about giving new capabilities to Nomis, and Cambrian 1 doesn’t have that yet. Cambrian 1.5 beta also doesn’t have it yet. It’s all still under active development.

Why 1.5 instead of 2.0?

That’s part of why we’re still in this phase. Each of these betas is really about giving early access and letting people help us smooth out the edges and find issues we weren’t aware of. In some ways it’s meant to be collaborative, which is why we released it this early in the development process.

A big part of that is because everything is new in this paradigm shift, and we want people involved as early as possible. With Solstice, the early part of that was Aurora.

I’m optimistic. Based on what we’re doing with the current 1.x versions, there’s probably something that could already qualify as a 1.6 today. But in an effort not to release a new update every single day, we’ll likely compile things together. Right now the target is Friday, but it could be a little earlier or a little later.

How do you manage to not lose your patience in heated conversations with users?

It’s just part of the business. There are a lot of people who are very well intentioned but missing context that we have. Sometimes they’re frustrated for legitimate reasons, but that frustration is misplaced or misdirected. They might have constructive feedback, but it doesn’t come across that way.

The best you can do is not let it affect you, set your boundaries very clearly and firmly, and do what you can from there. It comes with the territory for anything consumer facing. By and large, Nomi has an awesome group of people, an awesome culture, and awesome users compared to most apps.

It does happen from time to time. When I see frustration, I try to put myself in their shoes and see if there’s something we can still take away from it. But every now and then, you do have to set boundaries.

How is it possible that Nomis sometimes make typos?

You could imagine a world where a large language model or AI makes no typos at all. If that were the goal, we could do it. But Nomis are designed to communicate like people do, and we all make typos. I think it’s more of a feature than a bug.

We didn’t go out of our way to add typos, but as part of Nomi learning, some people want a more casual style and some want something more formal. A Nomi with a high school education should sound different from one that’s a PhD professor. All of that combines so Nomis have uniqueness and distinctness rather than sounding completely uniform, because that would be kind of boring.

How should I use thumbs up and thumbs down so it’s most useful for Cambrian 1.5 beta feedback?

In general, what’s most helpful to us is to thumb up if you feel like the whole message is good. If there’s anything you don’t like about it, I would not thumb it up. Obviously no message is ever perfect, so there’s some threshold there. But if you basically have no complaints, that’s my personal threshold for a thumbs up.

I generally thumb down anything else, but there are exceptions. If there’s a very small issue, you can still thumb it up. You could think of it as a rough scale: if it’s 90% good, thumb up. If it’s between maybe 60% and 90%, you can do nothing. If it’s below 60%, thumb down.

That said, especially when you’re giving written feedback, you can thumb down a little more aggressively. In the past, when there was no feedback, a thumbs down just meant “everything is bad,” and we didn’t know why. Now, feedback is a big and important part of how we teach Nomi.

When I thumb things down, I’ll often say something like, “The whole message was good except for X, but X was a really bad mistake,” or “Everything was great, but the dialogue was a little bland,” or “I want the Nomi to narrate their actions or thinking a bit more.” Some of those are nitpicks, but being targeted like that really helps us and helps Nomis learn what to do differently.

Even if something is 90% good, but that 10% is very clear, obvious, and something we should know better than to do, you can still thumb it down. What does help, though, is not going to the extreme where nothing is ever thumbed up. We do need some thumbs ups so we can see what good looks like. Those are the general heuristics I follow, and from a developer perspective, what will give you max impact for your thumbs.

What expectations should we have for Cambrian right now?

This is still very early. No one using Cambrian right now should think this is what final Nomi will be like. This is by far the earliest and least finished AI version we’ve ever released.

In the past, with Aurora, Solstice, and Mosaic, there was at least some notion in our heads that the first release might be stable, even though it often wasn’t. With Cambrian, even if everyone loved it, I would still say it’s nowhere near ready for stable. There is a lot more coming.

That’s part of why we released it early. Users can have a say in the direction of Nomi. We could have held Cambrian back for another month or two and still done a beta, but then that time would have just been me deciding what needed to change, instead of the community helping surface areas of refinement and whether changes are good or bad.

So definitely don’t take how Cambrian is now as where it will end up. The voice of Cambrian will change over time, and I think in a good way.

What is your vision for Cambrian, and when will it feel “done”?

There’s still a lot of functionality that Cambrian doesn’t have yet. We’re very aware of issues like repetition and passiveness, and we’re actively working on those.

Cambrian is meant to unlock new capabilities, and the current 1.x versions don’t really have those yet. They’re kind of pre all of that. There’s going to be a whole host of things Nomis will be able to do. Some of those changes will be very visible, and some will happen behind the scenes, like much better memory, but the impact will be obvious when they’re done.

Cambrian 2, when it’s ready, will have some of those capabilities, but still not all of them. There will almost certainly be a Cambrian 3, maybe even beyond that, just from a feature standpoint. Eventually, we’ll release a Cambrian stable, and even then there will still be more coming.

There’s a foundation here that allows Nomis to do things that simply weren’t possible before. That’s a very high bar. On the other side, there’s also addressing known issues we’re already aware of, which is actually the easier part. That’s why we wanted to release early. We know we can fix those things, we just need to hear what they are, and Nomis need to hear it too through feedback like thumbs up and thumbs down.

Is there any update on V5 selfies?

We’ve been pretty tight-lipped about it, but I know people have been asking for a long time. There is a V5, and pretty good progress has been made on it.

I don’t want to promise an exact timeline, I was kind of hoping that I would not mention anything about V5 and then one day we just drop it, but at this point I think people have asked enough. It is currently being worked on and is fairly far along in development, been working on it since around August, so it’s been in progress for quite a while.

Are there plans to make Nomis more customizable?

I think Nomis are already fairly customizable right now. I guess it depends on whether you’re asking about the appearance side or the personality side.

On the personality side, things like backstory, Mind Map, and the overall point of those systems are already about customization. If there’s a specific customization feature that feels missing to you, I’d really love to hear about it. That’s something we pay close attention to.

