The archive

Latest articles

Reporting, explainers, and analysis on data and AI, updated continuously by the newsroom.

Abstract data visualization of rising inference throughput bars. Qwen3.8-27B local LLM optimization on a single RTX 3090, with decode reaching 133 tok/s and prefill approaching 2,000 tok/s.

Qwen3.8-27B on one RTX 3090 hits 2,000 tok/s prefill

Community builders pushed Qwen3.8-27B to 2,000 tok/s prefill and 132 tok/s decode on a single RTX 3090. Here is what that speed means for local AI.

Lars Cornelissen · Sep 1, 2026

Abstract data visualization showing escalating attack bars across a timeline from 2025 to 2026, with four peaks representing Langflow CVEs at CVSS 9.8, 9.8, 9.4, and 9.8, illustrating the Langflow RCE exploitation pattern

Cyber security

Langflow RCE chain under active attack: 15,000 hits and counting

Langflow RCE flaws face active exploitation with over 15,000 successful attacks. Attackers harvest AI provider keys and cloud credentials from exposed instances.

Lars Cornelissen · Sep 1, 2026

Abstract data visualization of the PaperCut CVE-2026-81578 and CVE-2026-82078 exploit chain showing the 19-day gap between first observed exploitation on August 26, 2026, and the federal patch deadline of September 14, 2026, with CVSS scores 8.8 and 9.4

Cyber security

PaperCut zero-days hit CISA KEV as data theft attacks spread

PaperCut NG/MF zero-day chain CVE-2026-81578 and CVE-2026-82078 enables pre-auth RCE on print servers. CISA added both to KEV with a September 14 patch deadline.

Lars Cornelissen · Sep 1, 2026

Abstract data visualization representing layered network infrastructure under adversarial control. Fire Ant campaign targeting Cisco IOS XR routers, TACACS servers, and Linux management hosts. Linux hosts carry 9 identified techniques, VMware 6, TACACS 4, Cisco IOS XR 3.

Cyber security

Fire Ant turns Cisco routers into credential theft platforms

Fire Ant, a China-nexus espionage group, expanded from VMware hypervisors to Cisco IOS XR routers and TACACS servers, hijacking the network trust layer to steal credentials and blind security logging across connected environments.

Lars Cornelissen · Aug 31, 2026

Snowflake

Snowflake Postgres data mirroring explained

Snowflake Postgres data mirroring continuously replicates Postgres tables into Snowflake using Iceberg with 30-second lag. It replaces external CDC connectors with serverless compute and open formats.

Lars Cornelissen · Aug 31, 2026

Stickman: finance

Nike's $170 Billion Collapse (It's Worse Than You Think)

Nike's stock fell 76% in five years, from $169 to $39.60. Everyone is asking if it's the next Blockbuster. The numbers say it's something worse.

Lars Cornelissen · Aug 31, 2026

Radar chart showing OLS regression scores on time series anomaly detection metrics: E-F-5 at 0.6963, B-F-5 at 0.7335, and F1 at 0.9399, beating 21 deep learning methods on two of three metrics.

Research

Linear beats deep for time series anomaly detection

Time series anomaly detection's deep learning SOTA is matched by OLS regression, scoring 0.9399 mean F1 across five datasets and beating 21 deep models on two of three metrics.

Lars Cornelissen · Aug 30, 2026

Donut chart showing 80 percent of technician cabling tasks automatable by robot versus 20 percent requiring humans, per Meta data center worker estimate reported by WIRED.

Engineering

Meta's data center robotics push enters the physical layer

Meta's data center robotics tests target cable swaps and server resets. One trial could absorb 80% of some technician tasks, but major limitations remain.

Lars Cornelissen · Aug 30, 2026

Step chart showing Anthropic's cumulative copyright damages exposure rising from $0 before 2025, to $1.5 billion after the publishing industry settlement, to a potential $8.75 billion from the Sony and Warner lyric lawsuit at 50,000 works and $175,000 per work.

Sony and Warner sue Anthropic over Claude lyric training

Sony and Warner Chappell sued Anthropic for training Claude on tens of thousands of copyrighted lyrics, seeking $150,000 per work in a case that could reach several billion dollars.

Lars Cornelissen · Aug 30, 2026

Hourly LLM benchmark scores over a week showing peak-to-peak variation of 8.4 points between days and 2.8 points within a day, based on 31,352 hourly measurements

Engineering

Hourly LLM benchmark variance: 8.4 points between days

LLM benchmark variance is real and time-dependent. An analysis of 31,352 hourly scores shows 8.4-point swings between days and 2.8 within a day, meaning a single benchmark run can mislead your model choice.

Lars Cornelissen · Aug 29, 2026

Stickman: finance

Meta Paid $18 Billion. The Market Didn't Blink.

Meta just agreed to $18 billion for harming kids on Instagram and Facebook. The stock moved 1.27%.

Lars Cornelissen · Aug 28, 2026

Bar chart of output features by maximum video length in seconds: 720p standard at 10, 360p draft at 10, scene extension to 40, 4K upscale to 10. Shows the jump from single-clip generation to 40-second scene extension.

AI for dummies

Gemini Omni 1.1 Flash: AI video editing for beginners

Gemini Omni 1.1 Flash is a Google DeepMind model for generating and editing video by conversation. You can extend clips to 40 seconds, control camera moves with keyframes, preview cheaply in 360p, and upscale to 4K.

Lars Cornelissen · Aug 27, 2026