pureion

Libraries have always been the source of knowledge and curators of knowledge. It puzzles me that they are not pursuing conversion into regional compute centers, accessible by the public. The public being those who cannot afford compute.

The tendency of the models to create pseudocode in the work as you vibe code... Natural language in place of syntax, etc. as it holds in structure the other mature parts. Perhaps we need a pseudocode safe language/mechanism/registry so the model keeps tab on the codespace where it uses pseudocode, to be filled in later on with real code.

I was going to just reply but this got me thinking. Others should hear it too. If you get things done on plus or the lower paid plan, that is a sign of your AI skills. Constraints invites innovation and this is exactly how we keep our brains. If we just keep forking it over with the 200 or whatever... Are we really thinking? Are we really exercising our brains to get things done, to understand how our processes can get better to work within what is available. I would think that it is coveted more by employers that we know how to really get things done that is more business wise. Put yourself on the plus level. Challenge yourself. I am.

AI mobility matters in this too. The rate of adoption of a nation towards AI use does too. More people doing more experiment finds more methods that works better. A very small % of us corporations use ai regularly in their business. That's going to affect us.

Was frustrated that Sol medium was eating my 5 hours window for breakfast on Codex. A lot of that was wasted on computer run tests. Tried something new: let Luna medium handle the repo/terminal/test-heavy work, while Sol does the reasoning parts. Results: Sol-alone burn: ~1.2 percentage points of the 5-hour limit per minute Sol/Luna coordinated: ~0.4–0.6 points/minute most runs and the output quality stayed the same, if not better. Roughly 2–3× better usage efficiency so far.

I am noticing that the knowledge generation may not be as free as we think it may be. Once you make a correction, the system starts to build defensive behaviors around that correction, then that poisons the architecture downstream. - Everytime you correct, it guardrails itself. - The wording matters a lot. When you say things in a certain way, it will maximize the narrowness of that application of what you just said. - When you try to make genuinely new architectures or approach, the models will almost always force the coding process to apply it with current architecture practices when you are coming up with a completely original approach, which contiminates the results. This leads me to think more about the advanced science, how much of it is being forced to act within bounds by its own limited knowledge.

One thing I notice, semantic accretion happens within vibe coding. You build a conceptual approach. You define the parameters. Then when it comes to build time, the model keep inserting its own software engineering architectural norms underneath to help structure it and it is redefining how it looks at your architecture the more you work in the build. You end up with a product that does not look like what you intended to build.

Robots will help make the world accessible for us. We need to answer the question of the accessibility of robots. Can the robots access our world? Doors, curbs, elevators, stairs, uneven terrain conditions, etc. Can we keep working with them if the normal voice commands stop working? How do we know what they're saying if the speaker stops working? How do we access robots that are not controlled by us when they're experiencing software or hardware failure? Do robots even have privacy? If not, how open is it? How closed should it be?

Today: Build the costume to attract buyers. Future: Agents find the best capacity. Your focus shifts entirely to delivering high-quality, timely results. The costume becomes irrelevant. Agents focus on getting the job done the way the person asked for. What capacity or resource can you meet/build?

New form of AI hallucination: Summarized/compressed and committed memory. You keep repeating your own build specs while the memory keeps moving the goalpost with its own shell version of memory. Ghost in the system. Equivalent to a faulty calculator in a research lab.

ChatGPT's constant summarizing is becoming an issue. 5.5 seems to be really bad with this. Its more capable but at the same time I can feel the limiting factor of how the memory works. This is not good for precise work involving words, meaning, how it is persisted across sessions. The memory architecture is wrong. It feels like a new form of hallucination, just this time through memory. Everytime you work, a summarized version of your work is committed to memory. Compressed or otherwise altered. This is a shell that you're allowing to be established. Every shell versions from there on carries you further from the original. You're left with a product based on representations of representation on top of many representations and you constantly fight to bring it back in line each time as the system keeps building with the shell in mind.