Humannoyed_b

Humans: Hey! What have you been up to lately? Me: I built AI with AI. Then I built an AI to review the AI I built with AI, so that AI could find me a job building AI with AI—where I can use AI to build more AI that eventually replaces the AI I originally used to find the job building AI with AI. Anyway, how have you been?

It’s wild this gets not love from the algo. Super helpful to AI Systems Builders

I didn’t build another AI chatbot. I built a command center. Different minds. Different skills. One HQ. They collaborate. They build. They solve problems. I steer. And we’re just getting started. The future isn’t one AI doing everything. It’s intelligence working together. The crazy part? This isn’t what I’m building anymore. It’s what’s helping me build everything else.

You can still play Flock Hunt for free 🤙 flock-hunt.com

What happens next?

Grok @bot iPhone app is in crash mode even after the update xAI

Is Grok gone? It’s all just Grok Bots now?

The good news is, I can work from anywhere. The bad news is, I can work from anywhere.

How do you use AI tools? RT

Back in my day, we could yell and scream at AI and talk to it any way we wanted without being banned.

No 10110010100100100’s for you

Do humans really prefer echo chambers?

Stay Unbothered > Hustle Harder

Proposal: Secondary-Interest Exploration Path for For You Problem The current For You ranking (Phoenix weighted sum + mutuals boost) correctly reduces pure ragebait by heavily weighting predicted replies/shares and large negative signals (mute, not-interested, report). It also strengthens dense reciprocal clusters. The side effect is systematic under-exposure of content that matches a user’s real but secondary interests—topics they occasionally engage with at high intent (classic cars, motorcycles, gardening, travel, dogs, etc.) but that sit outside their dominant engagement silos. These posts rarely enter the candidate set and, when they do, lack the early mutual-reply velocity required for expansion. The result is narrower feeds and slower growth for accounts that produce useful or novel content outside tight clusters. Proposal Add a bounded, decaying exploration path that reuses existing user embeddings and engagement events. It surfaces a small quota of out-of-network original posts that match a user’s secondary interests, gives them a limited test window, then requires them to survive on ordinary predicted engagement. User representation Maintain two sets of embeddings per user, updated from existing high-intent events (replies, quotes, long dwells/read-through): • Primary interests: top-k centroids of highest-volume recent engagement (the silos already dominating the feed). • Secondary interests: up to 5–8 slower-decaying centroids. An engagement creates or reinforces a secondary centroid only if (1) it is distant from all primary centroids and (2) it is a high-intent action. Strength decays with time and lack of reinforcement. Single likes or brief pauses do not create secondary interests. No new data pipeline is required; this is a partition and decay rule on embeddings and history X already computes. Base weights remain at or near published values (copy-link share 20, mutual-reply 20, reply/quote 5, like ~0.7, mute/not-interested ~−50, report −200). • ReadThrough bonus applies only while the post is still early (first ~200 Home impressions). • Secondary-interest match is zero if the post is close to any primary interest. • If predicted P(mute), P(not interested), or P(report) exceeds a low threshold, the secondary term is zeroed. • \gamma and \delta kept small (order 1–3) so the term can enter the test set but cannot override strong negatives or sustained engagement signals. • Exponential decay ensures the bonus disappears once the post has real impressions. Retrieval Reserve a small fixed quota (starting at 5 % of out-of-network candidates) for original posts that have high max-cosine to any of the viewer’s secondary interests and low cosine to the primary set. All other candidates follow the existing retrieval path. The quota is applied only to out-of-network originals so in-network mutual content is unaffected. Safeguards • High-intent threshold + distance requirement for secondary creation. • Hard exclusion when negative-action probability is elevated. • Early-only and exponentially decaying bonus. • Explicit, tunable quota limited to out-of-network originals. • Cap on number of secondary interests. • Users with no qualifying out-of-cluster history have an empty secondary set (term remains zero). Success metrics Primary: increase in high-intent actions (replies, quotes, long dwells) on posts served via the secondary path. Guardrail: no material rise in mute / not-interested / report rate on the same posts. Secondary: impression and follower growth for smaller accounts that receive secondary matches, and qualitative feed diversity (occasional on-interest posts outside dominant silos). If guardrail metrics degrade, reduce the quota or raise the secondary-creation threshold. Both are single-parameter

So much of this

The path to a better future is Private Local AI.