Platforms Consolidate, AI Federates: Fox–Roku, Salesforce–Fin, and Europe’s Sovereign AI Path
NEWSLETTER | Amplifi Labs
Fox to Acquire Roku, Reshaping Connected TV Platform Landscape
Around the web • June 15, 2026
Fox Corporation plans to acquire Roku, combining a top connected-TV platform with a major broadcaster’s content and ad sales engine. For developers building Roku channels and ad integrations, this could bring policy, monetization, and advertising API changes as the ecosystems are aligned. Watch for updates to SDKs, billing, and data-sharing rules pending deal close and regulatory review.
Enterprise AI Strategy & Deals
Salesforce to Acquire Fin for $3.6B, Expanding AI Customer Support
Around the web •June 15, 2026
Salesforce signed a definitive agreement to acquire Fin for $3.6B. The deal brings Fin’s AI-driven customer support automation into Salesforce’s platform, strengthening native bots, agent assist, and workflow automation. Developers should watch for migration guidance and potential API/SDK changes as Fin capabilities are integrated across Salesforce clouds post-close.
Jane Street builds formal methods team to rein in agentic code
Around the web •June 14, 2026
Jane Street is reversing its long-held skepticism and forming a formal-methods team, arguing that agentic coding has shifted the cost–benefit calculus. The firm aims to embed modular specifications and proof techniques into OxCaml’s type system, integrate with tools like Lean, Dafny, Rocq, Agda, and Iris, and use proofs to ease the verification bottleneck and improve feedback loops for AI code agents. Roles are open in London and New York.
AI Infrastructure, Sovereignty & Trust
EuroMesh: Europe Could Train Frontier AI by Federating Public Compute
Around the web •June 15, 2026
EuroMesh presents a reproducible model and sourced report arguing Europe can train a sovereign frontier-class model by federating existing EuroHPC supercomputers and national AI Factories using low-communication (DiLoCo-style) training. With tens of exaflops already online and mean 7.6-year grid-connection waits for new 1‑GW campuses, the model projects a federated run by ~2028 versus ~2033 if waiting for new datacenters. Feasibility hinges on political coordination, usable HPC share, and unproven-at-scale distributed training, but it outlines a concrete near-term path for EU AI practitioners and policymakers.
Claim: Rio’s 397B LLM Is a 60/40 Nex–Qwen Weight Merge
Around the web •June 14, 2026
Nex-AGI alleges the Hugging Face model “Rio-3.5-Open-397B” is not original but an element-wise blend of its Nex-N2_pro and Qwen3.5-397B-A17B (0.57/0.43). As evidence, identity probes without the bundled “You are Rio” system prompt led the model to self-identify as Nex in 79% of 120 queries and reproduce Nex-AGI’s bespoke backstory, while tensor-by-tensor analysis across all 60 layers showed near-perfect collinearity (0.98–0.99) and a consistent mixing weight (0.571). If accurate, the findings highlight model-provenance verification techniques and raise compliance and trust concerns for teams adopting or redistributing “open” LLMs.
Designing Trustworthy AI Products
Apply IA to Agents: Taxonomy, Retrieval, and Memory Design
Nielsen Norman Group •June 12, 2026
The article introduces “context architecture,” applying information-architecture principles to agentic AI so systems interpret context and act more reliably. It outlines concrete practices—taxonomy, controlled vocabulary, unambiguous tool naming, ontologies, and structured memory with scoping/retention rules—to improve findability, retrieval, and tool selection. For teams shipping agents, this offers a blueprint to reduce retrieval noise, align to user mental models, and increase accuracy and trust in multi-step workflows.
Make Generative AI Usable: Define 'Good,' Then Judge-Evaluate-Iterate
Nielsen Norman Group •June 12, 2026
Because generative AI is nondeterministic, teams should replace exact UI specs with objective evaluation criteria that encode user research and design judgment. A judge–evaluate–iterate loop—often using an LLM-as-judge calibrated to human annotations (targeting ~0.8 F1), replay testing, and decomposed judges—helps refine prompts or finetunes and continuously catch regressions. While centered on conversational UX, the approach generalizes to AI features broadly, with visual evaluation expected to improve via linters and accessibility tooling.
Cognitive Inclusion Yields 1.8x More UX Issues and Actionable Insights
Smashing Magazine •June 10, 2026
A Fable-led study with UC Irvine ran 30 interviews across three prototype sites (5 cognitive and 5 gen-pop per site) and found cognitive participants identified 197 vs 113 usability issues and 93 vs 54 suggestions—about 1.8x more on both counts. Cognitive testers exposed deeper problems with content clarity, labels/buttons/links, and icon/visual comprehension and articulated why interactions created cognitive load, revealing barriers gen-pop users often power through. Teams should include cognitive users in routine UX studies, track energy/focus beyond task completion (e.g., AUS), and use cognitive accessibility as an on-ramp to clearer flows, higher trust, and better retention as aging demographics grow.
