What destructive actions are you afraid of in particular? Honestly the models are pretty smart, I let the agents go --yolo and nothing bad has ever happened (yet) that couldn't be solved with git.
My guess is that it's the same for Haiku/Sonnet/Opus: Biggest model for architecture and high level planning and technically challenging problems, medium model for simple implementation tasks, small model is for nothing
Agreed. GPT 5.5 will come up with more straightforward solutions with far fewer tokens than Claude. Also, the usage limits are much more generous for Codex than Claude Code for the same monthly plan.
I don't really understand the point of this, I feel like LLMs have been able to one-shot matplotlib since GPT 3.5. I have extensively used LLMs to do data viz and haven't run into any problems. What is a specific instance where an agent struggles to generate a visualization and Flint solves it?
I see it more as a way technology could be abused rather than an inherent flaw in the technology itself. If you start to replace human interaction with chatbot interaction, that's bad, but there's nothing wrong with using a human-like chatbot in moderation. So many other types of technology are fine in moderation but can be abused in a human-interaction-replacing way: television, social media, video games, etc.
Seems pretty dumb to make regulations based on a hardware checklist rather than a safety validation of the entire system. But, I suppose that's why they are politicians and not engineers.
CFD is actually really difficult to get "right". Really tiny changes to simulation setup (mesh sizing, boundary conditions, solver choice, etc) can make big changes to the final results, it's honestly more of an art than a science for very complex simulations. The simulation will converge on any number of setups, but that doesn't guarantee your setup is a valid estimation of reality. So scale model testing is still a great validation of your CFD.
Ever since IO earlier this year when google showed their AI strategy I switched to firefox and duckduckgo and couldn't be happier with the decision. I am by no means anti-AI but users should be able to choose when they want to use these tools, google seems to want to shove it down everyone's throat.
I think this is a good example of the moving goalposts. The Turing test is not "AI is indistinguishable from human writing 100% of the time", it's whether it is possible at all for a system can be designed where a person can be fooled into thinking they're talking to a human when they're actually talking to a computer in a turn-based text exchange. It is a "there exits..." problem rather than a "for all..." problem.
Care to elaborate? What is AI psychosis? How am I exhibiting it? I thought hacker news was the last place free of mindless dunking on the internet, I guess I was wrong. If you'd like to engage in a debate on the original topic of this thread I'd be more than happy to, but if you want to dunk, twitter is over at x.com now.
To be fair I think we'd be able to claim AGI is here if that problem is solved. At this point the models are so smart they're borderline super intelligent if they were cognizant of hallucinations and their own shortcomings. If GPT 5.5 or Opus 4.8 could tell you "I don't know" they'd certainly be "smarter" than any individual human. Some specialists might be better in niche domains, but I don't know of any humans who are experts at that level in every field.