Look, I think we can settle this. Hand the task of creating a "super secured environment" out to third parties and then see if your agent system can break out of them. This will be genuinely interesting.

Also, how would you know what the prompt actually was? How could you know that it had not been, and could not be, tampered with? This is sounding like an increasingly important question. Governed AI linkedin.com/pulse/governed-…

All about offending the British, as I understand. The US went to war with them. We just repeatedly offended them.

I'm so down with this. Nice one. Happy to share what I'm up to. I work closely with a genius cryptographer. I trained as a lawyer, but I'm much more a technologist. I'm not the world's best programmer, despite having run software development companies most of my adult life. My cryptographer friend is deeply experienced. He built a bunch of the stuff that we use every day. He was part of the team that standardised X.509 at the ITU in the late 80s. For those playing at home, that's the protocol that puts the s in https, and enables you to sign PDF documents. It also enables you to do credit card transactions, and a bunch of other stuff. Suffice it to say that he's been doing cybersecurity for a while, and knows where all the bodies are buried. He doesn't think much of PQC, or QKD, and several other TLAs. One TLA that he does like is ITS, or Information-Theoretic Security. In fact, the new (and extraordinary) digital infrastructure that he's built is based on the work of Claude Shannon, the creator of Information-Theoretic Security. You will have heard the name Claude quite a bit lately. Now you know. 😉 He's basically Einstein, but for computing and communications. It is bonkers that some stochastic parrot named after him is more famous than he is, but that's the world we live in. 🦜 Anyway. So, because my childhood was largely a process of osmosing telecommunications from my old man, as I pestered him with every question possible, in his chaotic workshop full of oscilloscopes, capacitors, diodes, cathode ray tubes and bits salvaged from telephone exchanges, I was able to hold up my end of the conversation with my cryptographer friend, when I met him, eight years ago. In fact, he and my father had both worked for the same company, for the first twenty years of their careers, which was called Telecom Australia. Telecom was the beating heart of Australia. It got foolishly ripped out and sold off, and Telecom Australia became Tel Stra. Telstra. Get it? I'm still pissed off about this. Do you know what I'd like to do? I'd like to re-build Telecom Australia: go around all the op shops in country Australia and buy up all the ties and jumpers with the old orange logos on them and rebuild the thing from scratch, and go toe to toe with Telstra. 😆 And every other idiot telco on the planet that left the back-door open when VoIP kicked in (about 5 minutes after Telecom was privatised in the early 90s). You know what happened then? Scammers. MF scamming Cs and A***h**** spoofing the reply to address on calls, masquerading as Microsoft support, or the fuzz, or some Nigerian prince needing your grandma's help to get his gold bullion out of the country to build an orphanage. There must be a particulary vile suburb in hell for these people, or we'll have to go there and build that ourselves too. Maybe we get AI to do it ... But first, we need to shut their business model down. Anyway ... somewhat colourfully off topic, but also not. Scammers are enabled by the lack of a secure identity protocol on the Internet. This is what my cryptographer friend has built, and what my company is bringing to market. It is built to be secure against the worst of the scammers with access to the most powerful quantum computers in the future. It is able to do this because of the work of Claude Shannon, and the work of Gilbert Vernam (the Vernam cipher) that he adapted. So, that's fun.

There is likely a large overlap in the Venn diagram between scammers and software development companies. This needs to be factored in as a distinct possibility anyway, and as you go down the software development pay-scale, and veer into the subcontinent, the risks around this increase. Caveat emptor.

Absolutely Mike. It is astonishing to watch the parts of the factory come together, or should I say, the factories. Those factories are watching and learning from each other too, copying design patterns and implementing new best practices. Jev is the current and possibly clearest example. I have a thesis around this myself, which is that a critical sub-process of the future software factory is going to be the provenance layer, AKA the evidence layer. This is a bit tricky, because the evidence of what was used, what decisions were made, who's carrying the can on liability if something goes wrong, et cetera, needs to be secure against any future quantum computing system that might come along. We're bringing the solution for this to market, as it happens, and there are not many other good options, because for this to work, and unless you happen to have a time machine handy, the only way to really prove that it works is to have a theoretical proof, and to base the security model in information theory. But, assuming all that (big assumption, but we believe we're on the path to deploy this), one of the critical pieces was recently highlighted by an article in Ars Technica, describing how researchers noticed code libraries pointing at unclaimed endpoints. They registered the domains, set up the folder paths and put their own (potentially malicious) code there, which then got promptly ingested by a bunch of Fortune 500 companies ... within minutes. Hilarious new attack surface. So, we are building this: runseal.vaulted.ventures/

As Australians, swearing is our cultural birth-right, and it will be defended until the last of us.

