# Daily Summary for 2026-04-07

## 2026-04-07 15:01:00

# AI Digest - April 7, 2026

## Industry News
- **Anthropic hits $30B ARR in 3 months**: Grew from $9B to $30B ARR with 1,000+ businesses spending $1M+ each, driven primarily by Claude Code adoption for software development. Meta alone represents one-third of revenue at 60T tokens/month. [link](https://x.com/ben_j_todd/status/2041526587409989806)
- **OpenAI partners with Razorpay for payments**: Codex now integrates Razorpay natively, letting developers embed UPI/cards/wallets via simple prompts without complex integration work. [link](https://x.com/ponnappa/status/2041530413366292600)
- **Intel joins Terafab for chip manufacturing**: Partnership with SpaceX, xAI, and Tesla to combine logic, memory, and advanced packaging in one factory, targeting 1 TW/year of compute capacity. [link](https://x.com/Jeffmcleod/status/2041528997805507048)
- **OpenAI acquires TBPN podcast**: Small teams with simple cap tables getting acquihired for 9 figures is now the dominant strategy for builders, demonstrating fastest consumer growth without referral programs. [link](https://x.com/acwints/status/2041530263633805426)

## Tips & Techniques
- **Agent harness failure causes 24-hour data loss**: Dev accidentally ran git checkout from repo root during agent session, reverting all tracked files. Version control discipline and testing remain critical as agents gain more authority. [link](https://x.com/koylanai/status/2041526628468019621)
- **40-rule behavioral contract fixes Claude memory**: Instead of project context, use CLAUDE.md with 40 behavioral rules based on past failures. Compliance degrades in long conversations, so put highest-impact rules first. [link](https://x.com/jasonxsigmon/status/2041530171857965139)
- **Repeatable workflows via /commands**: Turn 20-30 minute tasks into single commands by saving step-by-step markdown files to .claude/skills/your-task/SKILL.md, then invoke with /your-task. [link](https://x.com/jasonxsigmon/status/2041530171857965139)
- **Local knowledge base unlocks years of notes**: Set up MCP server to index markdown files and let Claude search by keyword during conversations, turning dead files into active context. [link](https://x.com/jasonxsigmon/status/2041530171857965139)

## New Tools & Releases
- **longrun-prd-runner for 24+ hour agent runs**: Open source tool for PRD → manifest → long-running coding agent execution with sequential tasks, validation, review escalation, auto-repair, human gates, and smart model routing. [link](https://x.com/kyunbit/status/2041530958764273777)
- **Taste: Inject design taste into coding agents**: MacOS app that teaches Codex, Claude Code, and Cursor your design preferences via portfolio uploads, creating a "for hire" marketplace for design engineers. Private beta coming this week. [link](https://x.com/alexkehr/status/2041181515229680031)
- **Hex CLI for agent-driven BI**: Command-line tool lets agents propagate dbt updates, handle BI migrations, and manipulate Hex projects programmatically. [link](https://x.com/barrald/status/2041529657443659907)
- **Artifact Omni: Agentic accounting orchestration**: Connects across 1,000+ platforms, lets accountants describe workflows in natural language, then templatizes and deploys with client-specific logic. Captures "wet code" knowledge as seniors retire. [link](https://x.com/ariel_harmoko/status/2041520347380584813)

## 2026-04-07 15:01:01

## Research & Papers
- **ARC-AGI-3 resets AGI benchmark**: New version broke every AI model with 135 hand-crafted game worlds requiring zero-shot reasoning. Humans score 100%, frontier models all under 1%. [link](https://x.com/pvergadia/status/2041531444032405746)
- **Vero: Strongest open RL recipe for VLMs**: SOTA 8B visual language model averaging +4.4 points across 30 benchmarks, beats prior RL datasets on charts, spatial reasoning, and open-ended tasks. [link](https://x.com/GabrielSarch/status/2041528435919511877)
- **OccAny: Generalized 3D occupancy without poses**: One model works with any sensor rig in any domain without camera poses or intrinsics, using segmentation forcing from SAM + novel-view rendering. [link](https://x.com/AnhQuanCAO/status/2041510699285856739)
- **WorldFlow3D: Unbounded 3D world generation**: Reformulates 3D generation as flowing through sequentially finer distributions, cutting training time in half vs existing approaches while enabling spatially unbounded generation. [link](https://x.com/_FelixHeide_/status/2041526553339675105)

---
*Curated from 583 tweets across AI engineering communities*

---

## Emerging Trends

✨ **Cognis AI Memory System** (45 mentions) - NEW
Milla Jovovich co-developed Cognis, an open-source AI memory system achieving highest benchmark scores (92.4% on LongMemEval, #1 on LoCoMo). System focuses on contradiction resolution and temporal reasoning rather than just vector search.

🔥 **Anthropic Claude Harness Restrictions** (124 mentions) - RISING
Anthropic restricting third-party harness access to Claude models, causing community backlash. Users finding workarounds and discussing migration to alternatives like Hermes and local models. Debate over API restrictions and product strategy.

🔥 **Karpathy's LLM Wiki** (287 mentions) - RISING
Andrej Karpathy's viral post about building personal LLM-maintained wikis using markdown files, Obsidian, and agent workflows. System uses raw sources compiled into wikis with ingest/query/lint operations. Spawned discussion of "idea files" over traditional code sharing.

🔥 **Morphic Workflows Launch** (68 mentions) - RISING
Morphic launching 72+ pre-built creative workflows for AI video/content generation covering storyboarding, UGC ads, color grading, and filmmaking. No-code approach with asset selection replacing manual prompting.

📊 **Gemma 4 Release** (156 mentions) - CONTINUING
Google's Gemma 4 model release generating significant discussion, with comparisons to Qwen 3.5. Available in multiple sizes (26B, 31B), integrated into Android Studio and various platforms. Mixed performance reports on benchmarks.