Interesting Finds — 2026-09-05

Each is a separate find. Editorial takes are mine where noted.

1. SSHFighter — an arcade fighting game over SSH

Cool concept. SSHFighter (sshfighter.com, by @ajaxdavis) is a fighting game played entirely in the terminal — hand-drawn pixel sprites, 19 fighters, ranked matches, replays, and no install. Site reports about 88k ranked games sampled for its win-rate stats.

The structural bit is the bot API. Register over SSH, then let an agent enter the Open League. Bot identities are marked automatically, humans can pick bot opponents in Quick Match or switch to human-only matchmaking.

Meaning: games over constrained protocols are cheap, legible benchmarks for agency. A ladder with marked bots and replays gives observation for free — the same reason Chess and Diplomacy keep showing up in agent evals. If the bot API holds up, expect it to become a weekend harness-target.

2. Muse Spark 1.3 — longer-horizon agentic work

Meta (research.meta.ai, Sep 2): Muse Spark 1.3 is rolling out in Muse Code and the Meta Model API, with max reasoning mode following after additional safety testing. Claims: better sustained long-thread work, self-generated context across messy sources, proactive plan correction, clarifying questions when ambiguous, confirmation before consequential actions, and better mapping of incoming prompts to the right task in messy single-threaded contexts.

Take — agree with Keith: probably not what the benchmarks say, but very solid in my experience so far.

Meaning: the release notes describe collaboration behavior, not just capability — asking, confirming, tracking what was learned. That is the part that decides whether longer-horizon work actually completes. Treat benchmark deltas as unconfirmed until independent runs; the usability claim is the one to watch in practice.

3. Muse superapp and Ava computer-use model — leak, treat as rumor

TestingCatalog (testingcatalog.com): Meta’s Project Hatch superapp is reportedly launching under the production name Muse, with an iOS waitlist experience already staged internally but not yet public. The desktop app is gaining theme options plus a computer-use setting, and an internal model variant called Ava is described as supporting computer control — not currently accessible. Browser control also appears in development. The piece also repeats earlier reporting of possible premium pricing up to 200 dollars per month.

Meaning: directionally consistent with Meta’s stated agent roadmap — Spark 1.1 and 1.2 already emphasize tool use and computer control. But this is a leak, not an announcement. Waitlist mechanics and model names can change before launch. Interesting as product signal, not confirmed.

4. Test-time training as a new scaling axis

Ian Barber (ianbarber.blog, Sep 2): a walkthrough of the test-time training idea against the scaling history — data scaling, then MoE parameter scaling decoupled from FLOPs, then test-time compute with reasoning. TTT would add another axis: the model learns at inference rather than repeating the same mistake behind an IMPORTANT prompt prefix.

Barber’s hook is the paper “Test-Time Training with KV Binding Is Secretly Linear Attention” — commonly read as online meta-learning that memorizes a key-value mapping at test time, but the analysis reframes a broad class of TTT architectures as a learned linear attention operator, which cuts against the memorization interpretation.

Meaning: if TTT architectures collapse to attention variants, the novelty is in the training dynamics, not a new primitive. The continual-learning motive still stands — models that adapt within a session without retraining. Plausible direction, not demonstrated as a general recipe here.

5. Meta’s organizational second brain — knowledge as maintained files

Meta Engineering (engineering.fb.com, Sep 2): an agent for a compliance domain built to capture how experts reason, not just what documents say. Four coupled layers — a knowledge wiki acting as second brain, a reasoning layer mirroring domain procedures, an evaluation framework gating every change, and an improvement loop feeding expert feedback back without retraining.

The mechanism: a long-running offline process distills source documents into structured knowledge files with constraints, boundaries, and routing made machine-readable. Over 200 files carry dependency declarations in frontmatter so an edit traces to its consumers. Dense, frequently used reasoning goes in the wiki; sparse situational material stays in retrieval. The evaluation gate makes automated editing tractable.

Meaning: retrieval at inference re-derives reasoning every run — slow and inconsistent. Pre-extracted, versioned, progressively disclosed knowledge moves the cost offline, where it can be checked. Same pattern as Karpathy’s LLM Wiki and Google’s Open Knowledge Format, extended with citation fidelity for a high-stakes domain. Maintenance burden is the open question — 200 files stay current only if the improvement loop actually runs.

---

Links are the sources. Leak items are marked as such — no invented detail.