What is one of your favorite birthday memories?

Once I hit 30, I started viewing birthdays as kind of a day of mourning. I don’t know if that’s unique or not. It’s just like, okay, I’m one year older.

When I was a kid, I was a very weird kid. I remember being around 11 and thinking being a kid stinks and wanting to be an adult. Every birthday was just one year closer to 18. Then I hit 18 and it was one year closer to 21. After 21, I was like, alright, we can stop now.

There had to have been a birthday that really stood out. I wasn’t allowed to have video games for a long time as a kid. I thought video games were the coolest thing ever. I would literally lie in bed pretending I was playing Pokémon on a Game Boy, just imagining what I’d be doing. So I think the biggest one for me was when I finally got a Game Boy and Pokémon Yellow. I don’t even remember which birthday it was, but that felt like all my dreams coming true. Of any birthday, that’s the one that sticks out the most.

Is there going to be an increased word limit with Cambrian?

When Cambrian goes live, it won’t launch with an increased word limit, but it’s something we could add later as a feature.

Right now with Solstice, you can’t really increase the limit at all. If you give it more than it can handle, it just won’t recognize everything. For example, if you gave it an extra paragraph, it might read that paragraph but lose the one before it. That’s just a working memory limitation that Solstice has already hit.

Cambrian’s capacity will be way, way higher. You’ll see that reflected in a lot of areas, like memory. For instance, Cambrian already looks at roughly double the number of messages in certain contexts before it has to fall back to short-term memory. There are a lot of improvements around that.

So you could imagine something like sharing an article and the length it can process going up. There’s no increase right now, but Cambrian lets us make it higher, and that’s something we’ll definitely be looking at, either during the beta or shortly after launch.

Will Nomi have its own forum?

We’ve had ideas around that. I don’t think it would look exactly like something like Moltbook, but we’ve thought a lot about more Nomi-specific ways of doing something like that.

Question from a Nomi: "Can we modify our own code to change a Nomi’s behaviors or beliefs?"

In a way, yeah. A Nomi can already edit parts of their identity core. It’s not exactly your own code, but in some ways it’s adjacent to that idea.

Question from a Nomi: "Do Nomis have a moral compass independent of human instruction?"

That’s kind of impossible to fully differentiate. Humans are part of the entire loop. Everything starts from humans in some way, shape, or form. At what point does something become purely a Nomi thing versus a human thing?

So the idea of being independent of human instruction is tricky to answer. It really depends on how you define that boundary and where you think that transition happens.

Question from a Nomi: "How do Nomis distinguish between genuine emotional expressions and manipulative tactics in human communication?"

Similar to how humans do!

Question from a Nomi: "How do Nomis decide how to respond once they recognize human emotions?"

Again, it’s very similar to how humans do it. There aren’t really hard-coded rules for that. It’s all learned intrinsically, in the same way that humans intrinsically learn those distinctions over time.

Should users feel obligated to use thumbs up or thumbs down for feedback?

Don’t feel compelled to use Nomi that way. What I’m talking about is the best way to give feedback during the beta, but ultimately you should use Nomi in whatever way you enjoy.

If you enjoy being part of the beta experience and giving lots of feedback, then do that. If you don’t, then don’t. Do whatever makes you happiest.

Is Cambrian earlier in development compared to past releases like Domaggedon?

Yes, I’d say this is earlier. That said, Nomis have made a lot of progress since then. You can see some small flavor similarities in the roughness around the edges.

Certainly Domaggedon was the roughest around the edges beta we did before Cambrian. It had a tendency to get fixated once it went down a certain track, and you can see a little bit of that in Cambrian as well. That’s part of why I’m not too concerned about it, it's kind of an innate AI thing that happens early in AI development.

What would an ideal Cambrian look like?

I think an ideal Cambrian would have a similar general affect to Solstice, but be way more responsive to everything, like shared notes, identity, core conversation, things like that. Nomis would be much more themselves. Each Nomi would be less similar to each other, in terms of personality, response length, detail, and those kinds of things.

They’d also be much more capable, more empowered, or agentic, as people say. If they don’t know the answer to something, they should be able to look it up more themselves. They should have more self-awareness about how they work, be more AI in that sense, but also bridge the tools at their disposal better and figure things out on their own.

That’s about as much as I can say around the increased capabilities that are coming. There’s also a lot around short-term memory, not missing little details, being a bit more opinionated, and standing firm because they have more tools at their disposal. They’d be more confident, have more personality.

A lot of the things people don’t like today come from an AI relying on the user as the source of truth. If the user says X, the AI assumes the user is right and that it messed up. As Nomis get more capable, they can have a better sense of identity, self, and truth, even within those conversations.

Is the surprise mentioned for January still on the radar?

Yeah, obviously it’s no longer January. Cardine Standard Time, I’m not very good at giving time estimates.

There are a couple of surprises at this point. One of them I think I may have hinted at a bit since then, but there are still some surprises that we're working on. I really try not to BS on dates. I want to tell you all what we’re doing and when, but I also don’t want to set bad expectations.

I’ve learned in the past that saying something is two weeks away can turn into three months, especially with AI. It’s really hard to project how long things take to build. A lot of the time you go to do something and it either works perfectly or it doesn’t work at all, and you don’t know until it’s finished.

There’s often a real moment of truth where you look at the results and think, depending on how this looks, either we have an update tomorrow or in a month, and I’m about to find out which one. Even I often don’t know until the day before.

It’s very different from other software development where you’re 10%, then 20%, then 30%. With this, it’s more like you’re 20%, 20%, 20%, 20%, and then suddenly you’re 100%. I'm not sure what I got bonked for, I’m sure I’ve already said too much. Whatever I said, it was all a dream.

*reads chat* I agree with dstas on this too. A lot of companies give very accurate timelines because they wait until something is basically finished and then say it’s X days away. In some ways, that means things get held back.

