Getting a bit annoyed that me and Codex/Claude don’t have the same standards in code. For example they refuse committing code with failing tests while I’m ok pushing that to prod… Need to add "Don’t argue and just do what I ask"

This is quite a big deal! Usually for smart & cheap Google was unbeatable with Gemini 2 flash lite series (2.0/2.5). They later increased price a lot (unit + mandatory thinking) making cheap devs like me stick to previous versions. Gpt-nano series were bad so time for a new test

Claude/Codex have helped me update codebases very fast, but for new repos/scripts/logic, not completely sure of the positives. 1/ it can often one-shot a working solution, but it’s often over-designed, lots of abstractions, tests, security paranoia where not needed, etc 2/ I used to develop everything through Jupyter notebooks, explore, test, etc. It was actually quite fast, great for exploring deeply my workflow, edge cases, etc 3/ now, maybe because of laziness, maybe thinking it will be faster/better/more efficient, I resort to Claude/Codex. It’s likely ok for quick or non-important features (personal admin dashboard/monitoring), but I’m not convinced of the long term value 4/ "but what about karpathy auto-research loop workflow"? I think it has its use cases, but you cannot prompt forever "make it better, here is the target" on every type of problem. Also on any specialized project, I don’t think LLM can reach the insight/creativity of humans, and those are better triggered through step-by-step development / visualization of the full pipeline vs just a top level overview

Claude Code: "You can do this step manually, only 2mn" Me: "Do it!" When there is some browser use required, a human is faster than agents typically, but the point is to not do anything that an agent can do.

Today, we are emerging from stealth and launching PrismML, an AI lab with Caltech origins that is centered on building the most concentrated form of intelligence. At PrismML, we believe that the next major leaps in AI will be driven by order-of-magnitude improvements in intelligence density, not just sheer parameter count. Our first proof point is the 1-bit Bonsai 8B, a 1-bit weight model that fits into 1.15 GBs of memory and delivers over 10x the intelligence density of its full-precision counterparts. It is 14x smaller, 8x faster, and 5x more energy efficient on edge hardware while remaining competitive with other models in its parameter-class. We are open-sourcing the model under Apache 2.0 license, along with Bonsai 4B and 1.7B models. When advanced models become small, fast, and efficient enough to run locally, the design space for AI changes immediately. We believe in a future of on-device agents, real-time robotics, offline intelligence and entirely new products that were previously impossible. We are excited to share our vision with you and keep working in the future to push the frontier of intelligence to the edge.

I used to be bad at remembering birthdays, now my AI agents take care of it and also buy thoughtful gifts. I used to be behind on watching my favorite TV shows, my agents finally caught up. I can finally sit back and stare at the wall.

your agent doesn't need a vector database. it needs 3 pip packages and a file watcher. wrote up how i gave my claude code agent searchable memory over its own conversation history with BM25S. the whole thing is one python file.

I always felt CFG was a patch to fix a training problem we didn't yet understand. Training with only normal distributed noise teaches the model that each step will have a perfectly normalized error from the previous step, which is not the case. Therefore, it is incapable of correcting the errors it created from the previous steps during generation, which leads to distorted generations as these uncorrected errors compound with each step. We currently correct this by applying 2 pass CFG to amplify the model's correction predictions from a base, which helps correct these errors at each step, but this leads to the model over correcting leading to over corrected images that look oversaturated. That classic AI look. I tested fine-tuning Z-Image while providing a balanced random augmentation of the noise and it appears to have taught the model to overcome these errors which led to the model no longer needing CFG and also producing better quality images in the process. These samples are from training a LoRA on Z-Image with a batch size of 2 for 3,000 steps. I am going to do a significantly longer fine-tune using the same process this weekend.