↑ 1 new · ↻ 2 carryover from yesterday
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 #
1 story · 🟢 +1.0 sentiment
- 6 🌱 🟢 ▤×1 🏷️ opensource, inference DeepSeek's permanent 75% discount on V4-Pro reframes open-weights economics. Sustained sub-quarter pricing turns "open weights" from a research-credit play into a structural pricing wedge. Sources: Bloomberg
📈 Prediction Markets #
5 markets tracked · big swings on lab-vs-lab June ranking
- "Will Google Gemini score ≥45% on FrontierMath?" Yes 49.0% (▲17.0pp) · $50K vol · Polymarket
- "Will a Claude Mythos model be released by June 30, 2026?" Yes 21.0% (▲10.5pp) · $113K vol · Polymarket
- "OpenAI IPO closing market cap above $1T?" Yes 83.0% (▲7.0pp) · $1.04M vol · Polymarket
- "Will Anthropic have the best AI model at the end of June 2026?" Yes 77.0% (▲5.6pp) · $1.02M vol · Polymarket
- "Will Google have the best AI model at the end of June 2026?" Yes 19.5% (▼5.0pp) · $670K vol · Polymarket
💬 Discourse #
Bluesky
- @wired.com on Bluesky — "The nine-member panel took only two hours to return a verdict in favor of OpenAI on Monday, which the judge quickly adopted." (103♥/22↻) — note on how fast the OpenAI jury moved.
- @sarahgooding.bsky.social on Bluesky — "AI is now driving both the production and consumption of open source software. AI-generated music ends in human ears..." (13♥/5↻) — clean framing of the feedback loop closing around OSS.