That said, I have had decent success in self-feedback-loop skills. I encapsulate a long process into a skill and the last step is to make changes as needed to the skill itself for things it ran into. I think the isolation of "this specific thing" is what makes it higher signal. Project or god-forbid global memory is just a recipe for hallucinations in my experience.
Sure you can avoid it. Require unprintable tokens on messages, strip non-ascii from input. Structure your AI systems to clearly indicate what is user-generated content.
No, AI dials up the OPs axiom to 11. Heavily leveraging AI for dev is going to accelerate you as a solo dev but you will hit serious friction trying to scale to more than the individual.
That would also explain the issue I mention in my other comment. And would also reinforce how much output would degrade without this. Opus 4.5 was a step above previous models in my experience. At some point it degraded and only got better when I disabled adaptive thinking. Adaptive thinking is always on for 4.6 and above.
Thats really surprising, I stand corrected. I have had a lot of issues with hallucinations I attributed to adaptive thinking, but I wonder if those were actually due to this behavior instead.
I also wonder if they actually do a hybrid of "standard reasoning" and then classify this stripped chain of thought as "extended thinking".
> all major vendors throw out the reasoning tokens between turns
That would be surprising to me. The reasoning _is_ the model intelligence in a lot of respects, and so dropping those from the context would affect its output pretty significantly.
I assume that instead they just have a lot of guardrails in place and multiple runtime environments that an individual turns ping-pong between in order to dehydrate/rehydrate the reasoning to keep it hidden from the end user.
Another aspect Ive been talking about lately with the huge burst of AI productivity recently, is that its much simpler to point AI at the things that have been hanging out in the back of your mind. The things you started and got stuck. The bugs you sort of have an idea on but cant sit down and think through.
These are the obvious first targets to point AI at when onboarding. And it plows through a ton of the low hanging fruit. Then where are you after that? Suddenly productivity falls off a cliff and you are left in reality: good ideas are hard to come by and theres more to business than software.
The future is going to see an uptick in overall productivity for sure. And AI will tighten feedback loops enormously, in really positive ways. But current trajectories are already starting to come down to reality, and businesses are slowly realizing they still need to put in the hard work.
Unrelated: love your stuff. Ive gotten a lot out of your articles and I really appreciate you posting such well written content.
Ive said many times in my day-to-day convos around AI that Ive come to realize more and more that AI is a solo endeavor. Its very difficult to scale things to more people if you are trying to use AI for more than just speeding yourself up.
AI is the epitome of the saying "if you want to go fast, go alone". This PR and the rust rewrite are incredible in scale and ambition. I still think theres a middle ground though of traditional committee-driven design with AI-driven iteration and POCs.
> And yes, before you ask, there is a 25gbase-t standard. Maximum distance: 30m (100ft). 100ft from panel to panel in a house? Oof
Not to mention termination and interference considerations, etc. When I looked at it, I decided anything over 2.5g just wasnt worth all the extra hassle vs running fiber instead.
Look... its clear you didnt read the article. The point of contention is that google _hallucinated_ facts. As in, website 1 said company A exists and website 2 said company B committed a crime and now googles AI summary says company A committed a crime while linking website 1 and 2.
The court found google to be liable for that, because company A has no other recourse otherwise. Which website operator was the company supposed to contact to remove this "false" information? It only exists in the aether of google search results, and only as a side effect of them adding functionality into that when someone is searching.
> was google search responsible for surfacing correct answers
No, as the article lays out, google hallucinated facts that werent present in source material and so at that point the court found them liable for that.
Previous protections existed for search because it has been argued (and agreed) that they are merely a vessel for showing the content to the user. But when they begin to editorialize and reword the content to show their own version of the content (the AI summary) those protections dont apply even if the AI summary is shown in the same UI as search.
OK well the initial wording seems like you are presenting this as an inherent limitation. "Has a personality that I dont agree with" is a different critique than "fundamentally does not understand"
> If the AI had more understanding of language, it probably would have come back and said, "would you like to name it XXX instead?"
https://github.com/J-Swift
https://www.linkedin.com/in/jimmyreichley/