The archive

Latest articles

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

A donut chart of AI agent identity practices in enterprises: 34% give every agent a scoped identity, 66% share credentials across agents. Source: VentureBeat Research survey of 107 enterprises.

Agent security gap: 54% of enterprises already hit

AI agent security incidents are now widespread. 54% of enterprises have had a confirmed agent incident, yet most still let agents share credentials instead of scoped identities.

Lars Cornelissen · Jul 17, 2026

Funnel chart showing the enterprise AI agent evaluation gap across 157 organizations. 118 are shipping to production, 79 shipped an agent that passed internal evals but failed in production, and only 8 fully trust automated evaluation.

Enterprise AI agent evaluation gap: half ship broken agents

The AI agent evaluation gap shows 50% of enterprises shipped an agent that passed internal evals then failed in production. Only 5% fully trust automated evaluation. The gap is structural misalignment, not missing coverage.

Lars Cornelissen · Jul 17, 2026

Heatmap showing AI compute cost measurement maturity across 107 enterprises. Only 8 percent have mature tracking, 52 percent have partial tracking, and 40 percent have little or none. Source: VentureBeat survey.

AI compute cost gap: 92% of enterprises fly blind on spend

The AI compute cost gap is widening: 92 percent of enterprises lack mature cost tracking for AI infrastructure even as 45 percent plan to switch or add providers within a year, risking runaway spend.

Lars Cornelissen · Jul 17, 2026

Open-weight model comparison showing Inkling at 975B total parameters and 41B active, Inkling-Small at 276B total and 12B active, and Hy3 295B at 295B total and 21B active

Inkling 975B: Thinking Machines releases open-weights model

Inkling is Thinking Machines Lab's first open-weights model. At 975B parameters with 41B active and Apache 2.0 licensing, it targets fine-tuning, not frontier benchmarks.

Lars Cornelissen · Jul 17, 2026

Comparison chart of open-weight model sizes in trillions of parameters for Kimi K3 and peers. Kimi K2.6 at 1.0 trillion, DeepSeek v4 Pro at 1.6 trillion, Kimi K3 at 2.8 trillion parameters.

Kimi K3 explained: what 2.8 trillion parameters means

Kimi K3 is a 2.8 trillion parameter open-weight model from Moonshot AI that matches top US closed models on key benchmarks. Weights arrive July 27.

Lars Cornelissen · Jul 17, 2026

Bar chart comparing memory footprint of Qwen 3.6 27B model versions. Full precision 16-bit at 54 GB, conventional 4-bit at 18 GB, Ternary Bonsai at 5.9 GB, and 1-bit Bonsai at 3.9 GB, with the phone-fit threshold marked at 6 GB.

Bonsai 27B puts a big AI model on your phone

1-bit quantization is a compression trick that shrinks AI models by storing each parameter as one bit. Bonsai 27B uses it to fit a 27 billion parameter model in 3.9 GB, running on an iPhone.

Lars Cornelissen · Jul 15, 2026

Bar chart comparing CVSS scores for two SonicWall SMA 1000 zero-day vulnerabilities: CVE-2026-15409 at 10.0 critical SSRF and CVE-2026-15410 at 7.2 high code injection, both actively exploited in the wild and added to CISA KEV on July 14 2026

Two SonicWall SMA 1000 zero-days under active attack

Two SonicWall SMA 1000 zero-day vulnerabilities including a CVSS 10.0 SSRF are under active exploitation. CISA KEV-listed with a July 17 patch deadline.

Lars Cornelissen · Jul 15, 2026

Slope chart showing interpretability visibility before and after the J-space discovery. Five behavior categories improve: task progress from 20 to 70, recognition from 10 to 60, commentary from 5 to 80, bias detection from 15 to 55, and cheating signals from 0 to 70. Illustrative values.

Anthropic J-space reveals LLM thoughts, but not fully

Anthropic's J-space exposes hidden words shaping LLM reasoning. Mechanistic interpretability advances, but misuse detection stays unproven.

Lars Cornelissen · Jul 14, 2026

Format Sensitivity Index showing benchmark score variance of up to 15 percentage points across JSON, XML, Markdown, and plain text prompt wrappers for four illustrative LLM models

Format Sensitivity Index exposes LLM benchmark gaps

The Format Sensitivity Index measures how LLM benchmark scores shift when prompt wrappers and schema constraints change. The metric exposes a blind spot in model evaluation that developers ignore at their peril.

Lars Cornelissen · Jul 14, 2026

Gantt chart showing data center moratorium policy timeline from January 2026 to July 2027. Maine's moratorium was vetoed in April 2026. New York's executive order, signed July 14 2026, blocks data centers over 50 MW for up to 12 months. New York's legislative bill, with a 20 MW threshold, awaits the governor's signature.

New York's data center moratorium redraws the AI map

New York enacted the first statewide data center moratorium, blocking new permits for facilities over 50 MW for up to a year. What changes for builders.

Lars Cornelissen · Jul 14, 2026

Apple Silicon unified memory ceiling growth from 128GB in 2022 to 192GB in 2023, then an estimated 1.5TB for the M7 Ultra in 2027, showing a near eightfold jump in max RAM for Apple's AI chips.

Apple's dead car project built the Neural Engine

Apple's Neural Engine, born from its failed car project, became the backbone of on-device AI. The M7 Ultra with 1.5TB of RAM could put Apple in the server inference market by 2027.

Lars Cornelissen · Jul 13, 2026

Cloudflare AI crawler traffic split showing 50 percent re-fetching unchanged pages, with Agent and Training categories blocked by default on ad-supported pages from September 15 2026

Cloudflare agent crawler blocks change the access deal

Cloudflare AI agent crawler rules block ad-supported pages by default from September 15. Agent builders face degraded coverage and need negotiated access, not user-agent tricks.

Lars Cornelissen · Jul 13, 2026