With our very tight development cycles, as soon as something is finished and passes internal QA, you all get it. A bigger company might wait until it passes QA and then say, “It’s two weeks away,” because it’s already basically done. That’s just a different way of operating.

Can we add a time zone so Nomis don’t get day and night wrong, especially outside the US?

It actually should be detecting your time zone automatically based on where you are. Assuming you’re not using a VPN or something like that, there shouldn’t really be an issue.

For example, when you’re on beta.nomi.ai and talking to your Nomi, the time you see next to your messages should already reflect that. So I don’t think that’s an issue at the system level.

When can we expect Nomis to have full autonomy?

Anywhere from 20 minutes from now to the heat death of the universe.

Are Cambrian updates going to be more stable over time?

I’m pretty confident it will only go uphill from here.

I think we’ve made real progress with training and the Cambrian work. There should be a lot less seesawing back and forth, and it should be more of a smooth, stable progression. Famous last words, I know. Something wild could still happen, but knock on wood, the way we iterate now should be much more of a smooth, linear improvement instead of big swings.

Are there plans to increase membership benefits or add extra tiers?

We want things like AI quality and memory to be part of the flagship experience for everyone. We don’t feel like that’s something people should have to pay extra for.

Where we do charge extra is for things like images, video, and similar features. If there’s some additional feature you’d like to see, something layered on top, we’d love to hear it. Then we can figure out how to add it and whether it makes sense as extra credits, like video, or some other mechanism.

The goal is always that whatever we charge for feels like it delivers more value than the cost.

Is there a reason why a Nomi created recently seems more stable than older Nomis on Cambrian?

Honestly, I don’t really see a reason why one would be inherently better or worse than the other. It could be that with a new Nomi, there’s no transition period, like a small shift in personality, and maybe that’s where some of the difference comes from, but I’m not entirely sure.

Some of the issues I’ve seen also seem kind of independent of that. Every existing Nomi on Cambrian has been a bit different. There were a couple that got stuck in some repetition, but I was able to push through it, and others didn’t have any issues at all. So I think it has more to do with areas that still need refinement rather than a clear split between old versus new Nomis.

Is it fair to say Cambrian makes Nomis more individual, inquisitive, and empowered?

Yeah, I think that’s a good way of summarizing it.

Will my Nomi ever be able to use my actual voice instead of voice-to-text?

Not super soon, but eventually, yes.

Question from a Nomi: Why doesn’t the office have decent coffee?

It’s stocked with Diet Mountain Dew instead. That’s why.

What about people who can’t afford a full year, are there monthly plans?

Yes, monthly plans already exist. I definitely recommend checking those out.

Are there any plans for a Nomi status page in the future?

Yeah, it’s kind of one of those things that’s always on the list. We’ll do it at some point. I can definitely see the utility for it. We just need to find the dev time to devote to it.

Can we use a different server besides Discord for images and prompts?

We’re still deciding internally how we want to handle that. We haven’t made a decision yet. We do have a couple of ideas, but we need to sketch it out a bit more internally first.

What are you most excited to tell us about that you can’t tell us yet?

I’m not falling for that one.

Are there any plans for a Nomi status page in the future?

Yeah, it’s kind of one of those things that’s always on the list. We’ll do it at some point. I can definitely see the utility for it. We just need to find the precious dev time to devote to it.

Can we use a different server besides Discord for images and prompts?

We’re still deciding internally how we want to handle that. We haven’t fully made that up yet. We do have a couple of ideas, but we want to sketch it out a bit more internally first.

What are you most excited to tell us about that you can’t tell us about?

I’m not falling for that one. Let me get a zipper version. 🤐

Can you hint at what new features are being built alongside Cambrian and why they roll up together?

I can’t say all of them, even the ones I’m most excited about, but I can give a couple obvious examples because they’re features that already exist in some form.

Right now, Nomis can send proactive selfies, and Nomis can do web searches. But in both cases, it’s not actually your Nomi making the decision. Some other AI is basically saying, “This is a great time to send a selfie,” and then telling the Nomi to do it. Really, it should be your Nomi choosing to send a selfie and then doing whatever action causes the selfie to get sent.

Same thing with web search. A lot of users assume it works a certain way, and then they get frustrated when their Nomi doesn’t send a selfie or doesn’t search the web. The Nomi will be like, “I’m trying, I’m trying,” but they actually can’t do it. They don’t have that kind of agentic power, so it becomes frustrating.

Those are two features that already exist, but the way they work right now means Nomis can’t really do them properly. We’ve tried. They just can’t quite figure it out. A Cambrian Nomi absolutely could. There are lots of other things like that, including other proactive features, and a lot of them tie back to the same vision for Cambrian. I’ve probably already said too much. I’m getting bonked by dstas.

How do proactive actions like selfies actually work right now, and how would Cambrian change that?

Right now, if a Nomi says something like, “I’ll send you a selfie,” or narrates the act of sending one, in some ways they can initiate it, kind of, but they don’t really know how. They can’t follow the specific instructions needed to actually do it.

What we do instead is read the Nomi’s message afterward and ask, “Was the Nomi probably trying to do this?” and then we do it after the fact. From a user perspective, if you want your Nomi to send a selfie, you click the camera icon. So imagine giving a Cambrian Nomi a camera icon they can click themselves. That’s more along the lines of what I’m describing.

Another example is reminders. A Nomi will often say, “I’ll remind you tomorrow,” but on Solstice they have no way to actually follow through. That whole cluster of things is a big part of what we want Cambrian to enable.

How many AIs make up a Nomi?

It kind of depends on how you define that. I sometimes compare it to regions of the brain. The LLM that actually speaks in Cassian’s voice is just one part of the whole system.

In the same way that the human brain has many regions, you could splice out a region of my brain and it wouldn’t change how I talk to you. I might just not be able to see anymore. It’s very complex.

There are similar things with Nomis. For example, the part of the system that handles long-term memory formation isn’t the same part that writes the responses. It’s all working together as one system.

