05-11-Daily AI News Daily

Daily Summary

Redis creator built a local inference engine specifically for DeepSeek, running on a 128G MacBook—Agent token bills can finally hit zero.
Meanwhile, Codex autonomously completed a security audit bounty task with zero human intervention, while ByteDance axed half its AI apps, keeping only Doubao—the money's running out.
Today's info density is off the charts; items 1, 2, and 3 are must-reads, and don't miss the one about Hengdian film studios emptying out.

⚡ Quick Navigation

• 📰 Today’s AI News - Latest updates at a glance

💡 Tip: Want to experience the latest AI models mentioned in this article (Claude 4.5, GPT, Gemini 3 Pro) right away? No account? Head to Aivora , grab an account in one minute, hassle-free support.

Today’s AI News

👀 One-Liner

Redis creator just stuffed DeepSeek into a MacBook, and Agent-era token bills might actually disappear.

🔑 3 Keywords

#LocalInferenceRevolution #AgentBurningMoney #AIStealingJobs

🔥 Top 10 Headlines

1. DeepSeek on MacBook: “Lobster Freedom” Without Spending a Dime

Heavy Agent users rack up token bills in the tens of thousands monthly. Then someone says: stop paying.

Salvatore Sanfilippo (antirez), Redis’s original creator, open-sourced ds4 on GitHub—a few thousand lines of C code for an inference engine custom-built for DeepSeek V4 Flash. Runs natively on a 128G MacBook Pro. Run code locally, run Agents locally, token consumption drops to zero.

The shock isn’t just “free.” Antirez is a top-tier internet infrastructure programmer globally. The fact he’d build an engine specifically for DeepSeek is an endorsement in itself. DeepSeek’s influence has already penetrated elite developer circles.

2. Interesting — Sam Altman Retweets: Codex Made $5 on Its Own

Someone told Codex to “make $5” and then left it alone.

Codex found an open-source security audit bounty path, submitted a PR, followed up with maintainers—zero human intervention—and the money landed. Sam Altman saw the tweet and replied with one word: “interesting.”

That “interesting” deserves unpacking. Not “amazing,” not “incredible”—it’s that quiet tone you hear when an inflection point arrives. Agents making money autonomously just went from concept to screenshot.

3. ByteDance Pulls Back Hard on AI Apps, Doubles Down on Doubao, Bets on PICO + AI Hardware

The intel’s from Jike, but the logic checks out: at 2025’s burn rate, ByteDance’s cash runway doesn’t make it past 2027.

So ByteDance’s move: keep only Doubao in the app layer, bet on PICO + AI hardware at the device layer, cut everything else. Meanwhile, several AI app companies with ARR over $100M are quietly laying off staff, and million-follower creator Dan Koe’s AI startup Eden stopped iterating due to unsustainable burn.

The real value here isn’t “ByteDance laid off again”—it’s the brutal rule it exposes: using internet logic to build AI products, chasing DAU and scale, is a dead end. AI products have no economies of scale; bigger = more losses.

4. Everyone Making Short Dramas Moved to Travel Vlogging—AI Just Emptied Hengdian

Q1 2026: Hengdian short-drama production starts dropped 75% year-over-year.

Not industry decline—AI killed the cost advantage of real-actor shorts. A decent live-action short drama costs hundreds of thousands; AI-generated versions cost pennies. Cinematographers, costume designers, gaffers, script supervisors, post-production crews—the entire supply chain is pivoting to travel vlogging and wedding photography.

This is AI job displacement in its most concrete form. Not “someday in the future”—it’s Hengdian in 2026, and the sets are empty.

5. 9x Explosion! SK Hynix Operating Margin 72%, Dethroning Nvidia as Global Leader

Nvidia’s margin: 65%. TSMC: ~54%. SK Hynix: 72%—the memory chip company just beat everyone.

SK Hynix hit 500 trillion won in 2026 revenue, market cap peaked at $807B, up 9x in a year. Employee year-end bonuses estimated at 3M RMB minimum, possibly 6.1M next year. The biggest winner in the AI compute arms race isn’t the model builders—it’s the memory sellers.

HBM (High Bandwidth Memory) is AI’s oil, and SK Hynix is the world’s largest oil field.

6. 3 Years, 5x Growth: Nvidia Pours $4.5B into Corning—175-Year-Old Glass Maker Becomes AI Infrastructure Core

AI data center fiber demand exploded 75.9% year-over-year. Supply-demand gap widened from 6% to 15%, fiber prices tripled in months.

Nvidia’s response: consecutive investments in Lumentum ($2B), Coherent ($2B), Corning ($500M)—$4.5B total—locking down the entire optical interconnect supply chain from lasers to photonic chips to fiber. Corning, a New York glass company founded in 1851, saw stock surge 316% recently, market cap hitting $160B.

AI infrastructure money is flowing to places you’d never expect.

7. Persistent Agents: Two Paths—Replay Model vs. Snapshot Recovery

Agent crashes mid-run—how do you recover? Every Agent infrastructure builder is losing sleep over this.

