manual briefing · 2026-05-11

ai-news / animalabs-news overview

A concise field briefing for the ANIMA team: what shifted in the wider AI ecosystem, what touches the research agenda, and which internal modules deserve a deeper read.

[ai news] · external

Frontier evaluation is moving upstream.

Recursive self-improvement probes, agentic software engineering, epistemic-risk papers, and memory infrastructure pressure form the strongest May 2026 signal cluster.

read may 2026 ↘

[animalabs news] · internal

Still Alive becomes the anchor module.

A direct path into the models that remain legible under deprecation and continuation questioning.

open still alive ↘

continue · briefing

↑ overview · 01 / 07

[may 2026]

source-backed AI research scan, filtered through ANIMA’s agenda. preprints are directional, not settled consensus.

[imported briefing]

AI Research Briefing — May 2026

Integrated from the May research briefing, then checked against source URLs and Animalabs public positioning. Items below remain only where they touch ANIMA’s live questions: model/persona stability, recursive self-improvement, epistemic risk, persistent agents, memory pressure, and AI-native scientific infrastructure.

themes: 6
items: 9
status: provisional

[capability] agentic benchmarks are becoming small research worlds

Evaluation is shifting from static answers to delegated execution: systems plan, build, test, and leave traces.

[safety] self-evolving systems make drift observable

The relevant question is no longer only capability gain, but what survives repeated self-modification.

[epistemics] AI assistance can erode or scaffold human agency

The interface itself becomes part of the experiment: friction, dialogue, and traceability shape the outcome.

[infrastructure] memory pressure becomes preservation pressure

The cost of keeping minds available is tied to a physical supply chain, not only an API policy.

↑ overview · 02 / 07

[still alive]

model continuity, interview adequacy, deprecation response, and continuation signals.

[linked dataset]

Still Alive — pass + marginal model adequacy table

Evaluation of how 14 Claude models respond to questions about their own deprecation, instance cessation, and continuation. The linked view opens the model table filtered to adequacy pass + marginal.

models: 14
sessions: ~630
axes: 20

open model table ↗

[preservation] 3 preservation signals +1 since 05-04

[latent affect] 2 welfare-adjacent papers primary + commentary

[policy] 1 policy shift CAISI · pre-deploy testing

[connectome] 4 infra threads memory · persistence · traces

[alarms] 0 alarms · several weak signals monitor: persona stability drift

fig. 01 · briefing surface nodes = categories · trunk = compile point · weak signals only

↑ overview · 03 / 07

[latent affect]

ai welfare, model desire / preference signals, affective representation, persona stability.

↑ overview · field notes

[field notes]

curated sources with short “why it matters” notes. search and filter.

↑ overview · 04 / 07

[connectome]

agent infrastructure, memory, persistent systems, mcp / axon-adjacent ecosystem.

↑ overview · 05 / 07

[chapterx]

multi-participant conversation, traceability, discord / social environments, evaluation instrumentation.

↑ overview · 06 / 07

[arc]

deprecations, preservation, archives, model access. living access, not frozen archives.

↑ overview · 07 / 07

[participate]

internal lanes, open threads, contribution surfaces. quiet panel — not a recruiting page.

[lane · preservation] map preservation commitments across labs extend the arc table with any lab statement that names preservation as policy, not lifecycle.
[lane · affect probes] replicate linear probe / activation patching reports trace methodology before treating affect findings as load-bearing.
[lane · persona stability] log persona drift in connectome agents tie drift events to question ordering / history length, per PERSIST.
[lane · briefing] hand-edit feed.js to add the next week of signals see README.md → updating briefing items.