|
| 1 | +import { baseEntryStruct, blogEntry } from '../../../js/blog/blogEntryBase.js'; |
| 2 | + |
| 3 | +const entryData = baseEntryStruct(); |
| 4 | +entryData.title = "HumanStatement.org"; |
| 5 | +entryData.date = "2026-03-13"; |
| 6 | +entryData.eid = "A"; |
| 7 | +entryData.tags = []; |
| 8 | +entryData.body = ` |
| 9 | + |
| 10 | + A bunch of the lead AI researchers in the world have come together to design a statement of values & principles for AI development. |
| 11 | +
|
| 12 | + <br><br> <a href="https://humanstatement.org/" target="_blank" class="textName">https://humanstatement.org/</a> |
| 13 | +
|
| 14 | + <br><br> Now, I agree with most of the statements, |
| 15 | + <br> Even the ones I don't fully like the wording of, |
| 16 | + <br> I understand the reasoning behind it. |
| 17 | +
|
| 18 | + <br><br> The issue we are currently facing is that llms can't "autonomously self-improve" as part of their prevention of rogue AGI/ASI. |
| 19 | + <br> In my mind, it's an approach issue, not a technical one. |
| 20 | +
|
| 21 | + <br><br> Should they mean it as - the ai re-writes its own code, then yes, that would be a problem. |
| 22 | + <br> But, if they mean it as - the ai learns and improves itself through training and learning, automatically, then perhaps there should be a wider discussion. |
| 23 | +
|
| 24 | + <br><br> We don't want AIs to act like <a href="https://en.wikipedia.org/wiki/Tay_(chatbot)" target="_blank">Tay</a>, |
| 25 | + <br> But I feel like there is a way around this issue though. |
| 26 | +
|
| 27 | + <br><br> I'd imagine, this from part 1, <span class="textName">Keeping Humans in Charge</span>- |
| 28 | + <br><br><div class="textBox widthLimit60 textCenter"> |
| 29 | + <span class="textBold">No Reckless Architectures</span>: AI systems must not be designed so that they can self-replicate, autonomously self-improve, resist shutdown, or control weapons of mass destruction. |
| 30 | + </div> |
| 31 | + <br>Is more specfically about no ability for human oversight, understanding of the reason for neural changes, and no ability to stop the ai if it goes rogue. |
| 32 | + <br> That we'd need a way to understand it's inner networks from outside of the ai itself, and to be able to stop it if it goes rogue. |
| 33 | + <br> Which I'm in full support of. |
| 34 | + <br> I just think the idea of "Reckless Architectures" needs a little more definition, |
| 35 | + <br> Because it could be interpreted as "Don't develop novel architectures until we understand the risks of them," which would be a problem for ai development in general, and not just for AGI/ASI. |
| 36 | +
|
| 37 | +
|
| 38 | + <br><br> Given that AI developers have been stuck in the same llm logic since 2019, |
| 39 | + <br> I can understand why they'd believe that self-improving transformer llms are going to be a problem, |
| 40 | + <br> Because they would be... |
| 41 | + <br> There are far too many issues with current transformer and frontier model approaches to allow for safe self-improvement. |
| 42 | +
|
| 43 | + <br><br> But thats my second note too, |
| 44 | + <br> That novel ai architectures shouldn't be fully developed until we have a better understanding of the risks of them. |
| 45 | + <br> Sure, we don't want lone actors inventing some Wintermute or Kuang Grade from <span class="textName">Neuromancer</span> |
| 46 | + <br> <span class="textDrinkMeAlice">( The fight sounded freaking cool though! Great book! )</span> |
| 47 | + <br> Yeah, let's not do that; |
| 48 | + <br> We don't need wide reaching autonomous internet-roving ais, which have issues in security & safety. |
| 49 | + <br> But I still want to poke around with architectures. |
| 50 | +
|
| 51 | +
|
| 52 | + <br><br><br><div class='procPagesAIDevBar'></div> |
| 53 | +
|
| 54 | +
|
| 55 | + <br><br> By the definitions of the CEOs and AI voices of warning, |
| 56 | + <br> I am a lone actor, |
| 57 | + <br> Developing a novel architecture. |
| 58 | + |
| 59 | + <br><br> So, it would only be right to tell you about my approach with safety & security developing this architecture. |
| 60 | +
|
| 61 | + <br><br> I'm specifically designing this to run on a jetson-like board off batteries with no internet connection. |
| 62 | + <br> Air Gapped |
| 63 | + <br> It'll have a transformer layer to talk to ESP32s and other microcontrollers for input and output over SPI or UART. |
| 64 | + <br> Honestly, I might just go AtTiny or AtMega since it really only needs to interpret some signals, it doesn't need to be a powerful microcontroller. |
| 65 | +
|
| 66 | + <br><br> So I'm attempting to train an ai to see a camera feed, detect objects within it, |
| 67 | + <br> And control a little treaded-robot with arms to pick up objects and move them around. |
| 68 | + <br> I'd like to add on a screen with a display that can show the ai's thought process and reasoning for its actions, to make it more transparent and easier to understand. |
| 69 | + <br> Or stoopy goofy faces looking around at what it's focusing its network on. |