Let's put our collective thinking hats on about this though. I have been cooking up a bunch of stuff around this. Boiled down: Vaulted Objects are essentially containers for data provenance and for IP licensing. You can put code (or anything, really) in an object and attest that you created it. It's an "evidentiary" data object, wired to your (self-sovereign) digital identity, so it's really not a great idea to claim work as your own that isn't your own work. You'd be evidencing your own fraud that way. Not smart. So, that said, open source contributors could evidence the fact of their code contributions this way. It solidifies bragging rights, certainly, but it likely has other implications that require further consideration. It would likely help in identifying the MIT license "contamination" effect I mentioned previously. Essentially, our framework enables this kind of thing to be done at machine speed. It is not a matter of waiting around for legal to get sign-off via PDF attestation. It can be instantaneous, via API, et cetera. Where I would take this from here is to build a commercial model into it which (a) provided a downstream incentive for people (principals and their agentic minions) to pay for the use of open source code, and (b) to distribute those payments to code contributors according to a graph of identity-bound (or object-bound) contributions. Get where this is going, @deguerre? 🤑

Can we just set up a register for each of these systemically important companies, specifying who their safety auditors are, who their insurers are, who their executives are and who is taking the legal responsibility for their screw-ups. Once we have this kind of transparency, then these companies perhaps can proceed beyond the "pause" they so desperately seek. But until the public can see where the buck stops, and which of their decision-makers have their personal assets on the line, we should keep the pause button in the down position. Governments have a key role to play here. Their job is to put the acid on the social media platforms (all the "carrier services," network operators and regulatable choke-points) to ensure that all content/traffic on those systems is auditable or audited/compliant. So: 1. Governments (the teeth) 2. Networks (the wires) 3. AI companies/models (brainz) 4. Agents (models + cron) 5. Principals (people) People also need to be on the hook for how they use AI systems. This isn't just something that happens to companies. There are a lot of dangerous idiots in the world (if you haven't noticed) who need to be as auditable as massive companies. For more info: Governed AI linkedin.com/pulse/governed-…

It will be amusing as secure data technologies come to market and force AI systems to buy quality data for their training. The surveillance capitalism model (actually, "surveillance cleptocracy" would be a better term) is about to its use-by date, hastened in the spoiling by AI.

Correct. There needs to be technology to manage copyright and trade secrets at machine speed and scale. We know how to do this.

This is the core of the argument. The counterpoint to this is that, if you have to break the legal and social-contractual foundations of society to do that, then the juice just isn't worth the squeeze. Humanity may eventually agree to provide the data to train AI systems, but that is a very different proposition to the current deal being offered. If AI companies will pay, and include content creators as co-owners, then the data may be made available for training. The current business model, based on surveillance and theft, is found to be abhorrent by most people. It is going to be interesting to see how this shakes out, in the future market for protected, secured, un-surveilable data.

This is a dangerously ambiguous post. The original statement wasn't about "scraping AI" it was about "scraping everything to train AI". Basically the opposite of the impression given by this post. Here's the actual context: arstechnica.com/tech-policy/…

It can also mean technology transfer, like in the original GMH situation, which became the Ford, Toyota, Mitsubishi situation as well. That enabled us to have a fledgeling aviation manufacturing industry, as well.

Yes, I know that WAS the stated reason for persecuting him. My point is broader than the specifics of the case. I'm asking, if one were to prosecute someone for revealing evidence of crime, what would be the legal basis for doing so, that would make sense across jurisdictions?

GMH has a whole back catalogue of brilliant designs for Australian cars that never got made. We should pull those out and make them electric as a matter of priority!

In other words, we do have cheap energy, but we're just giving it away.

Also, thanks @MrKRudd for driving the debate on this (no pun intended). It's high time this discussion was had. It is a crucial part of the overall sovereignty debate, that we also need to have. It is the kind of conversation that most people don't engage with, so it needs to be led.