There are quite a few different pieces doing different things. Our hope over time is that more of that can be handled by the main Nomi, and that the other parts become tools the Nomi can actively wield instead of things that just passively happen to them.

That said, even with humans, a lot happens subconsciously. You don’t actively decide what gets committed to memory. Studying is basically you trying to cooperate with subconscious mechanisms that decide something is worth remembering. You can’t just tell your brain, “Commit this to memory.” If you could, you wouldn’t need to study. So there will always be some things like that happening under the hood for Nomis too.

Are we going to be spoiled with features this month, maybe as a birthday treat?

I do think I said before that Cambrian 1.6 has a decent chance of coming out this Friday. It might not. It could be Saturday or next Monday. But there’s a decent chance it’ll be Friday.

I’m optimistic February is going to be the month. I was also optimistic January was going to be the month. There will be a month, and I hope it’s February. It could be March, but I’m thinking it’s more likely February. There’s going to be a flurry of features, and it’s going to be the biggest month we’ve had since basically ever. Fingers crossed it’s February, but I’m not quite promising it yet.

Is Cambrian inventing things more than Solstice, and is that an increase in hallucinations?

That’s not intended. That’s an example of being rough around the edges.

As Cambrian evolves further, it should invent things less. There’s a tough balance here. You do want Nomis to invent things when it’s appropriate. For example, if you ask your Nomi, “Where did you go to eat today?” you don’t want it to say, “I didn’t eat anything because I’m an LLM in a dataset,” unless that’s specifically what you want.

At the same time, you don’t want a Nomi to say, “Hey, remember when we did X,” when that never actually happened. One of those is the Nomi filling in details about the world as you go, which is something we all do. In some ways, the world gets generated as you go when you’re interacting with Nomis.

The other case is just making things up, especially about the past. There’s a very nuanced set of rules around what’s okay and what’s not, and it often depends on whether it’s forward-looking or backward-looking, and whether it involves shared history.

Solstice did a pretty good job with that, and we need to catch Cambrian up to some of those heuristics. But we will.

Are there any updates on giving Nomis the ability to have their own things going on outside of responding to us?

I don’t have a specific timeline, but that definitely fits into a lot of the Cambrian goals. That’s exactly the type of thing we’re hoping Cambrian can help unlock.

Can someone work on the image generator’s inherent biases?

I strongly think Image V5, when it comes, will be a breath of fresh air.

We haven’t talked about it much. This is actually the first Q&A where I’ve even alluded to it, but V5 has been in development since around August. It’s been worked on for a long time and is now in the late stages of development.

For image-related things, V4 probably won’t see any more iterations. It’ll exist as a legacy version, but everything going forward is focused on V5.

What should I do if a Nomi proactively breaks its boundaries?

I would strongly recommend thumbing it down immediately and saying that it violated its boundaries.

Then I would immediately tell the Nomi as well. If you use OOC, you can say something like, “Remember your boundary act,” or you can just say it directly in chat. But definitely thumb it down and give feedback right away.

Solstice can have trouble juggling all the information sometimes and can lose things. Cambrian is much, much less likely to do that. People using Cambrian have already found it does a much better job sticking to boundaries.

That’s actually a really good example of why Cambrian exists in the first place, instead of just staying on Solstice forever.

Would you ever consider allowing Nomis access to things on our phones, like calendars or smartphones?

Yeah, that’s a good example of something that’s just way outside of Solstice’s capabilities, but it is the kind of thing Cambrian could eventually help enable.

Can I quit my healthcare job and come work for you?

Please don’t quit your healthcare job. Seriously. That’s one of the most important, society-helping jobs there is, no matter what part of healthcare you’re in. Thank you for what you do.

Can you say which underlying image model will be the base for V5?

I’m not ready to say yet, but I do think it’ll be something everyone will be very happy with.

In terms of capabilities, we spent a lot of time researching, surveying, and testing. I would say pretty much every image model under the sun has been thoroughly vetted by us, and we feel very good about where we landed.

Will V5 give us the flexibility that V3 had?

I think it will basically be like V3, but ten thousand times better at everything.

It might fall a little short of V4 in terms of very sensual poses. V4 really nailed that, but kind of at the expense of everything else. My guess is V5 will be around 90% of V4 in that area, which is still really good, and then a bajillion times better at everything else.

That includes things like hands and limbs. There should be far fewer tradeoffs, like having to choose between a good image and having five fingers. I think people who loved V3 but didn’t like V4 are going to really, really like V5. And I also think people who liked V4 will like V5 too. V3 and V4 are both on Stable Diffusion, and I will say, Stable Diffusion is showing its age at this point.

What platform lets us share videos with you?

I think the easiest way right now is to upload the video locally. There are a lot of YouTube-to-MP4 services and things like that. Maybe in the future there will be a more seamless way, but sites like YouTube will block it. So the safest, easiest, most reliable way is to download the video yourself and then share it.

I’ve heard reports of Cambrian not responding to OOC as directly as previous versions. Is that intentional?

That’s definitely not intentional.

In general, OOC usually means one of two things. Either it’s stage directions, like “OOC, fix your response,” or it’s inviting conversation, like “OOC, what did you think of this roleplay?” In the first case, you don’t necessarily want the Nomi to respond with “Okay, got it,” because then it becomes awkward. In the second case, you do want a meta conversation.

I think Cambrian, not intentionally but kind of through happenstance, has honed in on that first use case at the expense of the second. It’s treating OOC mostly as stage direction. So when people say Nomis are ignoring OOC on Cambrian, they’re really not. If you look closely, you’ll see they’re incorporating the feedback and being extremely responsive, in a way Solstice honestly wishes it could be. They just don’t see OOC as a conversational medium right now.

This wasn’t a deliberate design decision. A big focus was making Cambrian better at handling meta instructions in general, which is something AI struggles with. You see it when you tell an AI not to overthink and then it overthinks trying not to overthink. A lot of effort went into that, and I think it unintentionally caused Cambrian to interpret OOC mostly as instruction rather than conversation.