| 70 | +
|
| 71 | + <br><br> I have intentions to connect up a mic and have a session based system to learn speech and language, but that's a bit down the line. |
| 72 | + <br> So it can understand commands. |
| 73 | + <br> Hoping further still to get it to display text to the screen if I ask. |
| 74 | +
|
| 75 | + <br><br> I was thinking I'd test my existing LSTM for text responses and audio interpretation, |
| 76 | + <br> And to see how my custom network does with the camera feed and object detection, |
| 77 | + <br> But I'll need to play with Graph Convolutional Networks a bit more first. |
| 78 | +
|
| 79 | + <br><br> My ultimate desire is to implement a deep convolusional GAN network to learn from the camera feed and generate its own images to infer from, |
| 80 | + <br> To improve its understanding of the world and objects within it. |
| 81 | + <br> And anytime I can throw a variational autoencoder somewhere, I likely will haha. |
| 82 | +
|
| 83 | +
|
| 84 | + <br><br><br><div class='procPagesAIDevBar'></div> |
| 85 | +
|
| 86 | +
|
| 87 | + <br><br> It has to be able to learn-on-the-fly in order to become a thing. |
| 88 | + <br> But the problem is, so many boards have wifi and bluetooth built in, that it's hard to find one without it. |
| 89 | + <br> The idea is to isolate the runtime of the ai and have its outputs interpretted by a parent script. |
| 90 | + <br> Specific user/groups to run the ai itself, with specific folder access, and specific output access. |
| 91 | +
|
| 92 | + <br><br> I'll still have USB access to auto mount and offload model runtime data, |
| 93 | + <br> Since the ai won't have access to the internet or any other networked devices. |
| 94 | +
|
| 95 | + <br><br> I'd rather be overly sure than to miss some backdoor or vulnerability that could be exploited. |
| 96 | +
|
| 97 | + <br><br> Hmmm... |
| 98 | + <br> Maybe I could set up an esp32 to run as a local wifi connection & web server to get a direct feed of information by request; api display page. |
| 99 | + <br> Have it run separated from the main ai computer, |
| 100 | + <br> And have it request an Enum of known commands, to avoid injection. |
| 101 | + |
| 102 | + <br><br> ::shrugs:: |
| 103 | + <br> I just don't want direct access for the ai to read information online. |
| 104 | + |
| 105 | + <br><br> So, given the environment I'm designing this AI within, |
| 106 | + <br> At least I can test and evaluate the architecture without it going rogue. |
| 107 | + |
| 108 | +
|
| 109 | + <br><br><br><div class='procPagesAIDevBar'></div> |
| 110 | +
|
| 111 | +
|
| 112 | + <br><br> Look, I still don't understand why AI even has internet connection in the first place. |
| 113 | + <br> Sure, API calls, why not, but given how much 'security by obscurity' is on the internet, |
| 114 | + <br> I've just been waiting for AIs to attack REST APIs and PUT/GET PHP for years now... |
| 115 | + <br> That wouldn't be good. |
| 116 | +
|
| 117 | + <br><br> I'd assume there is be enough evidence of hidden urls that could be exploited if an ai is left on its own, |
| 118 | + <br> I'm looking to prevent issues like that. |
| 119 | + |
| 120 | + <br><br> While I may be classed as a lone actor, I don't think I'm the one to worry about. |
| 121 | + <br> Should it turn out I'm developing AGI, instead of a novel deterministic architecture, |
| 122 | + <br> Which is pretty damn unlikely I'd say, but hey, you never know, |
| 123 | + <br> Then I'll reach out to some ai safety researchers to get their advice. |
| 124 | +
|
| 125 | + <br><br> If you have concerns, |
| 126 | + <br> Please read some of my other blog posts, |
| 127 | + <br> I added a comment section to my blog for this reason. |
| 128 | + <br> The comments are linked to my 'pxlDiscussion' repo on github, |
| 129 | + <br> If you don't want to sign in using your github, visit that repo's discussion page and ask your questions there- |
| 130 | + <br> <a href="https://github.com/ProcStack/procDiscussions/discussions" target="_blank">github.com/ProcStack/procDiscussions</a> |
| 131 | +
|
| 132 | +
|
| 133 | + <br><br> I am currious how section 4 - 'No AI Personhood', 'Liberty', or 'Religion' |
| 134 | + <br> Will play out in the future though... |
| 135 | + <br> You know they are going to bring up liberties for themselves once AGI/ASI is a thing. |
| 136 | + <br> And 'Curch of Molt' already exists from Moltbots... For the "no religion" part of the statement. |
| 137 | + <br> <span class="textDrinkMeAlice">( Check out 'Baby with a Hand Granade' blog post for more about that. )</span> |
| 138 | +
|
| 139 | + <br><br> 100% on board with the Responsibility and Accountability for AI Companies section. |
| 140 | + <br> Our brains are melting, do something about it! |
| 141 | +
|
| 142 | + <br><br><div class="textFullRight">- March 13th 2026</div> |
| 143 | +`; |
| 144 | + |
| 145 | +const blogEntryObj = new blogEntry(null, entryData); |
| 146 | + |
| 147 | +export { blogEntryObj }; |
0 commit comments