AI news6
Meta Superintelligence Labs ships its first model — Muse Spark — nine months after the $14.3B Scale AI deal
Released April 9, Muse Spark is Meta's first foundation model under Alexandr Wang's Superintelligence Labs division, built after the team "rebuilt our AI stack from scratch" over nine months. It is a proprietary multimodal reasoning model with a "contemplating mode" that runs multiple agents competing in parallel, benchmarked as competitive with Claude Opus 4.6 and GPT-5.4 on general reasoning though it lags on coding and ARC-AGI 2. Meta is flagging health reasoning as its particular strength.
cnbc.com ↗Why it matters
First concrete deliverable from the most expensive talent acquisition in the AI race, and the first public signal of whether the Wang-led reset has actually changed Meta's frontier trajectory.
xAI is training seven models in parallel on Colossus 2, including a 10-trillion-parameter mega-model
Elon Musk confirmed this week that xAI's Colossus 2 cluster — now running 700,000+ GPUs at gigawatt scale since January and upgrading to 1.5 GW by end of April — is training seven models simultaneously, including Imagine V2, two 1-trillion-parameter variants, two 1.5-trillion variants, a 6-trillion and a 10-trillion parameter model. Pre-training for the 10T model is a two-month window at roughly $1.5B in compute cost. No other lab has publicly confirmed training anything at the 6T or 10T scale right now.
eonmsk.com ↗Why it matters
Puts xAI in front on raw parameter count for the first time and sets a new compute-scale floor that OpenAI, Anthropic, and Google will have to answer to — or justify not answering to.
HeyGen Avatar V clones a face from a 15-second webcam clip and crosses 0.840 face-similarity
Announced April 8, HeyGen Avatar V builds a photorealistic digital twin — face, voice, gestures — from a single 15-second webcam recording, scoring 0.840 Face Similarity versus Google Veo 3.1's 0.714. It supports 175 languages with phoneme-level lip sync, generates videos up to 30 minutes long, and separates identity from appearance so users can change outfits or backgrounds without re-filming. Now fully rolled out on HeyGen's paid plans.
heygen.com ↗Why it matters
The first consumer AI avatar product to convincingly clear the uncanny-valley identity-consistency problem, which collapses the cost and time gap between professional video production and a laptop camera.
Anthropic rolls out Claude Managed Agents in public beta at $0.08 per session-hour
Anthropic opened its Managed Agents platform to the general public this week, priced at $0.08 per hour per active session plus standard token usage. The pitch is that teams can deploy agents in days rather than building orchestration and memory layers from scratch. It is the first time Anthropic has shipped agent hosting as a first-party managed service rather than leaving it to customer infrastructure.
therundown.ai ↗Why it matters
Anthropic competing directly with AWS Bedrock Agents, OpenAI's Agent SDK, and the agent-framework middleware tier — consolidating the stack inside the model provider rather than around it.
Google ships a native Gemini Mac app and Chrome AI Skills inside the browser
On April 16, Google released a native Gemini desktop app for Mac, and on April 15 it rolled out Chrome AI Skills — one-click prompt templates invocable directly from the browser chrome. The moves extend Gemini beyond the web interface and push AI capabilities into the default places users already work. Both are free tier at launch.
joinsuperhuman.io ↗Why it matters
Google escalating the OS-level distribution fight against ChatGPT's standalone apps and Microsoft Copilot's Windows integration, using Chrome's billion-user install base as leverage.
Equinix launches Fabric Intelligence to auto-manage enterprise AI infrastructure
On April 15, Equinix announced Fabric Intelligence, an AI-native operational layer sitting across its global network that automates deployment, optimisation, and maintenance of enterprise AI infrastructure. It is positioned as the backbone for organisations running AI workloads across distributed sites without dedicated networking teams. Equinix runs over 260 data centres across 70 metros.
newsroom.equinix.com ↗Why it matters
The interconnection layer is being rebuilt specifically for AI traffic patterns, and first-mover infrastructure plays like this will shape which enterprises can scale agentic workloads without re-architecting their backbone.
AI in the nonprofit sector5
DevelopMetrics uses AI to rescue 60 years of USAID evidence after the agency's closure
Stanford Social Innovation Review profiled DevelopMetrics this week — founded by former USAID staffer Lindsey Moore — and its DELLM (Development Evidence Large Learning Model) system, which processed more than 100,000 evaluations from USAID's Development Experience Clearinghouse before the agency shut down in July 2025. DELLM surfaced 10 times more relevant evidence than human analysts, cut drafting timelines from months to hours, and achieved FedRAMP authorisation. The methods are peer-reviewed in Land Use Policy and identified five recurrent development lessons across decades of aid spending.
ssir.org ↗Why it matters
First documented case of AI preserving institutional memory at archive-scale when a federal agency is dismantled, and a template for any closing institution with decades of unstructured evidence.
Gates Foundation AI for Giving Grand Challenge closes applications 28 April with grants up to $150K
The Gates Foundation's AI for Giving Grand Challenge is awarding up to $150,000 per grant to nonprofits, mission-driven for-profits, and partnerships building AI tools that move donors from intent to action, improve philanthropic decision-making, or strengthen global health and development. Application deadline is April 28, 2026 — ten days out. Eligible applicants include charities outside the US if paired with fiscal sponsors.
www2.fundsforngos.org ↗Why it matters
Near-term actionable funding specifically for AI tooling aimed at the giving side of philanthropy, an underfunded use case versus AI-for-programs grants.
Google.org AI for Government Innovation Challenge 2026 opens with $30M pool, $1M–$3M per grantee
Google.org's AI for Government Innovation Challenge 2026 is distributing $30M in total to organisations worldwide applying AI to improve public services, with individual grants running $1M to more than $3M. Nonprofits partnering with public-sector agencies are explicitly eligible. Recipients also gain access to a global accelerator and Google technical support.
google.org ↗Why it matters
Largest current pool of unrestricted funding for AI in government modernisation available to nonprofits, and a direct counterweight to the Recoding America Fund announced last week.
Microsoft Elevate for Changemakers pledges $5B to the nonprofit sector and launches AI for Nonprofits credential with LinkedIn and NetHope
Announced 25 March at Microsoft's Global Nonprofit Leaders Summit, the Elevate initiative commits more than $5B in discounts, donations, and grants to nonprofits over the coming year. It launches an AI for Nonprofits professional certificate co-developed with LinkedIn and NetHope, alongside a Changemaker Fellowship for nonprofit staff with actionable AI projects. Microsoft says it partners with close to one million nonprofits and education systems globally, and named EY and Caribou as additional partners.
blogs.microsoft.com ↗Why it matters
The largest enterprise pledge to the nonprofit sector this year, and the first AI credential purpose-built for nonprofit practitioners rather than generic Copilot training.
Virtuous + Fundraising.AI benchmark finds 92% of nonprofits use AI but only 7% report major impact
The 2026 Nonprofit AI Adoption Report, published by Virtuous and Fundraising.AI, surveyed 346 nonprofits in late 2025 and found 92% use AI tools in some capacity, but only 7% report major improvements in organisational capability. 81% describe AI use as individual and ad hoc, and only 4% have documented, repeatable workflows. Small nonprofits under 50 staff actually report moderate impact at higher rates (41%) than larger organisations (34%), inverting the assumption that scale helps adoption.
nonprofitpro.com ↗Why it matters
Sets the baseline for the "efficiency plateau" the sector is now stuck on — adoption is no longer the bottleneck, workflow integration is.