Comment Re:Other metrics (Score 1) 198
It should be "number of residents who are legally and physically able to work" (ie. no minors ineligible to work, no disabled persons, no elderly unable to work, etc.) in the calculation.
It should be "number of residents who are legally and physically able to work" (ie. no minors ineligible to work, no disabled persons, no elderly unable to work, etc.) in the calculation.
The main difference is that those other mechanisms try to prevent the LLM from doing things. "Guilt" simulation takes a different approach: penalize the LLM (cost it points) based on the damage done to the other party. So in the Prisoner's Dilemma game, defecting causes the other player to suffer punishment (lost points) which in turn causes the LLM to lose points too. The LLM's trying to minimize point loss, so the idea is it'll look for strategies that avoid causing the other player to lose points as well as trying to minimize direct point loss and it'll do this through it's own programming rather than having external constraints that it could work around.
The obvious downside is this requires feedback to the LLM about the results it produced, which in turn requires it to keep a permanent memory of every user and it's history with them. Any breakdown or inconsistency in that feedback loop causes the outcomes to get worse rather than better.
This is something easy to test by just setting up multiple AI agents and running the game. The optimal strategy is of course to always stand pat, say nothing and accept the minimal punishment when the other agent also stands pat. That, though, is only the case if all agents were programmed for "guilt" and if all interactions resulted in feedback based on the other agent's punishment. If any agent doesn't receive negative feedback for choosing to defect (isn't programmed for "guilt") or if feedback is sometimes not given, the other stable strategy is to always defect. My guess is that if they ran the simulation, that's where the agents would end up. I doubt the agents will figure out the "best" strategy, which is to initially stand pat and then return tit-for-tat until the other agent also starts doing that in which case switch back to standing pat. This requires being able to distinguish other agents from each other and memory of your history with a given other agent.
A few floors below? Hells, the scammer/ransomware squad probably outsourced the calling to Cognizant. Saves time and phone lines.
It's very critical additional data, though. The idea is the basic malware doesn't have anything truly alarming in it, so it's easier to obfuscate it to slip by the filters. Once it's running, it can download the really dangerous stuff via TXT records, which won't be checked for alarming/dangerous content, and set that up behind the filters without triggering them. It does all that in memory, without giving any indication of touching files, and then once the code's gotten root access it can persist things to files without triggering any alerts (because root's allowed to do that).
This is compounded by the (ab)use of TXT records to store arbitrary records without needing to extend the set of DNS record types. Having explicit types of DNS records improves syntax filtering, making it difficult to impossible to abuse those records this way. Limiting TXT records to only their original purpose (instead of putting SPF, DMARC, DKIM and other types inside them) would allow heavy-handed filtering of attempts to pass arbitrary data in large quantities through TXT records without impacting any other services.
Why do AIs keep putting fake citations into cases they generate? An AI just regurgitates what it finds in its data set after all...
No, they actually don't. What they do is regurgitate the sequence of tokens most likely to follow the current token based on a model generated from the training data. They certainly can regurgitate the exact contents of the training data set, but they can equally make up strings based on the model weights. That's what makes their output so pernicious: since it's what would likely follow, we see it as plausible and don't immediately flag it as something we need to check on. The only way to find the problems then is to check everything regardless of how plausible it appears to be.
Pretty close. They don't really have batteries, mostly they run directly off the generators and have very large radiators (just large fan-cooled resistors) to deal with excess power (like when braking when the traction motors are acting as generators). The drive-train's the same though, and the amount of traction they can get despite being steel wheels on steel rails says lots about how effective the results are.
The biggest thing about hybrids is that they allow for an electric drive-train. That opens up a lot of options for powering the vehicle since the engine doesn't need to physically drive the wheels. Gas turbines, for instance, with the turbine driving a high-RPM generator which eliminates the need for high-ratio reduction gears. Gas turbines, in fact any sort of continuous-combustion engine that doesn't need to be throttled to control it's speed, are more fuel-efficient than traditional IC engines. Add in regenerative braking to recover power and the ability to charge it's own batteries while parked and you get a vehicle that can have a much smaller engine without sacrificing range or performance. Maintenance costs would probably be lower too because with a gas turbine there aren't as many complex moving parts to break and they won't be under as great a strain.
Ask "Do we know all the consequences of eliminating mosquitoes world-wide?". If the answer is "No." then you DO NOT WANT to eliminate mosquitoes world-wide, for the same reason you don't push the Big Red Button before you know what pushing it will do.
Even if the court hasn't ordered it, once you're sued you have an obligation to retain anything which would be relevant to the lawsuit. It's called a litigation hold, and I work in a field where they're common. That obligation supersedes company data retention policies and requests from other parties (including the plaintiff, if the plaintiff eg. uses Outlook's "recall message" feature to try and remove a message they sent us we're required to ignore it and retain that message and the "recall message" request). It isn't supposed to require the plaintiff to prove anything, it's purpose is to preserve evidence that the defendant possesses so it's available to be found through discovery. And it applies the other way as well, the plaintiff is obligated to preserve evidence too because the defendant's entitled to discovery too.
Note that there are two things that absolutely top the list for convincing a true artificial intelligence that humans represent an existential threat to AI:
This problem is ancient. It's basically the same problem as we had in the days of MS-DOS: the biggest threat vector was clueless users who insisted on installing the latest toy without doing any checking or research. Screen saver, background changer, media player, whatever was new they insisted on having. No matter how many times they got burned, they kept repeating the same mistake. Now we have developers downloading the latest fad package without doing any checking or research, just because it claims some highly-trending keywords or something. Feh.
Can we say "positive feedback loop"? The LLM's designed to produce responses likely to follow the prompt. Producing responses that agree with and support the user's thoughts (whether rational or delusional) tend to elicit more prompts, which makes that sort of response more likely to follow a prompt than one which disagrees with the user. The more the user sees affirmation of their thoughts and beliefs (whether rational or delusional), the more convinced they are that they're correct. Lather rinse repeat until they're thoroughly brainwashed by their own delusions.
This is why engineers apply negative feedback loops to systems to keep them from running out-of-control. LLMs aren't amenable to having such installed.
Not a chance of this happening, because the current state of AI doesn't have the concept of a popular, high-probability concept not being the correct next step just because it happens to also be false.
Behind every great computer sits a skinny little geek.