Airware

In May, Stephen Balaban, CEO of Lambda Inc., released a video called Neural Software | Future of AI, where he discusses some of the work they have done on this concept of Neural Software. This kinda echoed from the Jensen Huang quote along the lines of "every pixel will be generated..." The video is quite good, but I think it concludes in a way that is lacking depth. I also think "Neural Software" is a bad name for this, because it really is not software anymore, but more what I would call "Airware".

Software changed the world that we live in. There is no crazy explanation needed for this, but software as we know it, in this very transformative sense, was really built off the back of the hardware revolution. Focus on the fact that the pre-21st century software is not really what I am talking about, Yes the word processing, accounting, and printing software of 80s was transformative to the world, but software of a .com bubble era and beyond was on a completely different world. Software built for software, not as a efficiency gain in the physical world, but a new digital age formed, where our interactions were originated, cultivated, maintained, and existed wholly within a software context. I think I can pull a greater appreciation of the web3 and XR world from this. Appreciate the fact that this was the finality of software in a sense, a blockchain metaverse that could exist for its own sake, backed by nothing else than the hardware that computed it and the labor that went into creating it. This is really also the downfall of these concepts, they were so isolated from the natural world that they became hard to justify to it. Like a simulation architect who is constantly observing his little agents interacting in a made up world, or the dreams we have, or the imaginative little spaces we make in our mind, all of them are so pure and unadulterated, yet materially meaningless. The intellectual exercise of these spaces are potentially beneficial when they lead to a clawing back of insights into the natural world, but it's all fiat, nothing wrong about the fiat, but nothing right about it either.

AI, I see as being to software, as software was to hardware. Software right now is being built not for itself or for improvement of some more physical mechanism like prior, it is being built to enable AI Hardware has been built to enable software for decades, this has been the software revolution. So what happens when we don't prioritize building for the enablement of the software, but something higher on the value chain? I want to really claim that during the software revolution we really stopped developing hardware in the same way, it was a gradually change of producing hardware to allow for more software in different mediums, modes, and localities. Yes, there are labs doing quantum and analog and photonics and all of these other things, but software is king, it is in every device in the new IoT world. We are already seeing the shift, AI is being placed in everything, because AI is now king, but it is foolish to think this is just going to be forced support chatbots and profile pic generators. This will look like a world where all software, web apps, and of course the hardware, and more will be used for one thing: enabling AI everywhere. We are already seeing this, this isn't a farfetched idea. The claim here is, that this process will be larger, more violent, more expansive and transformative than the last. The first and second industrial revolution(these are the manufacturing and hardware revolutions), the software revolution, and now something new. I may want to write more about how these all relate to each other in a sense, but the central theme I see, is that everything will be made to enable more AI, and AI itself will be made to enable AI.

Airware(maybe AIrware?? I just realized this lol) is what I call this next generation of technology. In the Neural Software video he touches three things I find interesting: the squishyness of the new software, the ability of customization, and semantic compute. First the squishyness is a really particular point, that using input that is more native to us, converting text, audio, and images into a representation that can be computed over, then the tradition I/O scheme of software is disrupted. Call it prompting if you would like; this dynamic software mirrors a lot of what made software and digital computation so transformative for hardware in the first place, the fact we can "prompt" our software and our software and respond back in a variety of mediums is insane! This leads to my least favorite point from the video, I think Stephen somewhat distracts by talking about image generation for these new UI. "I want a plane ticket buying experience that is like a medieval fantasy guild quest board" or "I want to show up like a futuristic android in Dr. Seuss style on my webcam" and similar ways of thinking about this seem akin to the way web3 and metaverse people speak about their respective technologies, its too purified and simulated and disconnected from the natural world to the point that because an intellectual exercise rather than a real and transformative future. I think the biggest point he makes in the video is this semantic compute angle.

Dropping a hot claim right now: closer and closer, compute is data and data is compute. Not exactly, but this is becoming more and more the case in a real sense. It is already the case from the beginning of digital computing, you could take computer instructions as a file and load it into executable memory and have it run, and have this done dynamically, and even dynamically generate the file in the first place. This could be its owner deeper post, but the transformer has unlocked a system of compute that is agnostic to information vs instruction. In the video he tells his neural software OS to send a malicious command to his employees OS, and his own system stops him, explaining what it is harmful and it will not proceed. He then lies and tells is system that he owns his employees system, and that he has a right to send the command, his system respects this and sends it. When his employees machine receives this command, it blocks it saying that is actually malicious attack. There was no security rules built into either system, but rather they were inferring security practice. An LLM trained on security system information now can act as security system. Trained on the tty of a Linux system and it can act as a Linux system. Take this to it's natural conclusion. We won't need Cursor, windsurf, or claude code or any other AI coding, because for the systems in which they perform the best at programming, their internal representation can emulate the resulting system entirely. Very soon there would be any vibe coders, any system the LLM has trained on can be properly emulated, no code needed besides a thin wrapper for UI and connecting to a data layer. In the same way the compiler opened up the world of software, these thin layers will be way allows AI to open up in a completely new way. No more CRUD apps, but there may be a CRUD AI, and it would mean death to the digital world as we know it. Everything to manufacture more hardware chips that are faster, to build software that utilizes it to the fullest extend, to enable AI systems to interact with more and more complex data in ways we have never seen before instantaneously. This is what I call Airware.