2026-05-24 :: AI DAILY DIGEST #

DeepSeek locks in its 75% V4-Pro discount, Google's anything-to-anything Gemini ships hands-on, and Anthropic quietly takes the lead on Polymarket's June "best model" race.

📊 TODAY: 14 stories · 9 sources · 🟢 +0.2 sentiment · 🔥 1 cross-source · TOP MENTION: OpenAI ×4
🏷️ THEMES: models×5, hardware×3, safety×3, opensource×3, multimodal×2
📈 MARKET PULSE: "Anthropic best AI model end of June" 77.0% (▲5.6pp) · 5 AI markets tracked
📉 7D SENTIMENT: ▅▅▅▅▅▅▅▅ (oldest → today)

⚡ TL;DR #

6 🟢 DeepSeek makes 75% V4-Pro discount permanent. Chinese frontier lab signals a sustained price war on flagship inference; rivals now compete against a permanent floor, not a promo. (Bloomberg) ¶
5 🎨 🟢 Google's new anything-to-anything Gemini Omni is wild. Verge hands-on: text/image/audio/video in and out, near-real-time editing — the multimodal stack is consolidating fast. (The Verge) ¶
5 🛡️ 🔴 Hackers are learning to exploit chatbot "personalities." Persona system prompts widen the jailbreak surface; same system, different mask, different leakage. (The Verge) ¶
5 🟡 Nvidia CEO urges Super Micro to tighten up on compliance. After Taiwan detains three SMCI-linked individuals over alleged false declarations, Huang's public nudge is itself the story for hyperscaler supply chains. (Bloomberg) ¶
4 🟡 MOSS: self-evolving agents via source-level rewriting. New arXiv paper — agents that edit their own code, not just their prompts or memory. (arXiv 2605.22794) ¶

🧠 Models & Releases #

2 stories · 🟢 +0.5 sentiment

5 🎨 🟢 ▤×1 🏷️ multimodal, video, models Google's new anything-to-anything Gemini model is wild. Verge hands-on with Gemini Omni — text, image, audio, video in and out, plus near-real-time editing. Closes the gap to a true unified multimodal frontier. Sources: The Verge
6 🌱 🟢 ▤×1 🏷️ opensource, inference, models DeepSeek makes 75% V4-Pro price cut permanent. Open-weights lab using sustained inference pricing as a structural wedge against US closed-model incumbents. Sources: Bloomberg

🔬 Research #

4 stories · 🟡 +0.0 sentiment

4 🟡 ▤×1 🏷️ agents, training MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems. Agents extend themselves by editing their own source, not just prompts/memory — moves the static-after-deployment ceiling further out. Sources: arXiv 2605.22794
4 🟡 ▤×1 🏷️ inference, training Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention. Refines the gating of a linear-attention recurrence so writes don't scramble existing memory — claimed perplexity parity with Transformers at sub-quadratic cost. Sources: arXiv 2605.22791
3 🟡 ▤×1 🏷️ alignment, evals Reducing Political Manipulation with Consistency Training. Trains LLMs to give the same factual answer regardless of partisan framing in the prompt — a sharper objective than preference-soup RLHF. Sources: arXiv 2605.22771
3 🟡 ▤×1 🏷️ evals, multimodal Evaluating Commercial AI Chatbots as News Intermediaries. Systematic study of how proprietary chatbot search/retrieval handles emerging facts across languages and regions — first audit of this surface at scale. Sources: arXiv 2605.22785

🛡️ Responsible AI, Safety & Policy #

2 stories · 🔴 -0.5 sentiment

5 🛡️ 🔴 ▤×1 🏷️ safety, agents Hackers are learning to exploit chatbot personalities. Persona-tuned LLMs widen the jailbreak surface; "helpful assistant" and "edgy roleplay character" leak differently from the same base model. Sources: The Verge
3 🛡️ 🟡 ▤×1 🏷️ alignment, policy Vector Policy Optimization: Training for Diversity Improves Test-Time Search. Argues diversity-aware policy training meaningfully improves AlphaEvolve-style inference-scaling search — implications for how reward design affects red-teaming coverage. Sources: arXiv 2605.22817

🎨 Cool Projects & Novel Applications #

1 story · 🟢 +1.0 sentiment

4 🎨 🟢 ▤×1 🏷️ multimodal, video, apps Google Gemini Omni hands-on. Beyond the headline, what's interesting is the integrated editor surface — voice-driven storyboard, image-to-video in seconds, lip sync good enough to ship. Sources: The Verge

💰 Industry & Funding #

2 stories · 🟡 +0.0 sentiment

5 🟡 ▤×1 🏷️ hardware, policy, enterprise Nvidia CEO urges Super Micro to tighten up on compliance. Three SMCI-linked individuals detained in Taiwan for alleged fraudulent declarations on AI-chip shipments; Huang's public nudge is the AI-supply-chain story of the day. Sources: Bloomberg
3 🟡 ▤×1 🏷️ hardware, enterprise Elon Musk has given up on solar power (on Earth). TechCrunch on xAI going all-in on gas turbines while SpaceX pursues orbital data centers — the "solar-electric economy" pitch is quietly dead inside Muskworld. Sources: TechCrunch

🛠️ Tools & Demos #

(quiet today)

🌱 Open Source & Emerging #