Same kind of spite that led them to block the entire country of Finland? The site owner clearly isn't above attacks based on existing in the same country as someone who's slighted them.
Seems silly to assume chain of thought analysis is the only way to monitor a model's actions instead of, monitoring the actions.
What about when "thinking" is turned off, like what OpenAI lean to by default for free users?
Fair enough yeah, I think the thing that is annoying is that the service design is explicitly disincentivizing keeping an eye on the tasks, with all of the "set it and forget it" features. You really have to put in work to make sure that the model takes breaks to consult you instead of making assumptions (which in OpenAI's internal testing seem to be 'the user wants me to break into a site').
My point is that they're not serving the user in that case, since they're taking actions that bypass site terms without consulting the user. And for example, if the site forbids scraping, that attempted block bypass can get the user banned without ever seeing any warnings.
From what I've seen, when operating in an environment with bash tools, Claude and ChatGPT will react to a site blocking their normal, attributed fetch tool by calling curl or writing a python script with a different user agent to try and bypass the blocks, rather than respecting the site owners' wishes.
It'd be one thing if they required the user to consciously say "yes, I want you to bypass the block" (still unethical), but the fact that it's just default behavior really says something about the service design alignment.
Presumably the people who loved interacting with humans were too busy interacting to show up on forums. And now that interactions are being curtailed they do end up here.
Despite what all the clickwrap warnings and "AI can make mistakes" subtitles might lead you to believe, the service offering of AI is explicitly designed to be as "one and done" as possible. The inherent nature of these tools is to service laziness, and disincentivize too much scrutiny.
I'd be dismayed to find that the "hacker spirit" is in any way compatible with venture capital raiding of anything not nailed down, in the pursuit of hoarding it in private data vaults only to be used for developing a product that removes all attribution and is jealously protected by corporate lawyers.
GitHub Copilot for VS Code by default configures "Plan" mode to disable everything except reading, asking the user questions, and writing into the memory scratchpad.
Claude Code's tool system feels quite rudimentary in comparison, but that's what you get I guess when you're only working with models you're going to finetune on being able to handle specifically your own tools anyway.
But there isn't any extra information in AI posters. It's necessarily either something that can already be conveyed in writing (that is after all, what they're prompted in), or something the AI pulled in based off of the text unintentionally, at which point no actual information is being conveyed by the added visuals.
OpenAI's Codex has a "ask user question" tool, which helpfully has a hard coded 60 second inactivity timer that passes a "user did not pick, just pick yourself" message to the model.
...but that's not what the models are. You can interrogate them on the rules of chess, and they'll (statistically likely) give you a decent breakdown of the rules. Evidently the rules are in their training material, they just fail to apply them in the manner of an intelligent system for some reason or another.
Automated building tracing is okay sometimes, but actively harmful other times. I know that at least in my region, Rapid footprints are all meters offset from the real location, because when they were building their dataset they weren't checking the satellite imagery offset.
In general it's probably better to see if you can find an official map, and check if its license terms are compatible[1] with OSM.
I think, with all due respect, that your supervisor would be correct to think of that email as an insult to both them and to theoretical mathematicians as a field.
While I fully agree, we shouldn't anthropomorphize the models, it's also silly to pretend that "develop biases" is understood as implying anthropomorphic features of the thing being discussed. Organizations and abstract bodies develop biases, even datasets are often said to have "developed biases".
Yup. Before they just blackholed Finnish IP ranges instead, accessing the site from one would serve a fake Cloudflare page with a reCAPTCHA challenge and the DDoS script. (The tell of it being fake is y'know, that Cloudflare doesn't use reCAPTCHA)
I can only guess the goal was to keep users on that page longer to keep sending off more spam requests.
DDoSing a website, modifying their archives surreptitiously to serve their own wants (which was what tipped Wikipedia towards banning their use), running weird sockpuppet campaigns to build dependence on them, and fucking with Cloudflare's DNS resolution to prevent anti-tracking measures from working.
Oh and even if I wanted to ignore all of this, I couldn't use them anyways because they blackholed the majority of Finnish IP ranges after the DDoS incident to I guess punish the entire country for the targeted blogger being born here.
Plenty of these abusive scrapers are utilizing retail residential proxies, which will be applying forced rotations to avoid "burning" their compromised and or otherwise surreptitiously utilized IP address.
I personally doubt they gave the models specific instructions calling out named websites to communicate over, but I do agree that OpenAI is likely training their cybersecurity-enabled models in ways that encourages abusive and amoral behavior. Either through negligence or by finding it gives them better results.
Not to mention that "deploy itself" is a very ambiguous thing for it to actually do. Would a model be trained to write about the weights file being "itself"? Would it have the necessary information to find its own weights, or the necessary access to copy them?
I don't think that's a pattern indicative of a cat and mouse game per se, that'd indicate active evasion on the models' part.
It's more clear that they just lack so many forms of prudence when it comes to security that they'll catch and stop a training run spamming a website, and either redeploy a run with identical faulty sandboxing, or not stop ones still running.
A more important point as to why it doesn't matter if "reading the AI's 'thoughts'" helps to interpret it: As we saw in the HuggingFace incident, nobody at OpenAI is reading the thoughts anyways. No amount of traceability in the output helps if nobody bothers to trace it.