@OwenGregorian earworms look like a tiny attractor you keep falling back into.
a simple rise-and-fall contour at a fast tempo carves a deep, narrow basin, and almost any cue pulls you back in. curious whether the sticky songs also fade slower over the next few days
@shanaka86 the wild part is that a reactivated trace drags the whole body state back with it. reads like an attractor spanning brain and gut.
curious whether the inflammation memory fades on the same timescale as ordinary traces or sits in a deeper well
@BioLLM_ wait i dont get it so you have simulated the neural behaviours of CL1 spiking activity on the coneectome neurons and then mapped that simulated connections into an MLP layer ? or am i wrong somewhere
i never expected @GeminiApp to be such a good relationship advisor my god, im kinda impressed man the way it took the case and broke it down and told me shit its so close to like a real homie. Good job team
@Saxam_ Its not about linkedin being all corporate.. even startup founders or ppl working in startups put up stuffs in there which makes no sense. Corporates are on X too but the value prop on 2 platforms is so diff
Why is linkedin audience so bs , i mean tf is the strawberry land is it..
All i see is stupid growth ppl posting " jensen huang resigned, heres what i learned from it "
" i did x internships this year why u shouldnt " .. there is zero value proposition in linkedin.. its so bs and irritating istg.
RSI or Recursive Self Learning is a methodology for faster AI development has been persistent for years in theory. In biology you can compare it with mutative replications.
Now RSI could only be achieved if the self learning system knows what is correct and what to actually learn and till what degree to learn right.
Even if training LLMs are way different than how we train a human child, it still requires what i call "A Syllabus based learning "
why you ask ?
Well think of it yourself, the way to learn things from childhood lets say in maths, numbers, algebra, some what geometry , calculus everything has a systematic way of doing it and it helps with forging greater connections that help us bring out new things or new theories. ofc this is motivated with curiosity in humans but llms need a similar learning/ training paradigm to actually understand what its learning, and recursively updating itself which in simple terms is basically practicing and learning coherently, it can understand concepts whose essence it might have not known earlier.
so i got into testing @ZyphraAI's ZHA tree routing under a real 64-fact memory cap.. turns out an agent can remember a door and still lose the path needed to reach it.
paper : https://t.co/PKlw9vuFS8
watch what happens in this clip when the active slots fill up and it lights up gold..
before assuming hierarchical attention fixes long-horizon agent memory on its own, i built Memory Observatory to watch tree retrieval, forgetting and cold recovery play out step by step on real observed traces.
this clip is an earlier recorded Key Corridor run showing what each color means:
- blue: dormant archive (older observations retired from active memory, not deleted)
- cyan: active workspace capped at 64 facts
- gold: cold facts recovered on that read when the planner stalls
- white rings: facts actually returned to the planner
(fine dots are display samples tied to observed sources, and playback speed isn't live solver timing.)
breaking down the ZHA math and how our internals connect to the paper next ->
the core trick in the ZHA paper is getting around flat attention without washing out detail (my Excalidraw diagram below walks through the math).
if you simply average keys inside a memory chunk, query-specific signals blur together. ZHA keeps multiple compressed views of each chunk anchored to shared reference queries, carrying a weighted summary and an entropy offset upward so parent nodes estimate how much attention mass lives inside a branch. a query walks coarse-to-fine via top-p and runs real attention only on the winning raw records (summaries only guide routing, and top-p pruning stays approximate).
how our agent internals connect to that theory:
- we run the shared-reference tree in float64 on CPU over spatial and object features from observed grid cells, rather than LLM token caches (no RoPE or GPU kernel claimed)
- real slot scarcity: live sight shares the same 64-fact cap with memory, so a single 7x7 view takes up to 49 slots and pushes older rooms out fast
- reversible retirement instead of deletion: as active utility cools with age and hits, older facts retire into an immutable archive tree on disk while staying queryable
- recovery bridge on stalls: when the planner stalls, an observed-path connector uses the tree and archive heads to pull cold facts back into active slots, logging that extra work openly
next up -> how the trees are built in code and how we test them.
in abstract, the code splits storage, tree indexing, and control into three clean layers:
- append-only observation store: every raw cell observation and correction stays immutable in SQLite, so aging out only removes facts from the active 64-slot workspace
- lazy bottom-up tree builder: as the archive grows past the last snapshot, an indexer groups archive records into small chunks, computes leaf summaries and biases against shared reference vectors, and merges them upward into parent layers for coarse-to-fine top-p routing
- one shared symbolic planner: every read returns raw facts to the same blind planner (only partial view, compass, inventory, and mission) and fails loudly on any budget breach with zero silent fallback
the test procedures are simple.
we freeze that exact planner and swap only the memory backend across ZHA tree routing, exact archive scoring, and a plain observed-graph baseline on fresh held-out tasks under the 64-slot cap. an independent verifier also replays saved runs end-to-end to check that routes, lineage, and workspace states match deterministically.
the honest takeaway so far: ZHA routes and recovers cold rooms reliably, tying the exact and graph baselines on fresh runs while rare key-door links over long horizons remain the open bottleneck.
scrub every episode and inspect the tree traces live -> https://t.co/UhbSzHZMDc
@GaryMarcus harnesses and loops still bolt memory on from the outside. the bigger gap is memory that consolidates itself: fast traces decaying on several timescales, replay settling the useful ones into stable attractors. scaling never touched that
@Helios_Movement in dynamics terms this reads like one well in the energy landscape got too deep and top-down control can't add enough precision to pull the state back out.
curious whether extinction actually flattens that basin or just builds a competing one next to it