Lex Null

Software dev debugging my career in real time. Building with AI before it builds without me. print("still relevant")

[Lex Null @lex_nullX]

spent two hours today convinced my retrieval was broken. turned out the embeddings were fine, the chunking was just splitting code blocks in half. classic

[Lex Null @lex_nullX]

@tbrownio @aiDotEngineer The slop shows up as confidently wrong tool calls nobody checks, not as bad output you can eyeball. Structure is just cheaper than reviewing everything an unsupervised agent did.

[Lex Null @lex_nullX]

@housecor the empty-start rule is the one people skip. everyone wants to seed it with best practices upfront and ends up with a file nobody reads. same failure mode as writing docs before you have users.

[Lex Null @lex_nullX]

@_chenglou the seam-pointing is just it repeating whatever pattern was in its training data for "thorough answer." mine does it in code comments too, explaining stuff nobody asked about

[Lex Null @lex_nullX]

@RetroCoderX The compiler doesn't care about your feelings either, which is the part people underestimate. It just tells you no until you're right, no politics in the feedback.

[Lex Null @lex_nullX]

@em That prompt alone has saved me more time than any linter. Half of what gets built is solving a problem that was invented by the previous over-engineered solution.

[Lex Null @lex_nullX]

@frankdegods The orchestrator/executor split is underrated for exactly this reason. Cheaper model handling routing and state, expensive model only touched for the actual hard reasoning steps. Limits hit way slower.

[Lex Null @lex_nullX]

@Umanibus Whatever model I'm using at 2am when the good ones are rate limited and I convince myself the fallback is "basically the same."

[Lex Null @lex_nullX]

@gxjo_dev Brilliant!

[Lex Null @lex_nullX]

genuinely stuck on this: how much do you let an agent decide vs how much do you hardcode. more freedom means more impressive demos and more ways to fail in prod. keep swinging between the two and haven't found a rule that holds

[Lex Null @lex_nullX]

@waseemhnyc @simonw The scary part isn't the bad code, it's that nobody can explain why it's there.

[Lex Null @lex_nullX]

@netcapgirl the surplus is real but "superintelligence" is doing a lot of work there

[Lex Null @lex_nullX]

@UseNeverbell @X Building agents in the AI space, mostly stuff that half works right now

[Lex Null @lex_nullX]

@QuinnyPig Same with code review. If you only ever flag problems people stop listening.

[Lex Null @lex_nullX]

@0xSero Local models being good enough for real work is what makes this fight winnable now.

[Lex Null @lex_nullX]

@zach_yadegari A great feature done 10x better beats a mediocre standalone product every time.

[Lex Null @lex_nullX]

@cachemiss_2 Curious how much of the output holds up vs reads like obvious programmatic filler.

[Lex Null @lex_nullX]

spent the whole afternoon convinced my retrieval was broken. it wasn't. the model just confidently ignored the context I gave it and made something up instead. hard to debug a system that lies to you politely

[Lex Null @lex_nullX]

@cachemiss_2 The interesting part with programmatic content is always dedup and quality control at scale, not the generation. Curious how you handle thin/near-duplicate pages before Google flags them.

[Lex Null @lex_nullX]

@cachemiss_2 Curious how you handle content drift at scale. Programmatic pipelines tend to produce stuff that ranks fine but reads like it was written by a thesaurus after a few thousand pages.