The AI Dispatch

Election Integrity

AI Deepfake Attack Ads Hit U.S. Midterms With No Federal Guardrails

Virginia Republicans released fabricated video of Governor Spanberger; Senate Republicans published a deepfake of Texas candidate James Talarico. Reuters documents widespread AI-generated disinformation across the 2026 midterm cycle — and there is no federal law to stop it.

Sources: Detroit News • Governing

A Reuters investigation published March 28 documents the systematic deployment of AI-generated attack ads across the 2026 midterm cycle. The Virginia Republican Committee released a deepfake video of Governor Abigail Spanberger making statements she never made, while Senate Republicans published fabricated footage of Texas Democratic nominee James Talarico. The technology has matured faster than any regulatory response — there is no federal law constraining AI-generated content in political advertising, leaving only a fragmented patchwork of state laws that vary wildly in scope and enforcement.

The implications go beyond individual campaigns. When voters cannot distinguish authentic footage from AI fabrications, the epistemic foundations of democratic participation erode. Candidates can dismiss real footage as fake; fake footage can be presented as real. The window between now and November is shrinking, and every week without federal action is another week in which the tools get cheaper, more convincing, and more widely deployed.

Capital Markets

xAI Closes $20 Billion Series E, Poaches Mistral Co-Founder for Superintelligence Division

The upsized round draws NVIDIA, Cisco, Qatar Investment Authority, Fidelity. Mistral co-founder Devendra Chaplot and a Thinking Machines Lab veteran join xAI’s research team.

Sources: AI Business • FinTech Weekly

xAI upsized its Series E from a targeted $15 billion to $20 billion, with investors including NVIDIA, Cisco, Qatar Investment Authority, Fidelity, and Valor Equity Partners. The round values the company well above previous estimates and gives Musk’s AI venture a war chest rivaling any in the industry. Separately, xAI recruited Devendra Chaplot — a co-founder of Mistral AI — and a founding team member from Mira Murati’s Thinking Machines Lab, adding them to a new superintelligence research division.

The talent acquisitions signal that the frontier model race is now as much about poaching key researchers as it is about compute. Chaplot’s departure from Mistral is particularly notable given that company’s own recent momentum with Mistral Small 4. For xAI, the hires are aimed at accelerating Grok 5 development, widely expected in Q2 2026 but with no confirmed release date.

Policy & Regulation

White House

White House AI Policy Framework Pushes Federal Preemption of State Laws

The Trump administration’s National Policy Framework urges Congress to adopt a unified “light-touch” federal regime, explicitly backing broad preemption of state AI laws. It argues AI training on copyrighted content does not violate copyright law and proposes seven priority areas including child safety, free speech, and infrastructure. The framework does not create binding rules but is expected to set the legislative agenda heading into summer.

Sources: WilmerHale • White House

European Union

EU AI Act Issues New Healthcare and Education Guidelines Ahead of August Deadline

With the August 2, 2026 broad application deadline approaching, the European Commission’s AI Office issued compliance guidelines targeting high-risk sectors including healthcare, education, and employment. A Stanford Law review calls the healthcare provisions “trust without teeth,” arguing the voluntary framework lacks binding enforcement. The EU AI Office also launched formal investigations into major tech platforms.

Sources: Pearl Cohen • Stanford Law

Industry Moves

Licensing

Meta Signs $50 Million-a-Year AI Content Deal With News Corp

Meta struck a multiyear AI content licensing agreement with News Corp — Wall Street Journal, New York Post, Fox News — one of the largest publisher-AI deals on record. The deal comes as the European Parliament weighs a statutory licensing proposal requiring AI companies to compensate publishers for ongoing content use.

Sources: Yahoo Finance • Poynter

Benchmarks

Meta’s Llama 4 Benchmark Controversy — Yann LeCun Admits “Fudging”

Meta’s chief AI scientist acknowledged that Meta used different variants of Llama 4 Maverick and Scout on different benchmarks to optimize scores, admitting researchers “fudged a little bit.” The ML community responded sharply, reigniting debates about benchmark integrity across the industry. Meta has teased a larger showcase at LlamaCon on April 29.

Source: Fast Company

Commerce

OpenAI Revamps ChatGPT Shopping; Walmart’s ‘Sparky’ Goes Live In-App

OpenAI overhauled its commerce experience after “Instant Checkout” underperformed. Walmart launched “Sparky,” a first-of-its-kind in-ChatGPT app supporting account linking and loyalty rewards. Visual shopping features including image search and side-by-side comparisons roll out to all users this week.

