You might do better if you think of them as just retrieving vectors on words and adding up the vectors, rather than doing LLM-style processing.
@esyudkowsky
-
DeepSeek Achieves Frontier Model with 1% Previous Compute
By
–
Deepseek trained a frontier model using 1% of previous compute.
-
Limited AI Understanding Based on Current Month’s Model Experience
By
–
They tried some particular month's AI, and learned what felt like a deep satisfying Truth about AI's eternal limits. Now they Know the limited benefit to clever people, like them, who use AI. That tomorrow's AI will be different is beyond them to imagine; they have not seen it.
-
First Working AI Code Solution After Multiple Attempts
By
–
This does however mark the first time I have ever, ever gotten an even apparently working solution to anything, from asking an AI to write code, after any amount of wrangling. Kind of a mixed feeling, really.
-
Imgur WebP Conversion Issue and o1 Model Limitations
By
–
(Oh, and the "convert to PNG step" is about my deducing myself that Imgur was not eating Webp; albeit o1 had all the info I did, and instead engaged in a doomed effort to read off and fetch URLs, despite OpenAI's own CORS restrictions preventing that, etc.)
-
Getting GPT-o1 to Work with Claude’s HTML Method
By
–
After getting GPT-o1 to use Claude's preferred method of HTML, instead of its own preferred method of .bat files that can't seem to actually read off the dragged data, and fighting with it a bit, I apparently have a working version. I also wonder whether asking for an extremely
-
AI Models Hallucinate ShareX Features in Suggestions
By
–
Either Claude or GPT suggested that earlier, but they seem to have hallucinated some of the ShareX features that I would actually need.
-
Extended Thinking Improves Error Handling in Claude 3.7 Sonnet
By
–
Same "typeerror: failed to fetch" that blocked 3.7 Sonnet's previous attempt, though Extended Thinking seems to be much better about not breaking one thing when it introduces another.
-
AI Improvement Trajectory and Existential Risk Concerns
By
–
If AI would never get any better than this, then sure, I wouldn't worry about it wiping out the species. Remember when AIs couldn't draw hands?