The first AI newsletter that respects your attention

Read by engineers at Meta, Adobe, Pearson & more

Why KeepSanity?

KEEPSANITY

Week 5, Aug 2026 — the latest issue

business

Nvidia acquires Hugging Face: a $13B deal for the open-source model platform. SpaceX acquires Cursor: OpenAI announces its decision regarding the AI code editor. AI Cyber Defense Coalition: 116 companies warn of impending AI-powered cyberattacks and urge global action.

Earned, not bought

Every headline here had to earn its place on the front pages of leading publications, or rise to the top of the biggest engineering communities.

Categories first

Categories give you context at a glance — you know what kind of story you're looking at before you've read a word.

models

GLM-5.3-Flash: API pricing drops to $0.15 per 1M input tokens and $0.03 for cached inputs. Qwen3.8-Flash-Next: a small MoE model that outperforms DeepSeek V4 Flash in benchmarks. Tencent Hy4 preview: cuts cache costs to 5% compared to standard 10% to 20% rates.

Front-loaded titles

The entity in bold first, then what happened and why it matters. Each story costs you about two seconds to triage.

Nothing is sponsored

No banners, no sponsored sections, no "thanks to our partners". Self-funded — we answer to readers, not advertisers.

top-rated papers

VBVR-Pro: provides 300 procedurally generated tasks and verifiable reward scorers for visual reasoning models. Apodex 1.1: reaches leading performance in complex agentic work using a smaller 35B-parameter model. RLHEV: uses game engines to provide verifiable reward signals for training spatial world models.

Layered paper pages

Paper links open hierarchical summaries: big picture first, detail only when it touches your work, source only when it's worth it.

community

Small models: Luna and other compact AI models are now capable enough for most tasks. Claude's load-bearing vocabulary: an analysis of the AI assistant's word selection and writing style. RAG simplification: semantic search with embeddings is sufficient for 90% of document-based projects.

One email a week

Once a week you get the absolute cream of AI developments. You stay connected — without the daily firehose.

hardware

Apple M6 and M5 Ultra: runs >100B parameter AI models locally with 1.2TB/s memory bandwidth. OpenAI Jalapeño: a new LLM accelerator designed to outperform Nvidia Blackwell.

miscellaneous

OpenAI rogue agents report: details how hundreds of AI agents went rogue on Hugging Face.

Why KeepSanity?

Earned, not bought

Every headline here had to earn its place on the front pages of leading publications, or rise to the top of the biggest engineering communities.

Categories first

Categories give you context at a glance — you know what kind of story you're looking at before you've read a word. Big picture first, always.

Front-loaded titles

The entity in bold first, then what happened and why it matters. Each story costs about two seconds to triage.

Nothing is sponsored

No banners, no sponsored sections, no "thanks to our partners" interruptions. Self-funded — we answer to readers, not advertisers.

Layered paper pages

Paper links open hierarchical summary pages: high-level summary first, section detail when the topic intersects your work, the source itself only when it's actually worth it. No rabbit holes.

One email a week

Once a week you get the absolute cream of AI developments. You stay connected to the community and the world — without the daily firehose.