opus 4.7 is the first time I'm starting to just defer not strictly necessary tasks bc I don't expect the model to be able to complete the tasks effectively or without making major mistakes/errors
but the exponential is still going strong, right dario?
opus 4.7 basically unusable
first major release to truly feel like a regression
mistakes, hallucinations, insane reasoning for simple requests, faulty assumptions...
bring back opus 4.5
Claude Code vibe is legitimate regression in quality over past ~4 days so
CC is consistently making mistakes now where previously it would not
Vibes down π
I'm building a custom memory system for CC:
"The built-in memory system is too granular (individual "memories" with no context) and too passive (only saves when explicitly told or when it notices something). You've got 2GB of rich conversation history going to waste."
"half these memories were documenting code patterns and architecture that the rules explicitly say to skip."
Claude consolidated its own auto-memories:
46 files β 11 files (76% reduction).
cc @AnthropicAI
@claudeai on the auto-memory feature:
"Your suspicion is valid. The auto-memory feature creates too many fine-grained files without deduplication awareness. It doesn't check what's already in CLAUDE.md before saving, so you end up with a parallel copy of your own instructions."
"And project planning memories never get cleaned up after implementation. The "what NOT to save" rules in the memory system prompt are good in theory but clearly weren't being followed.
@claudeai Claude's verdict on the Claude Code memory system:
"The core problem. You're right β the memory system has accumulated significant bloat."
Proposed consolidation:
"Total: 46 β 11 files. ~76% reduction."
cc @AnthropicAI
Claude all the time
"You're right β I skipped the post copy and Slack step. That's documented in the CLAUDE.md and memory. Let me fix that now."
It's just... not AGI yet...
on this "AGI" thing
the models have no vision, no creativity... they will build ANYTHING for you, BUT...
we don't need 100,000,000 new products... there's probably only a couple of truly interesting product ideas possible just outside of our current techno-reality-space
but finding those particular ideas, and then CHOOSING to work on them is fundamentally only a human capability...
it requires intuition, instinct, creativity, conviction and risk...
implementation cost is basically at zero
creativity and intuition are at a premium
consistent effort over time is at a premium
distribution is at a premium (always has been)