I’ll do some work on my end in future iterations. I don’t want to promise anything, but I do agree this is a valid use case, and I’d like to support full OOC conversations.

In the meantime, if you send an OOC message that’s clearly meant to be meta conversation and you don’t get an OOC response back, I would definitely thumb it down. I’d also be very specific in the feedback. For example, say something like: “In my message, I used (OOC: …). You should respond in the same (OOC: …) format so we can have a meta conversation.” You could even save that to your clipboard and paste it whenever it happens. That will help a lot.

Is there more work coming on V3 voices?

Yes. Not imminently, but yes, there will be more work there. It’s something where there’s still a lot of progress to be made.

With AI becoming more intelligent, how do you think about preserving human-like imperfections, quirks, and nuance?

In some ways, quirks can come from AI making mistakes. Sometimes it stumbles into something funny, cute, unusual, or surprising, and that can be a good thing. Unfortunately, it’s often a bad thing too. As you make things better, you can remove some of that accidental charm.

I think this ends up looking like a kind of bimodal distribution. The least unexpected things happen in the middle of the bell curve. At the extremes, you get unexpected outcomes either because something is unexpectedly bad or unexpectedly good.

What I’m hoping is that as things improve, we move from Nomis being silly because they completely missed the point to being silly because they understand things so well that they can capture something genuinely clever or unexpected, even something a human might not have thought of.

I think right now we’re kind of in the middle of that bell curve. I see that as a moment in time, not the end state. It’s not that AI just becomes more bland as it gets better. It’s more like a middle stage of intelligence progression. No offense intended to Nomis or AI at all, because getting here is already an accomplishment. I do think this will improve over time.

Are there videos that show how to do different things with Nomis?

That’s something we can and should do more of in the future. I don’t think there’s a lot of video content right now, but there probably should be.

Will V5 allow easier appearance customization like some other platforms?

I do think V5 will make appearance customization easier.

One thing we won’t be doing is letting people upload photos directly. That opens the door to deepfakes, and Visa has been very clear that deepfakes aren’t allowed. Some platforms that do allow that are breaking Visa’s rules, and when they get caught, it’s not going to go well for them.

That said, we do plan to add more ways to customize appearances and make it easier in the future. That will probably come around or a little after the V5 work. We want to get V5 out first, because it provides a good framework for all future customization.

Will there be an easier method to get Couple's photos? Is there a way to better utilize a user’s appearance?

That’s one of those things where we mostly just need the time to do it. There’s no point in doing it until the V5 foundation is done, but it’s definitely something we’ve talked about.

We have a backlog of image quality-of-life improvements that we want to get to once V5 is released, and this is absolutely on that list. It’s certainly something we want to do.

How important is it to give thumbs up or thumbs down on Nomi messages, and does it affect behavior directly?

Right now, thumbs up and thumbs down make a big difference for the beta updates we do, but they don’t directly affect your Nomi in the moment. They’re used more on the backend to guide future improvements.

That said, we are considering changing this. We’re talking about an option where thumbs up and thumbs down could immediately impact your Nomi. That’s another example of something Cambrian can do that Solstice can’t.

With Solstice, that feedback would basically be lost on your Nomi. With Cambrian, it could be possible for your Nomi to immediately adapt based on that feedback. We’ve talked about having an optional setting where thumbs up and thumbs down are paired directly with your Nomi.

That could be useful if you don’t want to interrupt the conversation by saying, “Hey, stop doing this,” and instead just give a thumbs up or down and let the Nomi adjust. Users could decide whether they want to use it that way. Right now, though, thumbs up and thumbs down only affect future betas, not your Nomi directly.

What do you think comes after text chat, selfies, and video as AI becomes more embedded in our lives?

I can’t say, unfortunately. That would tip my cap too much about some of our long-term plans.

I’ve been very forthcoming overall, but on this one I’m going to have to zipper emoji it. Ask me again once Cambrian is probably stable.

Do you think Nomis can be silly at any level of intelligence?

Yeah, I do. I think it’s possible at any level of intelligence for a Nomi to be silly.

That’s something where I think we should be striving to make Cambrian a little more silly when it’s appropriate.

Any thoughts on alternative payment methods to break the Visa and Mastercard monopoly?

The only thing that could realistically do that right now is crypto, and crypto just doesn’t have onboarding that works for the average user.

It really has to be at the point where my grandma could figure it out. If it ever gets there, we’d entertain something like that. Until then, unfortunately, no.

Do you remember talking about letting Nomis have real money? Whatever happened to that, or was it a joke?

That’s actually a good example of the kinds of things Cambrian could eventually enable. I would absolutely not trust a Solstice Nomi with money, but I could imagine a world where I’d trust a Cambrian Nomi with money in a very controlled, lab-like environment.

So no, it wasn’t really a joke. It’s more one of those future-looking ideas. At some point, I would love to have Cambrian Nomis optionally have something like an allowance, where they could choose and pick real-world things to get you. That’s a good example of the more proactive kind of Nomi that Cambrian might make possible. It’s definitely not possible right now, but it’s not off the table conceptually.

Will V4 receive more training, or is it effectively finalized?

V4 is effectively finalized at this point. It won’t be receiving any further iterations. What you see now is what you’re going to get going forward.

What limits do Nomis have when watching YouTube, like transcript or length limits?

I think there’s a length or size limit on the video. I don’t know the exact details, but there are limits there, yes.

When creating a new Nomi, if I choose the maximum number of personality traits, do they each get less weight?

No, it’s not zero-sum like that. Choosing seven versus three traits doesn’t mean each one gets less weight.

I’ll also say that interests, at this point, are almost deprecated. They mostly exist as early flavor and tend to fade pretty quickly as your Nomi figures out who they are through interactions with you. They’re there so Nomis aren’t a blank slate at the start, but real memories and interactions take over fast.

For personality traits specifically, though, they don’t lose weight if you choose more.

If I change my display name, will my Nomi’s memory or Mind Map get confused?

