# Daily Summary for 2026-03-26

## 2026-03-26 15:01:22

# AI Digest - March 26, 2026

## Industry News
- **Hark launches as new AI lab**: Brett Adcock introduces Hark, building "the most advanced, personal intelligence" with backing from major investors. [link](https://x.com/adcock_brett/status/2037180562184585544)
- **Modal's sandbox revenue milestone**: Modal's sandbox revenue now equals the company's total revenue from 9 months ago, showing explosive growth in AI infrastructure. [link](https://x.com/bernhardsson/status/2037181965225660913)
- **Runway Characters gets real-time inference**: Partnership with Modal enables global avatar spin-up in seconds, dramatically faster than previous implementation. [link](https://x.com/kamilsindi/status/2037182523239752111)

## Tips & Techniques
- **Ask agents if they're missing context**: Simple prompt "Are you missing any context?" significantly improves agent performance when stuck. [link](https://x.com/TTrimoreau/status/2037182784456769933)
- **Context window vs tool count**: Adding more MCP tools doesn't make agents better—it makes them confused. The context window fills up regardless of tool quality. [link](https://x.com/byteHumi/status/2037179941478170920)
- **Real-time video captioning with Qwen3.5**: Run 0.8B model on Mac Studio M2 Ultra at <1s per frame for 269 frames from a 3m49s video. [link](https://x.com/HuggingModels/status/2037179453018571090)

## New Tools & Releases
- **Cohere Transcribe (Apache 2.0)**: #1 on Open ASR Leaderboard, beats GPT-4o and Whisper v3 with 5.42% WER across 14 languages including Arabic, Chinese, Japanese. [link](https://x.com/aidangomez/status/2037172942803701838)
- **TRIBE v2 from Meta**: Foundation model predicts human brain responses to sight, sound, and language—a digital twin of neural activity with zero-shot generalization. [link](https://x.com/AIatMeta/status/2037178782495498668)
- **Lica launches with editing layers**: AI image generation that breaks outputs into editable layers instead of flat PNGs, ending "prompt doom loops." Backed by Accel and Replit CEO. [link](https://x.com/pritopian/status/2037167928114896987)
- **OpenEvidence Coding Intelligence**: Auto-generates CPT, ICD-10, and modifiers from physician notes, handling medical coding alongside clinical AI. [link](https://x.com/EvidenceOpen/status/2037178697812164749)
- **Expect for browser testing**: Let agents test your code in real browsers—run Claude Code/Codex to QA, watch video replay of failures. [link](https://x.com/aidenybai/status/2037182150601302416)

## Research & Papers
- **On-Policy Self-Distillation breaks under uncertainty**: Paper shows self-distillation improves acceptance rates but collapses when navigating uncertain environments—internalizes feedback but loses adaptability. [link](https://x.com/ShashwatGoel7/status/2037172974198153728)
- **Ego2Web exposes web agent limitations**: Google DeepMind shows today's web agents fail badly on real-world video demonstrations, highlighting gap between benchmarks and reality. [link](https://x.com/rohanpaul_ai/status/2037182472283152825)
- **LEAP: Local ECT-Based Positional Encodings**: Graph neural network approach that outperforms message aggregation with learnable positional encodings. Accepted at ICLR 2026. [link](https://x.com/Pseudomanifold/status/2037171745140920750)

---
*Curated from 478 tweets across 5 lists*

---

## 2026-03-26 15:01:23

## Emerging Trends

✨ **Cloudflare Dynamic Sandboxes** (32 mentions) - NEW
Cloudflare announced lightweight, fast, truly ephemeral sandboxes with virtual filesystems for agent execution. Multiple developers highlighted this as a paradigm shift for containers/sandboxes, making them resource-light and ridiculously fast, positioning Cloudflare as the default compute provider for agents.

✨ **LiteLLM Security Breach** (28 mentions) - NEW
The LiteLLM package on PyPI was compromised with versions 1.82.7 and 1.82.8 containing malicious code that steals credentials including SSH keys. The package has been quarantined by PyPI, with widespread warnings for users to rotate tokens if they installed the compromised versions.

🔥 **Figma MCP Integration** (45 mentions) - RISING
Figma released a major MCP update enabling AI agents like Claude Code to directly design on Figma canvases using the use_figma tool, with full design system context. This reverses the previous code-generation focus to allow agents to create and edit designs, components, and styles directly in Figma.

🔥 **RunClaw Launch** (38 mentions) - RISING
RunClaw launched as a competitor to OpenClaw, offering a $1 no-setup alternative compared to OpenClaw's $700 setup cost. It provides agents via DMs with enhanced capabilities including website building, video generation, and slide decks, while claiming better security with zero CVEs.

🔥 **Mirror Mirror AI Marketplace Launch** (22 mentions) - RISING
Mirror Mirror AI launched a marketplace where fashion models license their likeness for AI-generated brand imagery. The platform provides commercially licensed, model-approved AI fashion photography featuring real agency-signed models, with an open call for independent models globally.