@alxfazio normie workflow fr, but maybe structured chunks solve some hallucination, anyone see better approaches yet?

@leodoan_ this is literally the basics dawg

@alxfazio RAG discourse has been happening for 2 years normies still don’t know it exists they think AI is just “upload file, ask question” the gap between what’s possible and what people actually know how to use is massive “we’re still early” undersells it most people don’t even know

@alxfazio lol the fact that people still dont understand context windows and chunking strategies tells you everything like bro you wouldnt read a 100 page doc in one sitting and expect perfect recall either. gotta break it down and use the right tools for the job

@alxfazio Then the LLM should tell them that the pdf is too big. What a strange inversion of responsibility.

@alxfazio They also think it immediately learns from everything uploaded to it and adjusts all of its future generations.

@alxfazio I'm personally waiting for what the next step is. Remember, kurzweil mentions that all throughout history the stats are in agreement: the improvements continue despite the changing of tech. Not that I'm unimpressed by LLMs, but even ask GPT5 or Grok4 what their thoughts are and

@alxfazio You do realize that every bit of marketing copy available confidently said that this ALREADY WORKED in 2023? The limits of LLMs are somewhat popularized now only because enough SWEs tried to use Claude Code for their projects and saw it stating to choke at 1000 LoCs.

@alxfazio My flow. I'm sure there's something better but it works. Throw the codebase at the model in a single markdown file. Have it debug. Have it tell me which files need fixed. Give it those exact files in the "fix it" prompt. Haven't had any issues.

@alxfazio Ah yes, the classic "let me feed War and Peace into an AI and then act shocked when it makes up a chapter about flying toasters" workflow.

@alxfazio "I don't get it! I completely overwhelmed its memory in 3 seconds flat and it started hallucinating. How could this happen?!?"

@alxfazio We’re way earlier than everyone seems to think

@alxfazio Or worse trying to invalidate 100 page frameworks in one shot.

@alxfazio Do you believe i can make it NOT hallucinate?

@alxfazio they don’t even know what llms are yet 😭🙏

@alxfazio Tech twitter is the real bubble.

@alxfazio What irritates me most are the so-called AI influencers dunking on gpt-5-mini for not grasping the subtleties of case law.

@alxfazio We're eventually all going to realise that asking an LLM to be lazy for us in the laziest way possible is going to end up with them taking shortcuts. They're getting more like us by the day.

@alxfazio blame the lie sellers

@alxfazio Genuinely what is the point of an LLM if it can't do the work people want it to?

@alxfazio as of nov 2025, chatgpt still treats pdfs like plaintext graveyards. it just yeets the rest, all the figures and images inside them. give it a scanned paper and it’ll just errors 'no text detected' 💀

@alxfazio It didn't hallucinate dude, it averaged that pdf out

@alxfazio Can confirm

@alxfazio vibe parsing

@alxfazio hey wtf?

@alxfazio we’re early like “explain temperature to my uncle” early