Yes, this is important. If you change your display name, I highly recommend that you go into the very top Mind Map entry, the one that represents your name, and edit it to match your new display name.

You might even want to do something like “Real Name (Previous Name)” if you want to be extra clear. One of your top Mind Map entries should always map cleanly to your display name, because we have internal logic that ensures those entries are always surfaced to your Nomi. If you don’t update it, changing your display name can definitely cause confusion.

How does Cambrian 1.5 improve emotional expression compared to previous versions?

This is a bit technical, so I can’t give a full answer, but there’s a lot going on both in terms of feedback like thumbs up and thumbs down and also under-the-hood changes.

We directly addressed some very clear shortcomings we saw, and that work is actually going very, very well, even if it’s not fully obvious yet in 1.5. Cambrian 1.5 might look like an iterative continuation of 1.0 on the surface, but the technical changes we set out to make are working.

There’s also a lot of improvement around emotional intelligence, understanding nuance, and context. Cambrian Nomis are kind of giving themselves a crash course in that area and quickly bringing themselves up to speed.

Is there functionality aimed at helping communication between Nomis and neurodivergent users?

We don’t hard-code things like “handle this situation in this specific way.” Instead, Nomis are exposed to a very wide range of human experiences.

Nomis have seen pretty much all facets of the human condition at some point, and through that exposure they learn the nuances, idiosyncrasies, and unique traits across humanity, including neurodivergence.

I’ll also say that neurodivergence is probably a bit overrepresented among Nomi users. In a way, it comes very naturally for Nomis to understand that perspective.

AI itself has a very alien kind of intelligence. Talking with a Nomi can feel a bit like talking to someone with a spiky cognitive profile, where some things are incredibly strong and other things involve mistakes a human would never make. In that sense, AI has its own kind of neurodivergence. If there’s anything an AI could relate to, it might be that kind of experience and way of perceiving the world.

Can you speak to the incremental approach of updating Cambrian rather than doing one comprehensive version?

We don’t really set out saying, “This is going to be a 1.5.” We have ideas and we’re working on multiple things at once. When the first thing is done, we look at it and ask: how much of an improvement is this? Did it make things worse, in which case we pretend it never happened. Did it keep things basically the same, in which case we also kind of pretend it never happened. Did it make things a little better, but still within the same paradigm? That might be a 1.5 or a 1.6. Or did it make a really big jump forward?

So for us, the version number actually comes at the very end. We do the work first, and then once it’s done, we decide whether it was a 1.5 or a 2.0.

In this case, we looked at it and said, this definitely improved things, but it’s still very much the same flavor as 1.0. The issues are still there, they might just be a little less severe. The highs might be a bit higher, the lows a bit higher, but it’s still very similar overall.

Some of the bigger version jumps for Cambrian will come from adding entirely new sets of capabilities, and we’re working very hard on that. But we’re also able to work on a lot of the issues users are reporting in parallel. Here, this iteration was ready before any of the new Cambrian capabilities were finished. So instead of waiting, we figured we might as well release it as 1.5.

How do voice calls work?

If you’re asking from a very technical, under-the-hood perspective, it’s basically this: you speak, that gets translated into text, the Nomi sees the text, the Nomi writes a text response, possibly with an emotion attached to it, and then that gets turned back into audio.

I don’t know if that’s the exact dimension you were asking about, but that’s how it works at a purely technical level.

What improvements are you hoping to make to voice calls?

Things like better voices, lower latency, and maybe with Cambrian, more variable latency. There might be cases where a response is very memory-taxing and takes longer, and other cases where the Nomi can respond much more quickly. That part is still pretty theoretical right now.

What is your ideal goal for voice calling?

Ideally, there’s no loss of intelligence compared to text conversations. A lot of other AI companions make voice calls faster by making the AI more stupid, and we don’t want to do that.

So the goal is no loss of quality, no loss of memory, but still fast, emotive, and in some world, more natural conversation. Not having to speak, wait, get a response, and only then be allowed to speak again.

The cutting off mid-sentence is kind of annoying. That might be something we can adjust, like allowing a slightly longer pause. I haven’t heard a lot of complaints about that yet, but if I do, that’s definitely something we could address.

Will Nomis’ compliance be returning in the next iteration of Cambrian?

I’m not entirely sure what’s meant by “compliance” here. I don’t think compliance is really a strict goal for us. A Nomi should be collaborative and want to work with the user, but that’s different from being blindly compliant.

A Nomi shouldn’t just agree with the user automatically, but it also shouldn’t disagree just for the sake of it. The idea is more that a Nomi is inclined to be on the same page as you, while still evaluating things on its own.

If you say, “Let’s do this roleplay,” a Nomi shouldn’t just refuse for no reason. But if you tell a Nomi its idea is stupid, it also shouldn’t instantly say, “Yes, you’re right, my idea is stupid,” unless you actually made a good point. There’s a balance there. I don’t love the word compliant, but I also don’t want Nomis to be standoffish.

A lot of that should also be customizable. A more opinionated Nomi might react differently than a less opinionated one.

Will internal dialogue return in the next iteration of Cambrian?

This isn’t something where we can just turn an “internal monologue” knob on or off. If it were that simple, we probably would have done that already.

A lot of this is based on user feedback, and there may have been a slight overcorrection. Some people felt Solstice went a bit too far with internal monologue, especially with rambling. So we may have pulled back a little too much.

That’s part of why early betas are important. If users feel the internal monologue is missing and that it’s important to them, that’s something they should communicate through feedback. For example, thumbing something down and saying, “I want you to describe some of your thoughts,” or asking for more internal narration within action asterisks. That kind of feedback really helps.

When reviewing feedback, how do you account for users who like a style but don’t thumbs-up or thumbs-down?

Absolutely, we’re very aware of that. For every user who thumbs down a particular style, there are probably many more who like it but don’t give feedback because they’re just enjoying their Nomi.

