AI Weekly Update - June 3, 2025
Perplexity's agent is here, Apple skips AI for this year's WWDC
what to know for now
🤖 Perplexity Labs rolls out agent workflows. The Pro-tier feature chains search, reasoning, and document creation, generating live spreadsheets, dashboards, and PDFs from a single prompt. Perplexity pipes results through its in-house Mixture-of-Experts stack and stores workflow graphs for reuse. Read more
🗣️ Claude gets full voice chat. The beta voice mode streams recognition and synthesis together, cutting latency to 150ms and supporting context windows up to 90 seconds. Mobile clients tap the same audio encoder that powers Opus but with Anthropic’s safety filters in front. Enterprise users can pin domain vocabularies. Read more
🖼️ Flux Kontext brings in-context image generation. BFL’s FLUX-1 suite mixes text and reference images, extracting visual concepts on the fly for precise edits. The flow-matching pipeline swaps diffusion for faster sampling and exports editable latent vectors that slide into Photoshop. Read more
🎵 Napster resurrects brand as AI platform. Infinite Reality rebranded itself “Napster Corp” and unveiled a generative music workbench that turns stems into interactive songs for web storefronts. Read more
🧪 AI Research of the Week
Compress, Gather, and Recompute: REFORMing Long-Context Processing in Transformers
From KAIST and Amazon AGIJake’s Take: REFORM lets language models read million-token files while keeping GPUs calm. It slices text, stores a compact cache, then recomputes only the pieces in use; tests show higher accuracy, 30% less peak memory, and up to 80% faster inference.
This approach could help whole codebases, legal briefs, and hour-long meeting transcripts fit on a single H100 card, trimming hardware bills and nudging agents toward edge deployment. The downside is that cache mixing can blur token-level provenance, so auditors will need fresh tools (before regulators come knocking).
what to know for later
🍏 WWDC buzz points to only modest “Apple Intelligence” demos. Apple shelved a wide Siri upgrade and slowed on-device model work, so next week’s conference should spotlight incremental OS tweaks without grand generative debuts. Cook says the team needs extra time to “get it right.” Read more
📊 Meta plans end-to-end ad automation by 2026. Internal roadmaps show image-to-ad pipelines where marketers upload a product photo and budget; Llama 4 handles copy, creative variants, and targeting. Meta is pouring billions into Arm-based data centers and open-sourcing guardrails to placate regulators. Ad-agency stocks slid on the leak. Read more
💼 Anthropic boss predicts white-collar wipeout. Dario Amodei warned Congress that LLMs could erase whole professional strata unless policy keeps pace. He urged payroll-linked AI taxes and “aggressive” worker up-skilling, estimating 60% task coverage by 2027. Lawmakers probed liability for model-driven layoffs. Read more
📰 NYT licenses archive to Amazon for Bedrock training. The multiyear deal (reported at ~$60M) lets Amazon fine-tune Titan models on 170-years of articles while adding a fact-checking clause that credits NYT in generated answers. Read more