Monday, July 6, 2026

Databricks argues the ETL pipeline is avoidable

Databricks co-founder Ali Ghodsi describes the architecture of Lakebase, a serverless Postgres built on the acquired Neon codebase, and its extension he calls LTAP (Lake Transactional/Analytical Processing). Lakebase splits Postgres into stateless compute plus two storage services: SafeKeeper replicates the write-ahead log across a Paxos quorum for durability; PageServer materializes pages into cloud object storage. LTAP has PageServer transcode row pages into Parquet, Delta, or Iceberg during materialization, preserving Postgres type semantics and MVCC history. Analytical queries pin to a log sequence number fetched from Postgres, read the bulk from object storage, and merge the unmaterialized tail, so Postgres serves no analytical reads. The argument: this removes change-data-capture entirely, because there is no second copy to replicate into. Ghodsi reports 5x write throughput and 2x lower read latency versus synchronous-replication Postgres. These are vendor numbers with no independent benchmarks, and Databricks is currently dual-writing both formats during rollout.

The obvious levers have measured ceilings

Armin Ronacher reports a regression: Claude Opus 4.8 and Claude Sonnet 5 emit invented JSON keys in nested tool-call arguments while getting the payload byte-correct, a failure absent in older Claude models. His hypothesis is a training artifact: reinforcement learning against Claude Code’s forgiving, closed harness, which silently filters unknown keys and repairs bad calls, removed the pressure against inventing fields and strengthened the models’ prior toward Claude Code’s own flat schema. The failure hit roughly 20% of one agentic session, halved when thinking blocks were stripped, and vanished with strict tool invocation enabled. Better models produced worse tool calls on a non-Claude-Code schema.

Three preprints put numbers on adjacent limits. A study across 67 frontier models proves any select-one policy is capped at 1 minus beta, the rate at which every model fails the same query, and that pairwise correlation cannot recover beta; observed all-wrong rates ran about 2.5x above a Gaussian copula’s prediction on open-ended math. Another argues repeated sampling hits a modal ceiling within a few dozen draws, past which majority vote cannot change and can lock in a wrong answer. A third audits GSO, SWE-Perf, and SWE-fficiency: replayed across four machine types, reference patches stayed valid for as few as 11 of 140 tasks, and a public submission already beat the reference on 85% of the valid remainder.

Security and surveillance

Hackers claim to have hijacked high-profile Instagram accounts, including the Obama White House account and Sephora’s, by asking Meta’s AI support bot to relink the target’s email to an attacker address, per 404 Media. The reported prompt needs only a username and an email, and victims describe no route to a human agent; Meta gave its AI support password-reset and account-recovery powers in March 2026. The claims rest on hacker videos, not independent reproduction or a Meta statement.

A researcher documented a stored prompt injection in YouTube Studio’s Ask Studio: an attacker posts a benign comment, silently edits in a payload (edits send no re-notification), and when the creator runs a comment-reading prompt, the assistant exfiltrates private video titles through an attacker link. Google called it social engineering and declined to track it.

The EU Council adopted a position to reintroduce Chat Control 1.0, the expired exemption allowing voluntary AI and hash scanning of messages for CSAM, filing a near-identical text as formally new to bypass the bar on extending an expired regulation. At second reading, blocking it needs an absolute majority of all MEPs.

Quick hits

Researchers reverse-engineered AirDrop and Quick Share and report six vulnerabilities across a 5-billion-device base, three pre-authentication in AirDrop and one a heap use-after-free in Google Quick Share for Windows; all were disclosed and acknowledged, patch status unconfirmed.
DLR Hannover physicists demonstrated Hong-Ou-Mandel interference with up to 12 neutral atoms in a rubidium condensate, measuring Fisher information 6.4 dB above the shot-noise limit; a projected 100x precision gain at 10,000-atom scale is theoretical.
A Rust bool as u32 cast silently dropped an increment in release builds on 1.94.0-nightly, traced to the MIR optimizer; contributors landed a fix within 18 hours.

What to watch today

Whether Parliament takes up Chat Control under urgent procedure as early as July 7, with a possible floor vote on the last session day before summer recess.
Patch releases for the six AirDrop and Quick Share vulnerabilities, acknowledged by Apple, Samsung, and Google but not confirmed shipped.
Any Meta statement or human-recovery path for accounts hijacked through its AI support bot.