Quote from: InterestedEngineer on 11/27/2025 04:32 pmQuote from: Twark_Main on 11/25/2025 03:27 pmQuote from: InterestedEngineer on 11/25/2025 02:48 pmAI may be a short-lived phenomena. As soon as it starts getting fed its own output, it quickly degrades to garbagehttps://alwaysthehorizon.substack.com/p/urban-bugmen-and-ai-model-collapseThis is a speed bump, not a brick wall.This is the modern equivalent of saying "64K ought to be enough for anyone." It's based on assuming a certain technological paradigm will continue forever, but that's not how it really works."that's not how it really works".Do, indeed, tell us how it works, how one avoids the AI feedback loop that will generate garbage in the long run, which numerous experiments demonstrate.Simply scaling data size has reached diminishing returns. The innovation in the next 18 months will be algorithms, not "throw more [synthetic] data at it."
Quote from: Twark_Main on 11/25/2025 03:27 pmQuote from: InterestedEngineer on 11/25/2025 02:48 pmAI may be a short-lived phenomena. As soon as it starts getting fed its own output, it quickly degrades to garbagehttps://alwaysthehorizon.substack.com/p/urban-bugmen-and-ai-model-collapseThis is a speed bump, not a brick wall.This is the modern equivalent of saying "64K ought to be enough for anyone." It's based on assuming a certain technological paradigm will continue forever, but that's not how it really works."that's not how it really works".Do, indeed, tell us how it works, how one avoids the AI feedback loop that will generate garbage in the long run, which numerous experiments demonstrate.
Quote from: InterestedEngineer on 11/25/2025 02:48 pmAI may be a short-lived phenomena. As soon as it starts getting fed its own output, it quickly degrades to garbagehttps://alwaysthehorizon.substack.com/p/urban-bugmen-and-ai-model-collapseThis is a speed bump, not a brick wall.This is the modern equivalent of saying "64K ought to be enough for anyone." It's based on assuming a certain technological paradigm will continue forever, but that's not how it really works.
AI may be a short-lived phenomena. As soon as it starts getting fed its own output, it quickly degrades to garbagehttps://alwaysthehorizon.substack.com/p/urban-bugmen-and-ai-model-collapse
https://twitter.com/leap_71/status/1989725508855287907QuoteNot designed by aliens. Designed by "the first AI that builds machines" as the press has called Noyron, our Large Computational Engineering Model. This hypersonic precooler we produced with @FarsoonAM uses a fractal folding algorithm to chill down superheated air.QuoteChromeKiwi@AshleyKillipAi engineered precooler. This is one thing i like about Ai being able to design new ways to do things and since it's computer based can likely do all its own flow path analysis and model simulations for total efficiency or material weight without compromising the structure. Maximising form and function.
Not designed by aliens. Designed by "the first AI that builds machines" as the press has called Noyron, our Large Computational Engineering Model. This hypersonic precooler we produced with @FarsoonAM uses a fractal folding algorithm to chill down superheated air.
ChromeKiwi@AshleyKillipAi engineered precooler. This is one thing i like about Ai being able to design new ways to do things and since it's computer based can likely do all its own flow path analysis and model simulations for total efficiency or material weight without compromising the structure. Maximising form and function.
Installed claude code this last week.A coding task that would nominally take a day or two took an hour - worked on third attempt (the only real problem was due to something neither I nor the AI anticipated). This is me using the tool for the first time and understanding nothing about its gotchas.It was good code. I had some nits about the coding style that were easily addressed.Watching it think is like your listening to your inner dialog when working out a problem. It even made a coding mistake I had made sometime in the past and fixed it in front of my eyes with no intervention.In 2025, AI got an inner dialog. This is an world-changing breakthrough. Larger models were showing asymptotic results. This is not an asymptotic result, this is a step function improvement.It'll get better once the speed of thinking improves and the cost comes down, because it'll be able to do more rounds of inner dialog. (There will probably be an asymptote to that too).Probably should anticipate more step functions coming down the road.TL;DR my coding productivity went overnight from 1.5x (AI "suggestions" and small amounts of code creation) to 4x productivity improvement.I've been trying to get AI to improve the algorithm for the multi-body trajectory problem particularly in cases where the hyperbolic orbits to planets break the existing pork-chop methods. So far that has been a failure, the AI didn't have enough real-world examples to get it right If i get a chance during the holidays I'll try to get it to tackle this problem again (I'm trying to see if polar coordinates makes the problem simpler/faster/easier to understand). I suspect the introspective AI is going to allow me to make more progress more quickly.
Quote from: InterestedEngineer on 12/21/2025 09:09 pmInstalled claude code this last week.A coding task that would nominally take a day or two took an hour - worked on third attempt (the only real problem was due to something neither I nor the AI anticipated). This is me using the tool for the first time and understanding nothing about its gotchas.It was good code. I had some nits about the coding style that were easily addressed.Watching it think is like your listening to your inner dialog when working out a problem. It even made a coding mistake I had made sometime in the past and fixed it in front of my eyes with no intervention.In 2025, AI got an inner dialog. This is an world-changing breakthrough. Larger models were showing asymptotic results. This is not an asymptotic result, this is a step function improvement.It'll get better once the speed of thinking improves and the cost comes down, because it'll be able to do more rounds of inner dialog. (There will probably be an asymptote to that too).Probably should anticipate more step functions coming down the road.TL;DR my coding productivity went overnight from 1.5x (AI "suggestions" and small amounts of code creation) to 4x productivity improvement.I've been trying to get AI to improve the algorithm for the multi-body trajectory problem particularly in cases where the hyperbolic orbits to planets break the existing pork-chop methods. So far that has been a failure, the AI didn't have enough real-world examples to get it right If i get a chance during the holidays I'll try to get it to tackle this problem again (I'm trying to see if polar coordinates makes the problem simpler/faster/easier to understand). I suspect the introspective AI is going to allow me to make more progress more quickly.Why did you go with Anthropic's Claude in particular?
Yeah, Claude is pretty incredible. I haven't used it, but in terms of what length of tasks it can handle (at 50% success rate), it's a massive leap above everyone else. https://x.com/emollick/status/2002208335991337467?s=20(Really makes me want to try it...)
Quote from: Robotbeat on 12/22/2025 04:11 amYeah, Claude is pretty incredible. I haven't used it, but in terms of what length of tasks it can handle (at 50% success rate), it's a massive leap above everyone else. https://x.com/emollick/status/2002208335991337467?s=20(Really makes me want to try it...)Just to be clear, I'm referring to the *agent* version of Claude Opus 4.5, which is called `Claude Code`.no more copy and paste from a chat window. High level requirements and some details in, code out.I have used Claude-Opus 4.5 using the chatbot version and it is definitely better than GPT 5.x.But the agent is a whole other level of productivity improvement.
Will AI Ever Be Conscious?https://youtube.com/watch?v=VAFEmFSMfTg
