Daily News · 5 min read

AI News: April 20, 2026

Listen

1. RealChart2Code Benchmark: Top Models Lose Half Their Score When Charts Get Real

RealChart2Code. A new benchmark of 2,800 test cases drawn from 1,036 Kaggle datasets (~860M rows) evaluates models on three chart-to-code tasks: replication from images alone, reproduction with raw data attached, and refinement of broken code through dialogue. Claude 4.5 Opus (8.2) and Gemini 3 Pro Preview (8.1) lead, while open-weight Qwen3-VL-235B and InternVL-3.5-241B collapse to 3.6 and 3.4. The headline finding is a “complexity gap” — Gemini 3 Pro Preview scores 96% on the older ChartMimic benchmark but only ~50% here — a useful warning for teams shipping chart-comprehension features that current saturated benchmarks dramatically overstate real-world capability. Code and data are on GitHub and Hugging Face. Source

2. NYT and Purdue Identify 300+ AI-Generated Pro-Trump Influencer Accounts Ahead of Midterms

New York Times / Purdue GRAIL / Alethea. The NYT identified at least 304 AI-powered pro-Trump accounts on TikTok since January, while Purdue’s GRAIL lab and security firm Alethea found additional clusters across Instagram, Facebook, and YouTube, with some accounts hitting 35K+ followers and individual posts reaching 500K views. Production cost is reportedly $1–$3 per post, repeating language patterns and character archetypes give the campaigns away, and no comparable left-wing networks were found. TikTok is treating them as spam rather than coordinated influence; Japan’s recent election saw similar campaigns where over half of surveyed respondents believed the synthetic content. The episode is a real-world stress test of platform detection ahead of November and a signal that synthetic-media policy needs to advance faster than detection models. Source

3. German Court: AI Comic of Copyrighted Photo Isn’t Infringement Without Replicating Creative Elements

OLG Hamm. A German Higher Regional Court (case I-20 W 2/26, April 2) ruled that converting a copyrighted underwater dog photograph into a comic-style image with AI did not infringe the original because the AI version did not adopt the protected creative choices — framing, perspective, lighting, sharpness — only the subject matter, which is not protected. The judges leaned on a prior CJEU ruling about specific creative-element adoption, and added that AI outputs themselves only get copyright protection when humans make distinctly creative choices (generic prompts don’t count). The decision aligns with positions from other German courts and the US Copyright Office, suggesting an emerging international consensus narrowing both the infringement risk and the protection scope of AI generations. Source

4. VisionClaw Study: Always-On AI Agents on Ray-Ban Meta Glasses Cut Task Time 13–37%

University of Colorado / GIST / Google. Researchers paired Ray-Ban Meta glasses (displayless) streaming audio and frames over a custom phone app with Google’s Gemini Live and the OpenClaw agentic framework, then ran a study across note-taking from physical documents, email composition, product research, and device control. The system completed tasks 13–37% faster with users reporting 7–46% lower perceived effort, though small-object recognition (e.g., receipts) dropped to 58% accuracy. A field study clustered usage into six categories — information retrieval, shopping, content saving, communication, memory support, device control — and the authors argue ambient agents shift the design center from reactive voice commands to opportunistic, context-driven assistance. Concrete data point for teams designing agent UX on wearables. Source

5. Palantir Publishes 22-Point Manifesto Tying AI Weapons to “Western” Defense

Palantir. A 22-point summary of CEO Alex Karp’s “The Technological Republic” went up on Sunday, arguing Silicon Valley owes “a moral debt to the country that made its rise possible” and dismissing “the shallow temptation of a vacant and hollow pluralism.” On AI specifically, the post contends “the question is not whether A.I. weapons will be built; it is who will build them and for what purpose,” framing the company as the answer in a new era of “deterrence built on A.I.” It is less a philosophy thesis than a positioning statement aligned with Palantir’s defense, intelligence, ICE, and law-enforcement contracts — useful context for engineers deciding where their work sits on the dual-use spectrum. Source

6. Uber Commits $10B+ to Owning Robotaxi Fleets Across Wayve, WeRide, Lucid, Nuro, Rivian

Uber. TechCrunch Mobility frames Uber’s pivot as an “assetmaxxing era”: $2.5B in direct investments plus $7.5B earmarked for robotaxi purchases over the coming years, abandoning the asset-light model in favor of owning physical AV fleets built by partners including Wayve, WeRide, Lucid, Nuro, and Rivian. Travis Kalanick has called shutting Uber ATG in 2020 “a mistake,” and the new model effectively rebuilds that bet through procurement rather than R&D. In adjacent funding, Slate Auto raised a $650M Series C for sub-$30K EV trucks, Glydways pulled in $170M for autonomous pods, and Loop took $95M for supply-chain prediction. The strategic question for the AV stack: with Uber as the buyer-of-record, do AV developers compete on tech or on per-mile economics? Source

7. Elad Gil’s “12-Month Window” Argues Most AI Startups Should Plan to Exit Inside a Year

Elad Gil / TechCrunch. Investor Elad Gil contends most AI companies hit peak valuation within a roughly 12-month window before foundation-model expansion erodes their defensibility, and recommends standing twice-yearly board agendas explicitly devoted to exit timing — removing emotional bias from the call. He cites Lotus, AOL, and Broadcast.com as historic exemplars of selling at the top, and quotes Deel CEO Alex Bouaziz’s tweet — “if you do come for us, call me first 🙏” — as the prevailing mood among founders watching GPT-class models walk into adjacent verticals. For practitioners and operators, the framing is a useful counterweight to the default growth narrative: differentiation drift is now a concrete planning input, not just a strategy-deck risk. Source