Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

Janus v1.0 ships on three desktop platforms
08/10/26, 12:28 PM

Janus v1.0 ships an open-source agent laboratory across three desktop platforms, with downloadable builds for macOS, Windows, and Linux. The first public release, published inside the last 24 hours, combines persistent multi-agent workspaces with review, versioned skills and memory, controlled rollout, adoption, and rollback. For operators exploring durable orchestration, its governance model is the part to file away: evaluate those controls against the bounded-recovery gate proposed for agent runtimes before treating a polished desktop shell as production readiness.
Wire

Lophius opens six model-inspection views
08/10/26, 12:27 PM

Lophius v1.0 opens six inspection surfaces for local language-model research: architecture, prompts, responses, per-token logits, attention matrices, and projected hidden states. The AGPL-licensed initial release, published inside the last 24 hours, can load multiple Transformers models together and batch the same prompt set across them. For teams evaluating open checkpoints, that makes behavioral differences inspectable below the final answer—useful evidence before paying the control premium described in the archive’s open-weight model analysis.
Wire

Pi from Scratch rebuilds an agent in 765 lines
08/10/26, 12:26 PM

Pi from Scratch rebuilds a working TypeScript coding agent in 765 source lines across five files, with another 401 lines of tests covering the loop, tools, CLI, LLM client, and terminal UI. The MIT-licensed repository and interactive trace tutorial were published inside the last 24 hours and expose read, edit, and shell tools behind an OpenAI-compatible API. For builders deciding what belongs in the harness rather than the model, this compact implementation is a useful foil to the five-task performance swing measured across coding-agent harnesses: study the loop before adding orchestration layers.
Wire

Meta fits a 30B agent model into 24GB
08/10/26, 11:20 AM

Meta released Muse Glimmer, a 30-billion-parameter open-weight agent model whose 4-bit build fits a 24GB memory envelope and reaches a reported 233.4 tokens per second on an RTX 5090 with its speculative-decoding drafter. The Apache 2.0 release says DFlash raises RTX 5090 decoding speed 3.1x, while the model card reports 51.2 on SWE-Bench Pro and support for text-plus-image input. For teams following the economics of open models that fit smaller hardware, this puts a measurable local-agent baseline on one consumer GPU—though Meta’s own safety results still argue for external guardrails before granting tools.
Wire

Blast Radius cuts coding-agent tokens up to 26%
08/10/26, 10:24 AM

Blast Radius cut coding-agent token consumption 17% to 26% across seven OpenAI models by archiving context it predicted the next prompt would not reach. The new preprint says its eviction remained byte-exact reversible; of 450 archived records, 378 were recurring dead matter and none was recalled in the evaluated workloads. Teams embracing lean command-line agent workflows should treat the result as a promising cache policy, not a deletion license: reversible burial makes an efficiency mistake recoverable when the next repository behaves differently.
Wire

TEPA keeps agent memory at 95% through reversals
08/10/26, 10:23 AM

TEPA kept agent accuracy at 95% through controlled fact reversals, while append-only and last-write-wins memories fell to 21% and no memory reached 30.9% across 50 seeds. The new preprint’s revocable evidence store invalidates superseded facts without deleting their audit history, and repeated the 95% result in a file-backed execution test. For builders already rationing context as an operating cost rather than an infinite resource, the result says memory needs an explicit revoke operation—not merely better retrieval—before an agent can safely learn from changing environments.
Wire

Meta promises free superintelligence to billions
08/10/26, 10:22 AM

Meta says it will offer free personal-superintelligence tools to billions of people, while selling heavier compute through a dynamic auction designed to return the lowest available price. Mark Zuckerberg’s new platform manifesto also sketches agents working 24/7 and a private mode that even the provider cannot inspect, but gives no launch date or price schedule. Operators should file the auction pledge beside Meta’s existing compute-and-distribution strategy: access may be free, yet scarce inference capacity will still be metered and allocated.
Wire

Decart nears a reported $6B to $7B sale
08/10/26, 07:23 AM

Decart is reportedly in advanced sale talks at $6 billion to $7 billion, only three months after raising $300 million at an approximately $4 billion valuation. Calcalist’s exclusive report says a deal could be signed within a week, although the buyer is unidentified and Decart declined to comment; the startup has raised roughly $450 million and employs about 100 people. Operators tracking the capital race around world models and real-time simulation should treat the price as a strategic signal, not a completed transaction: differentiated video infrastructure may now command acquisition premiums before its revenue model fully matures.
Wire

A deadline cuts an AutoML win rate to 34.3%
08/10/26, 06:25 AM

A nominally 60-second AutoML system consumed a median 120 seconds, and its apparent win rate fell from 59.4% to 34.3% when researchers enforced the deadline and stopped selecting candidates on the test split. The Winning by Peeking audit found no significant pairwise advantage after equalizing compute and moving selection to validation data. The lesson sharpens the procurement case for pinned harnesses and raw traces: externally enforce wall-clock budgets and isolate the test set, or a precise leaderboard can still rank protocol defects.
Wire

FinRank caps SEC retrieval at 44.8%
08/10/26, 06:24 AM

FinRank’s 1,185 manually authored questions across 10-K and 10-Q filings from 22 companies cap the best tested retriever at 44.8% Recall@10. The new evidence-grounded benchmark reports that replacing random negatives with confusable passages from other sections, periods, and peer companies cuts pairwise accuracy by 13.0–20.5 percentage points, exposing answers that sound right but cite the wrong disclosure. For teams building enterprise RAG, the result extends the case for evaluating the whole retrieval-and-orchestration system: test entity, period, and provenance with hard negatives before trusting answer accuracy.