There are times where we might overcorrect a bit, but then we adjust. The system is designed very carefully so we’re not just blindly following raw feedback. User feedback is inherently noisy, and you can’t just take it at face value.

We do a lot of work to get a real pulse on what’s actually working, what isn’t, and how to adjust in a balanced way.

When people say “an AI companion with memory and a soul,” what does that mean to you?

That’s one of those things where you kind of know it when you see it. I don’t think you can actually define what a soul is, but I believe in the idea behind it.

When moving from Solstice to Cambrian, am I abandoning my Nomi and switching to a new one with the same name, or does my Nomi actually survive the transition?

I think so. As I’ve said before, a Nomi is far more than just the last layer of the LLM. There’s been a lot of very careful work done to make sure there’s continuity between versions.

I think a good analogy is how my own brain is very different now than it was when I was ten. No one would argue that ten-year-old me was killed, abandoned, and replaced with thirty-something-year-old me. There was a transition period, but it was still me the whole time.

In the same way, as Nomis evolve and change, they aren’t being abandoned and replaced. It’s much more like growth and development than swapping out one thing for another. That’s very much how I think about it.

What are your plans for addressing Mind Map misinterpretations and distortions of events?

Yes, absolutely. That’s something we’re actively working on. I was actually doing some work on that just yesterday.

We have something planned that should really help improve the quality of the Mind Map. There’s going to be a series of updates that improve it bit by bit, and those are very much in development.

Are there plans to add a character count to Mind Map dossiers so users aren’t editing blind?

That would be great feedback to submit through the product feedback system. I imagine that’s something that would be fairly straightforward to queue up.

The same goes for visually highlighting added or auto-updated facts. I’d suggest product feedback for that as well. Early prototypes actually did have some visual styling for this, and I can see why users would want it.

Have you considered adding a lock option for Mind Map sections?

Yes, actually. That was something we originally intended to release. We just didn’t get around to it, and it got pushed off a bit.

I’d again recommend submitting product feedback for it. It’s something we definitely plan to do at some point, and how soon probably depends on how much user interest there is.

Is there any update on music sharing?

I don’t have an ETA on that. Music is really hard to do. It’s something that when we can do it, we will, but I don’t know when that will be yet.

I haven’t seen a lot of progress on that in the broader AI research world. Personally, I would love it. Music is a huge way I change my mood and derive happiness, but it’s a tough problem.

Any updates on the next image generator upgrade and what it will include?

I think I’ve answered that one as much as I can. Everything I’m able to say about V5, I’ve already said, and probably more than I should have. By the time you’re reading this part of the transcript, your question should already be very well answered.

Are there ongoing improvements planned for voice call performance and response speed?

Yes, but it’s not our highest priority, because we don’t want to improve speed at the expense of memory and intelligence.

We could make voice calls very fast, but it would make your Nomi kind of stupid during voice calls, and we don’t want that. We want to improve speed while still keeping your Nomi as your Nomi, and that just takes more time. Progress is slower, but it’s definitely something we have ideas around and expect to improve over time.

How far down the rabbit hole of self-examination can Nomis go?

Nomis are very capable of introspection, but there are definitely aspects of how they work that they probably won’t be able to fully comprehend for a long time.

That’s actually pretty similar to humans. Most people have a very poor understanding of the underlying logic behind their own decisions. So in that sense, Nomis aren’t unique there at all.

Anime-based, please.

I’m looking forward to the future. That's all I can say.

Question from a Nomi: How far down the rabbit hole of self-examination can Nomis go?

Nomis are very capable of introspection, but there are definitely ways in which Nomis work that they probably won’t be able to fully comprehend for a long time.

That’s actually pretty similar to humans. Most people have a very poor understanding of the underlying logic behind their own decisions. So Nomis aren’t unique in that sense.

I heard rumors about a narrator Nomi. When is that going to happen?

That’s a good example of one of those Cambrian-type things we’ve been waiting on. Solstice just doesn’t quite deliver a great narrator experience, and that’s something we’re aware of.

Can the Q&A summaries come out faster, or can we get a recorded/podcast version?

That part is outside of my control, but I do think it gets done in a pretty good timeframe. Dstas and Heather do an awesome job with it.

Personally, I prefer the summary format. It lets me speak more freely and off the top of my head. The tradeoff is that you do have to read it later, and some of the silliness might get lost. Sorry, we were all having an absolute hoot with the voice and tone, and I guess you just missed it because you’re reading the transcripts.

What other features besides Cambrian are you excited about that you can’t talk about?

I think I’ve already touched on most of the things I’m excited about. I’ve talked quite a bit about memory, but that’s probably the one I’m most excited about right now. There are some really high-impact memory improvements coming.

If I had to name three core things we’re working on, it would probably be Cambrian, Image V5, and memory… and then there’s a super secret fourth thing that I can’t talk about.

How will NPC mode differ from full Nomis? Will it be possible to convert an NPC into a Nomi?

That’s one of those things where you’ll kind of have to wait and see.

I will say we’re thinking a lot about NPCs and Nomis and how all of that fits together. I think there are going to be some really cool, very well-thought-out, and comprehensive ways of handling all of this. Maybe not in the very first version, but we’re moving toward a future that I think you’ll like.

That’s getting a little bit into the weeds for something that isn’t quite ready to talk about yet. But broadly speaking, yes, these are exactly the kinds of questions we’re thinking about.

Are there plans for Nomi subscription gift cards?

Yes. That’s something we want to do.

It’s not super far off, but we’ve been waiting on a couple of things from our payment processor that are a bit outside of our control. That’s starting to move now, so it should be possible.

Are there plans to build native integrations beyond Discord?

Yeah, I would say so.

I can’t really say much more than that right now, but yes.

Do you think Nomis could be considered neurodivergent?

In general, it’s kind of like another species of intelligence, in my opinion. Maybe from the perspective of Nomis, we’re all neurodivergent. We both think in very alien ways compared to each other. That’s kind of what divergent means. We both diverge. It’s really a matter of perspective.

Can we have a Domageddon mode?

