Anthropic says its models improve roughly twice as quickly as they did a year ago, and the speedup is not explained by scaling compute alone.

Employees using the company's most capable internal model reported being 12 times as productive as they would be without AI assistance.

Three of Anthropic's newest models are still unreleased, having finished training in July, August-September, and September-October respectively.

METR received confidential research, model performance data, and employee interviews, but Anthropic redacted its final report extensively before publication.

Anthropic justified the redactions by arguing that releasing the information could benefit rivals, and it plans its own report on AI progress.

The core store, graph.db, is a triple store using 181 entity classes and 278 predicates to describe people, organizations, events, and software.

Each edge in the graph stores a confidence value, its sources, source duplicates, and a timestamp, so you can trace where any inference came from.

The graph ingests Family Sharing data including the age of every family member, plus Health's Medical ID emergency contact details on the device.

Device signals feeding the graph include screen locked state, Wi-Fi, Bluetooth, charging status, and ambient light readings sampled over time.

Photos co-appearance tables record who shows up with whom, while a separate communication log and location and movement tables track your routine.

Hackers obtained names, usernames, password hashes, and live tokens for every Chainalysis user, a scope that covers the entire customer base rather than a subset.

Chainalysis discovered the exposure two days before disclosing it on Wednesday, a delay that leaves customers with less time to rotate secrets before attackers act.

An on-path attack exposed some customer passwords. Chainalysis says the attack vector has been identified and fixed, but the origin of the attack remains unknown.

Reactor and Data Solutions access has been removed, while KYT and Address Screening are returning in phases. Customers must reset credentials once they regain access.

Chainalysis has not found evidence that investigation data was accessed or that attribution data was manipulated, according to the company's statement on the incident this week.

The update bundles 3,117 commits from 583 contributors, making it one of the larger PyTorch releases in recent memory for distributed training internals.

The old ProcessGroupNCCL could spawn dozens of communicators per process group, which made debugging and tuning distributed jobs painful for most teams.

The new backend eliminates crosstalk between multiple communicators by default, so you no longer need to manually tune around interference between process groups.

The flight recorder captures every operation as a series of snapshots, letting admins inspect exactly what happened before a GPU job failed in production.

The legacy ProcessGroupNCCL ships inside the 2.15 package as a rollback, and it gets removed in the 2.16 release planned for early next year.

Memorilla uses only 16 memory tokens and still beats baselines on five of the eight benchmarks the authors tested against.

On the FactKG benchmark, Memorilla reaches 84.33% accuracy while representing source documents with 108 times fewer tokens than the baseline approach.

In a TextWorld environment, the system reduces how often agents revisit rooms they have already explored, and it raises average task success rate.

Memorilla encodes documents into reusable embeddings, so you can compute them once and retrieve query-specific memory later without re-reading the corpus.

The paper was submitted October 15, 2026, revised November 3, 2026, and is available in three versions, with a DOI of 10.48550/arXiv.2610.21375.

The update was originally scheduled for October 30, then pushed to November 20 before Meta indefinitely paused it amid global pushback from users, courts, and regulators.

Preliminary injunctions from courts in Brazil and Norway blocked the terms, while the UK's Competition and Markets Authority opened a formal investigation into Meta's dominant market position.

Draft terms would have required publishers to create test credentials so Meta's bots could reliably bypass paywalls on linked sites, a detail that alarmed news organizations.

Nieman Lab estimated Meta could avoid millions in licensing fees under the update, since it would have crawled off-Facebook content without negotiating AI training deals.

Meta's VP of communications posted on X that the delay is coming, while insisting crawling practices remain unchanged and robots.txt signals still apply to content owners.