Trigger.dev co-founder Eric Allam proposes: traditional “replay models” (replaying operation logs) work for short tasks, but Agents can run for hours or days, context balloons, replay costs explode. His solution: two-layer architecture—context logs + Firecracker microVM-level execution snapshots, recover from crash point directly, no re-run needed.

Stateless computing ruled backend infrastructure for 30 years. Agent era is smashing that paradigm.

8. Embodied AI’s Journey: Google RT1, RT2, SayCan Author Ted Xiao Recaps Three Eras of Robot Learning

The hesitations, pivots, and breakthroughs you don’t see in papers—that’s where the real value lives.

Ted Xiao, core author of Google’s RT1, RT2, and SayCan, walks through three eras of robot learning: early brute-force era of hand-designed reward functions, data-driven imitation learning era, and today’s embodied AI era powered by large models. Each paradigm shift carries stories of “we thought we had it right, turned out it was a dead end.”

Essential reading for embodied AI builders, valuable for AI product makers too—tech route selection was never linear.

9. PyTorch Lightning: 1 to 10,000+ GPUs, Zero Code Changes

Training large models: the real pain isn’t algorithms, it’s environment setup and multi-GPU adaptation—change code, hit error, change again, error again.

PyTorch Lightning solves exactly this: same training code, unchanged, scales from single GPU to 10K-GPU clusters. Currently 31,128 GitHub stars, supports pretraining and fine-tuning at any scale. For teams without dedicated MLOps staff, this project saves massive infrastructure debugging time.

If you’ve been scared off large model training by environment config hell, take a serious look.

10. Behind the Explosion of Kaipai: What AI Tools Do Regular People Actually Need?

Luo Zhenyu started making video diaries, then publicly recommended a mobile app called Kaipai.

Small story, but the logic behind it matters: in the AI era, tool friction eventually drops to near-zero. Regular people don’t need complex workflows, just a phone. The core insight: AI tool competition ultimately converges on “who’s simpler, who’s more intuitive,” not “who has more features.”

For AI product builders, that’s a serious warning to take to heart.

📌 Worth Watching (5 Items)

[Product] Clearly: Free Open-Source Minimal Markdown Editor — Obsidian too heavy? This one supports KaTeX, Mermaid, global hotkey invoke, iOS version, Agent-friendly, lightweight with zero overhead.

[Product] HuggingFace Official CLI: Read Latest AI Papers from Command Line — One-line install, hf papers read [paper URL] reads directly, supports both arxiv and HuggingFace, research efficiency supercharger.

[Business] Samsung’s Retreat and Advance — Samsung appliances lost 940M RMB in China in 2025, first sustained losses across all categories in 30 years, formally exiting China’s consumer market; but it still commands half the ultra-premium global market—strategic pullback, not collapse.

[Other] Ditch Blind Workouts, Follow This Weekly Plan — Barely AI-related, but after scrolling through all this anxiety-inducing news, this article reminds you: your body still needs training.

😄 AI Fun Facts

🔮 AI Trend Predictions

Local Inference Becomes Standard Agent Infrastructure

• Prediction Timeline: Q3 2026
• Confidence: 72%
• Rationale: Today’s news on DeepSeek on MacBook shows elite developers already building custom wheels for local inference; once ds4-class projects mature, they’ll integrate into mainstream Agent frameworks fast. Zero token costs become the core Agent selling point, pushing more devs toward local deployment.

AI App Layer Consolidation, Independent AI App Companies Cut in Half

• Prediction Timeline: Q3 2026
• Confidence: 68%
• Rationale: Today’s news on ByteDance’s AI App Pullback signals a systemic issue: even ByteDance can’t sustain the burn rate, and $100M+ ARR AI app companies are laying off. The internet-scale-logic approach to AI products is collectively failing. Consolidation or shutdown wave incoming.

Agent Infrastructure Funding Boom

• Prediction Timeline: Q2-Q3 2026
• Confidence: 75%
• Rationale: Today’s news on Persistent Agents shows Agent long-run infrastructure problems (persistence, recovery, state management) are now clear technical bottlenecks. Trigger.dev-class projects are filling gaps. Capital typically floods in 1-2 quarters after bottlenecks get clearly defined.

Embodied AI Enters “Data Flywheel” Phase, Leaders Pull Ahead Fast

• Prediction Timeline: Q3 2026
• Confidence: 60%
• Rationale: Today’s news on Ted Xiao’s Robot Learning Recap maps embodied AI’s evolution from hand-tuned rewards to large-model-powered systems, currently at phase-three inflection. History shows teams accumulating real robot data first build insurmountable advantages.

❓ Related Questions

How to Experience DeepSeek V4 Flash Local Deployment in China?

DeepSeek V4 Flash can currently run locally on 128G MacBook Pro via the ds4 project, completely free, no API Key needed. But setup friction is high—requires self-compiling C code and downloading model weights, not friendly for regular users.

Want to experience DeepSeek or other mainstream AI models (Claude, ChatGPT, Gemini, etc.) without local setup hassle? Visit Aivora , ready-made accounts, instant delivery, hassle-free support.