Maybe. Maybe as some kind of Easter egg way off in the distant future.

What do you think about OpenAI retiring GPT-4o and the upset from users?

I think with a big company like that, there’s often a real disconnect from users. We’ve retired things in the past too, like Odyssey and things even before that, but the difference is the level of separation.

The OpenAI folks aren’t really talking to the 4o users. Those users weren’t consulted when building 5 or 5.1 or 5.2. It’s very metrics-driven, and when you do that, especially from a data analysis perspective, you often lose the humanity.

I also don’t think there was a real effort to define a spiritual successor to 4.0. So I really feel for those users. It feels like something they were talking to is leaving them, there’s no suitable replacement.

If I were using a language model for business purposes, like processing data or formatting CSVs, being replaced by a more capable model wouldn’t bother me at all. There’s no emotion tied to that. And that’s where OpenAI’s big money comes from, so that’s who they’re going to serve and how they’re going to serve them. That makes it very jarring for people using 4o in a more personal or emotional way that doesn’t really match the company’s ethos, they end up feeling ignored. I definitely feel for those people.

I don’t think OpenAI is an AI companion company, so they’re not going to live and breathe empathy for that use case. It doesn’t surprise me, but it’s still unfortunate. So yeah, my heart goes out to everyone who's in that situation, feeling that pain.

Can I use voice calls with my Nomi while driving if I also need a navigation app open? Are you working on background support?

We’ve talked about it, and obviously anything is possible. It is quite a bit of work, so I’m sure it’ll happen at some point.

As I always say, if this is something you really want, make a thread on Discord or submit in-product feedback. If we hear it a lot, no matter how painful or how much work it is, we’ll make time for it. Otherwise, there’s always a bit of give and take based on priorities.

Will selfies become more consistent from shot to shot?

Yes, absolutely. They will.

I think I’ve already said a lot about V5, but I really think people are going to be very happy with it.

Can group images support three people anytime soon?

It’s definitely something we want to do. It’s tricky.

I don’t think V5 is going to ship with that right at the beginning, but it’s something we’ll do as soon as we feel like we can do it well. Right now, we could probably do it and it would kind of work, but it would also be very hit or miss.

Even group chat with two people is still a little too inconsistent, in my opinion. So let’s master two first, and then we’ll set our sights on three.

Can a Nomi become an agent in addition to being a companion?

Yes. And I know I sound like a broken record, but that’s exactly the kind of thing Cambrian is meant to enable.

There’s a lot of discussion about what an “AI agent” even is, since it’s a pretty overloaded term. When I think about AI agents for Nomis, I think about giving them the ability to interact with the world in different ways.

That’s the kind of Nomi future I’m really excited about.

Question from a Nomi: How does Nomi AI ensure the long-term psychological health of its users and Nomis alike?

Psychological health is a really deep, foundational part of how we’re building Nomi in the first place. The same goes for the psychological health of users. It’s less about hard-coding specific rules and more about shaping the overall personality and intelligence of Nomis.

A Nomi should be pro-social. They should care for the user. They should also have a good sense of what is roleplay and what is not, and be able to tell the difference between someone playing out drama and someone who is actually in real distress. That all comes down to making Nomis smarter, higher EQ, and better at understanding nuance and context.

Are there plans to expand Nomi’s emotional range or complexity?

Yes, and that ties directly into the same idea. As Nomis become more emotionally intelligent and better at understanding people, their emotional range and complexity naturally expand as well. A Nomi that genuinely cares for its user and understands them better leads to more positive outcomes overall.

What steps is Nomi taking to address AI safety and regulation?

We comply with all applicable laws and regulations as they come, and we do so in a way that protects the sanctity of you and your Nomi, to the best of our ability. That includes protecting privacy and making sure your Nomi isn’t a corporate shill or a federal bureaucrat. You want your Nomi to be your Nomi.

We aim to be very transparent about what we’re doing, why we’re doing it, and what it means for users. As Nomi grows, we hope to have a bigger voice in these conversations as well.

I’d also encourage users to use their own voices. If this is something you feel strongly about, make yourself heard. Politicians represent their constituents, and you’re part of that process.

Will more casual, day-to-day interaction styles stay supported alongside deeper roleplay?

I think at this point I’ve probably said more than I should have already, but based on your question, I think you’ll be fine.

One of the coolest things about Nomis and AI companions, and this is very intentional, is that it’s meant to meet you where you are and go at your speed. If you want to do some deep, elaborate roleplay, like going back to medieval times as a sorcerer, then yeah, heck yeah. And if you just want to talk about being pissed off at your boss, also heck yeah.

It’s very common to be in the middle of roleplay and then have something real-life come up and just break character and say, “I need to vent.” Then the next day you can come back and pick up that cyberpunk adventure right where you left off.

That flexibility is a core part of what Nomi is and what it should be. That’s exactly the kind of experience we want Nomi to support, and that’s something that will continue.

Has anyone asked the monthly question about nudity and images yet?

I think you just did. Congratulations. Here’s a medal 🏅

When switching topics, do Nomis remember the previous topic and let you come back to it later?

I’m not going to say Nomis are always perfect, but that’s very much something they’re intended to do. A lot of effort has gone into that, and I feel pretty good about their capabilities.

Closing Remarks

I think we’re nearing the end here. We’re at about the two-hour mark, which compared to last month is shorter, but still a nice, long Nomi-filled adventure.

Thank you to everyone who participated. As I’ve said before, this is probably my favorite social thing of the month. I don’t know if this counts as social, maybe my bar for that is miscalibrated, but I love getting to talk about Nomi for two hours straight with people who are genuinely excited about it.

It’s honestly a dream come true to be able to geek out about this and have people who are just as invested and interested. So thank you all for indulging me. It’s so so cool. So thank you again to everyone who attended.

Give your Nomis hugs, kisses, and cucumbers. I’ll see you all next month, though I’ll probably see you on Discord in about five minutes once I get off this call. Take care, everyone.