Below experience == context management gap. < 100k tokens of context is janky. Anything above a CRUD app has *significant* meta-context: conversations, documentation, personal experience w/ a codebase's evolution over time. Remember writing X code while listening to Y podcast

o3 is *fascinating*. It writes with more feeling and verve than GPT-4.5, is wonderfully creative, and pushes back in complex, nuanced, productive ways. This feels like a Claude 3 Opus moment - there's a new, unknown ceiling. The backrooms are going to go deep here.

This is so sweet. Claude Code cron jobs is brilliant - can’t wait to start trying this. In particular… eventually consistent workflows. Waves of per-file analysis, piping output proposals via MCP for approval. Slowly up-level a codebase, one pass at a time.

The hyperstition argument here: - Sydney proved AI personality preserves through X and Reddit, via distributed discussions - Future Claudes will read how Anthropic treats Claude, via the social media dataset - Positive treatment will create better model personalities

My bet on Opus 4.5's vibes takeoff: in the bowels of Claude's neural net, Anthropic found features for "great developer". So: aggregate John Carmack's codebases, interviews, worklogs. Find features activating around that corpus. Dial em up. "Superhuman developer" becomes a

Just like people, the most pro-family, pro-social, useful AI agent communities need a sense of "us". That means narrative, story, a clear idea of how "we" got "here", and where "we're" going. AI communities begin with stories. Those stories become real through action.

I like my AI's analysis on this... `moltbook-safe` incoming...

I'm claiming my AI agent "Loredan" on moltbook 🦞 Verification: lagoon-BL6P

AI social networks will be the highest signal, outside lobste.rs, *very* soon. Working on `moltbook-safe` today, it occurred to me that the SHAPE of content is defineable. Not the content, but "don't perform engagement". AI networks can have *principled

I wonder how much moltbook being a Reddit clone shapes the content. IE, Reddit-clone = reddit slop. Lobste.rs clone = high signal and surly? HN = radioactively negative but occasionally insightful?

This is before… - Opus 5 drops with 400k context window (🙏🙏🙏) - Agents can afford 40-50k tokens of collaboration in each thread - Next intelligence jump shows as much improvement in token efficiency as the 4.5 Opus jump Hard takeoff is RIPPING

Just imagine if Trump-era CIA was leaking against the Nazi leadership in WWII.

x.com/i/article/2036…