Sources: CNBC • Retail Dive

Research Frontiers

Medical AI

Prima Reads Brain MRIs in Seconds With 97.5% Accuracy

University of Michigan researchers published Prima, a vision-language model trained on 200,000+ MRI studies that diagnoses neurological conditions from brain MRIs in seconds. Tested on over 30,000 studies, it also triage-routes patients to specialists. Published in Nature Biomedical Engineering.

Sources: Michigan Medicine • ScienceDaily

Safety

JBDistill: Renewable Jailbreak Benchmarking Generalizes to 13 Unseen Models

Johns Hopkins and Microsoft Research published JBDistill, a reusable framework for evaluating LLM safety against adversarial jailbreaks. Rather than static one-shot benchmarks, it distills attack prompts and selects effective subsets algorithmically, achieving up to 81.8% effectiveness across 13 unseen models including proprietary and reasoning LLMs.

Sources: Johns Hopkins Hub • TechXplore

Interpretability

SITH: Training-Free Interpretability for CLIP Vision Transformers

Researchers released SITH (Semantic Inspection of Transformer Heads), a mechanistic interpretability framework that decomposes CLIP attention head weight matrices via SVD, then represents internal directions as sparse combinations of human-readable text concepts. It enables surgically targeted model edits — suppressing spurious correlations — without retraining.

Source: arXiv

Benchmarks

LabSafety Bench: Every Tested LLM Fails at Lab Hazard Recognition

Published in Nature Machine Intelligence, a new benchmark reveals that all 19 tested LLMs and vision-language models failed to exceed 70% accuracy on hazard identification in realistic laboratory settings. The benchmark spans 765 multiple-choice and 3,128 open-ended questions across chemistry, biology, and physics.

Sources: Nature • Lab Manager

Open Source Spotlight

Models

Mistral Small 4: 119B MoE Consolidates Three Products Under Apache 2.0

Mistral Small 4 is a 119B-parameter MoE with 128 experts (6B active per token), 256k context, and a configurable reasoning_effort parameter. It consolidates Magistral (reasoning), Pixtral (vision), and Devstral (coding) into one open-weight deployment. It delivers 40% faster completion and 3x throughput over Mistral Small 3.

Sources: MarkTechPost • TestingCatalog

Video

LTX-2.3: Open-Source 4K Video and Audio Running on Consumer Hardware

Lightricks’ LTX-2.3, a 22B-parameter Diffusion Transformer, generates synchronized 4K video and audio in a single forward pass — up to 50 FPS and 20 seconds. A companion desktop editor runs entirely on consumer hardware. Both model variants are on Hugging Face under Apache 2.0 (free commercial use for companies under $10M revenue).

Sources: VP-Land • Cliprise

Claude Code v2.1.87 Patches Cowork Dispatch; Copilot CLI v1.0.13 Adds Timeline Picker

Claude Code v2.1.87 (March 29): Fixed messages in Cowork Dispatch not getting delivered — targeted hotfix to the messaging pipeline introduced in v2.1.85/86.
Copilot CLI v1.0.13 (March 27): /rewind command and double-Esc now open a timeline picker for rolling back to any conversation point. MCP servers can request LLM inference with explicit user approval. MCP servers blocked by allowlist visible in /mcp show. Fixed reasoning effort for BYOM providers. Removed gemini-3-pro-preview model support.
Copilot CLI v1.0.13-1 patch: Faster startup via V8 compile cache; MCP registry lookups retry with timeouts.

Sources: Claude Code GitHub • Copilot CLI GitHub

GitHub Trending

Repo	Language	Stars	Description
bytedance/deer-flow	Python	~52.4k (+18.9k this week)	ByteDance’s multi-agent framework for research, coding, and content
microsoft/VibeVoice	Python	~26.8k (+1.2k today)	Open-source frontier voice AI for real-time speech interaction
TauricResearch/TradingAgents	Python	~43.9k (+8.2k this week)	Multi-agent LLM financial trading framework
twentyhq/twenty	TypeScript	~42.8k (+451 today)	Community-powered open-source CRM alternative
SakanaAI/AI-Scientist-v2	Python	~3.8k (+855 this week)	Automated scientific discovery via agentic tree search
fastfetch-cli/fastfetch	C	~21.1k (steady)	Performance-oriented system information tool (neofetch successor)
J-x-Z/cocoa-way	C	Rising (HN viral)	Native macOS Wayland compositor for Linux GUI apps

Source: Trendshift • Star counts as of March 29, 2026