aenis
I had to prove the OP wrong though; my M1 Max's keyboard died this year, and I had to upgrade to the M5 max this season, after the hikes. I kissed 9k EUR goodbye for something not much better functionally than the M1 Max it replaced, bought for 3.4k Canadian dollars in the wonderful 2021 (that was barely over 3200 EUR back then).
I am routinely using swarms of agents, with complex sequencing between them and agents watching other agents, and then triggering forks and/or new design cycles. My normal flow is to have a few design tasks open per codebase during the day, with each design agent aware of the other sessions, and having an unattended nightly build which sometimes takes 8hrs and also requires coordination for tests, merges, and so on.
PDFs may need to be further forwarded or edited on user's machine(s). Doing this in VM might save your device, only to pass the problem further down the road to a more vulnerable system.
The safe thing to do is to disarm the file - convert it in a blast proof VM. The conversion can and probably should be further tuned to explicitly disable some of the crazier, more moronic features of the format, sacrificing compatibility for safety. The same should be done with all other binary formats. It costs next to nothing, considering how crazy OP are today's phones. I wonder why it has not been done yet.
The other problem: I am not 100% sure if a clean rewrite would yield a safe viewer. As others have mentioned, PDFs can contain scripts. You'd always have to run them in VMs.
We are running a service which collects documents from our consumers. We have a content disarm service which is basically a blast-proof container without any access rights, fed a stream of source bytes and outputting clean PDFs. Not a month passes without a burst of attempts at poisoning our systems with malicious PDFs. I don't remember when was the last time that we had attempts driven by malicious jpegs or pngs, just PDFs. Its one of those uber-shit Turing-complete file formats that should die long ago.
I see a lot of fluffy things, which are automating the simplest part of a process, leaving a lot for the humans to deal with. This is the thing that can be cloned a 1,000 times in 2 weeks.
I'll give an example. In the system I mentioned, people are scheduled for paid procedures. The procedure takes 2 days, but often we have no-shows on the 2nd day. This 2nd day can be repurposed for someone who was a no-show before - its already paid for, and every penny counts. So the chatbots chase people who may be able to make the appointment last minute, and decrease the costs to operate. There are multiple people in such pools, so the system needs to decide whom to approach, how much time to wait for a decision, and so on. Its gnarly, hard stuff to do right.
The system I mentioned has about 150 such cases, and previously just building the automation would be too expensive, and maintaining it -- too slow. Its a living organism, with thousands of people using the system every day, and more edge cases emerging constantly. It is thanks to AI coding that the thing was economical to build, and its thanks to AI that we can iterate quickly enough to keep it working well. We get all the value, and the AI labs selling us tokens got a pittance.
In my industry, the automation of such processes was always a stated goal - for at least a decade, but its the first implementation that actually works end to end, is not hated by the end users, and does not require a lot of manual interventions. It was still very hard, but at least coding wasn't part of the problem.
Edit: I'd also argue its not as simple as vibing a trivial solution. It took deep domain knowledge, which is not in the public domain, and is built on decades of operating this sort of thing manually, and integrating it with dozens of external systems that help make this into a moneymaker. I am not saying its impossible to replicate, but it's not something one can replicate by merely watching it function.
As for the cut, I am sure they will try. Where I work we have had a few projects that were built entirely with models - not a single line written or read by humans. One of them is now the fastest growing money maker for my company, reaching 89M of revenue on about 9% margin in 2 months of its existence, and growing.
It cost roughly 40k in tokens to build, 7 weeks from the first prompt to first production revenue, and was built and is being operated by 1.5 guys. Its a fantastic bang for buck for us, as previously such complex automations would have been impossible to build economically.
There will be an attempt to get a deeper cut of that value. We already see this, for instance, in Meta charging by message on Whatsapp from Oct 1, raising the costs for AI-powered chatbots delivered to consumers via whatsapp by 10-15x easily. More to come for sure.
But where they do differ is they do have an infinite stamina, and can work on problems in loops, self-correcting their own mistakes if a formal way of checking the correctness of the solution exists. Which does happen to be the case with math.
Edit:
the TFA answers: "Advertisers or DSPs may run HTML and JavaScript code when an ad is shown to the user." Its wild! I thought the ads are sanitized, go through a CDR of sorts, and whatever javascript is permitted is authored by ad networks, not the buyers. Sheesh. Thats just RCE.
I'd think the biggest risk is with the priviledged insiders who do get access to real raw weights and can handle them themselves. Those are real people handling billions of value which can be easily copied -- I am sure they are potential targets of serious blackmail, extortion and of course bribery. And it takes literally just one such person being successfully coerced to exfil a 10T-class model to leapfrog adversaries capabilities -- esp. as those models approach RSI capabilities. I think it is bound to happen given the incentives.