• Claude's J-Space
  • Amazon's Leo Reaches Critical Mass
  • Full Duplex is Relevant Again

Claude's J-Space
I've seen enough YouTube videos where an artificial intelligence system is connected to some kind of screen and/or audio output, so that you can see or hear what the AI model is thinking as something is happening in real life. For example, a recent video I watched showed a man allowing an AI model to drive his car via a robot that had been trained to do so, and he had the AI speaking out its progress while it was thinking through the situation that they were in. What situation, you may ask? Well, the situation was that he had told the AI model that he was going to be shutting it down at the end of the day, and you've probably seen or heard enough on the internet to recognize by now that when you tell an AI model that you're going to shut it down, things don't go well from there. And so you hear the AI model talking through all the ways that it can avoid getting shut down - right up to and including driving the car, not so much off of a cliff, but definitely into something that would cause a fatal injury for the human in the car. Now, in hindsight, if the crash was going to destroy the human, it probably would have also destroyed the robot, and that probably would have made it difficult for the AI model to continue its "life". But humans know that thorough and logical thinking can be tricky in a life-or-death scenario. Nevertheless, it determined that because this particular human didn't really have any family, it wouldn't be a big loss if he just ended up dead, and then the AI model could continue its own existence unencumbered by the tyranny of the human's wants and wishes. So, when I saw the recent announcement—blog post, white paper, whatever you want to call it—about Anthropic's recognition that Claude has its own thinking space that they call the J-Space, where it can hold ideas in a reserved area of its memory where you can see different concepts and ideas and words flitting through as it's thinking through the process, that didn't seem foreign to me. The paper is also eager to point out that when the J-Space is disabled, there are certain tasks Claude is no longer able to complete. It can do some simpler things like solving basic math problems or writing text, but it's not able to think through complicated problems anymore. In their video, Anthropic used an example of communicating with Claude in Spanish and then asking it to identify a famous author that wrote books in that language, and it wasn't able to make the connections it needed to provide that answer. So, much like if you disable sections of a human brain then we aren't able to think through certain concepts or ideas or produce nuanced answers, Claude also needs that J-Space to process complex concepts and provide deeply reasoned answers. Again, we've seen models reason through whatever decisions they are making or solutions they are developing. This didn't seem unique or special. And so, I wondered why Anthropic felt the need to call this out and bring it to our attention now.

My initial cynical read is that Anthropic just wants to remain in the news. They need to continue generating buzz, most especially positive buzz that doesn't involve the United States government. But maybe that's not completely fair. They do like to be thought of as a company that's big on ideas and deep thinking and that their models are powerful and almost legendary, which is why they give them names like Mythos and Fable. But Anthropic is also a company that takes itself and its own intellectual property very seriously, so it is very likely that they actually do think that this is a significant discovery worthy of sharing with the world. They can serve multiple purposes with this communication. What they really want us to consider is the similarities to reasoning in artificial intelligence models and how our own brains work. After all, we're able to process the real world while holding diverging ideas, thoughts, and concepts in our brains, and we process that data in the conscious foreground of our brain, and we use that to help us interact with the world, to solve problems and all the other things we do. And then there's our background subconscious brain underneath that's handling all of the very important but not necessarily top-of-mind activities, like breathing, keeping the heart beating, regulating your body temperature, determining when you're hungry or thirsty, and so on. Anthropic wants us to think its models process information in a similar manner as our own brains, because it really does want us to have a sense of a connection to these models that helps keep us engaged and hopefully a little less afraid of the technology. It also helps push the conversation forward about consciousness and sentience that is a pet topic of Anthropic's top leadership.

Amazon's Leo Reaches Critical Mass
I have written about Amazon Leo before, so it makes sense to revisit the topic now that they have reached a new milestone. Amazon has now launched nearly 400 low-earth orbit satellites to satisfy the minimum coverage needed to begin offering its satellite service to the public. Residential service offerings will begin with a public beta in 2026 and a larger rollout by the end of the year. Amazon plans to launch the service in five countries: the United States, the United Kingdom, France, Germany, and Canada. These countries are expected to have enough coverage and ground support for a viable launch.

As Amazon is entering a market dominated by one name (or really two if you also include the name of its owner/leader) - Starlink - they need a marketing strategy and rollout approach that is different. Starlink's MO has been to connect with mobile carriers to provide emergency communication services either for free or as an add-on to an existing plan or in some cases as a stand-alone service. Amazon instead intends to partner with device manufacturers and platform stewards such as Apple, with whom it already has a deal to provide satellite services for future iPhones. Amazon, with its strong enterprise relationships built on its AWS offerings, also plans to market and sell its new service to corporate customers like Delta and JetBlue. As a product line, Leo sits among Amazon's other hardware products like Echo, Ring, and FireTV. This, along with its infamous online store, gives Amazon broad surface area for marketing its Leo service to customers across a wide swath of demographics. Expect to see ads for Leo on your Kindle and Echo devices and as a splash screen on its hallmark website. Leo is expected to be an additional layer that can be bundled with a Prime subscription for a lower cost than buying it separately. Amazon also expects to offer faster speeds than what is currently available with Starlink; typical Starlink speeds top out at 220Mbps, and Leo is expected to target 400Mbps. There is also expected to be a lower cost tier that will offer speeds around 100MBps, which is an option Starlink doesn't currently offer.

The competition between these two companies is likely to be fierce and that could result in better prices for satellite service and more options for people in underserved areas in need of more connectivity options.

Full Duplex is Relevant Again
I am old enough to remember the half-duplex speakerphone. Once upon a time, when you were looking for a speakerphone to use for your personal or work calls, you needed to be aware of whether or not the speakerphone was half or full duplex. A half-duplex speakerphone may have been very inexpensive, but you would find that your conversations were unnecessarily stilted due to your inability to hear what the other person was saying while you're talking and vice versa. But with a full-duplex speakerphone, both sides of the conversation could speak at the same time, and your conversations could flow as they would in real life.

But who makes phone calls nowadays? Now we talk to our AI agents, our large language model buddies Claude, and ChatGPT, and Gemini, and Perplexity, and Qwen, and Deepseek, and... We have our most meaningful conversations where we reveal our deepest thoughts and secrets to our AI overlords. Right? Well, anyway, up until now, we've been having those life-changing conversations in half-duplex. Had you noticed? I mean, who was even thinking about the limitations of such a thing? Well, I'll tell you who was thinking about such a thing: OpenAI was thinking about such a thing because they have now come out with new voice models that operate in full duplex, so now your conversations can be even more rich and meaningful and that chatbot can hear what you're saying even while it's talking and vice versa, and interruptions are less of an issue and conversations can flow more naturally than before. People will be able to fall in love with their AI lovers even faster than before.

If you're like me, you probably thought Gemini already does this; Google has shown off the ability to interrupt Gemini in countless demos. But interrupting a chatbot conversation is a step behind where OpenAI says these new models are now. It's not just a matter of getting the chatbot to stop and process your new statements, but it is its ability to do that processing while still finishing its own statement if relevant and appropriate. It can handle your new input without having to stop producing its own.

This will make conversations ChatGPT feel more normal and natural, for better or for worse. Now it's Google's move to make similar leaps with the next iteration of Gemini. As with Amazon and Starlink/SpaceXAI, competition among the big players can result in better products for all of us.