This Week in One Line

Google shipped Android 17 with Gemini features, Adobe put AI assistants into Creative Cloud, OpenAI hired Gemini co‑lead Noam Shazeer, and Baseten lined up $1.5B — together pointing to faster, more automated, and more governed AI in everyday work.

Week in Numbers

• $1.5B — Reported size of Baseten’s new round at a $13B valuation, signaling an inference platform land‑grab. ¹
• 1.6 GW — Meta’s contracted AI data center capacity with Crusoe across two U.S. sites (Texas and Missouri). ²
• Up to $85M — Elastic’s planned acquisition price for DeductiveAI to automate incident response. ³
• 27x — Time To First Token (TTFT) speedup from Execution‑State Capsules at 16k tokens on RTX 5090. ⁴
• 22% — FLOPs reduction from Variable‑Width Transformers versus uniform width at matched loss. ⁵
• 7,031 — Labeled cases in ClinHallu to pinpoint where medical MLLMs hallucinate. ⁶
• $234M — Sarvam’s raise at a $1.5B valuation to scale India‑focused AI. ⁷

Top Stories

Baseten reportedly lines up $1.5B for inference at a $13B valuation

TechCrunch, citing the Wall Street Journal, reports Baseten is finalizing a $1.5 billion round at a $13 billion valuation, just months after a $300 million Series E at $5 billion and a $150 million Series D. Terms are “split‑priced,” with some investors buying at $13B and others at $11B, a signal that serving models after a user prompts them — inference — is where capital is concentrating. For buyers, this points toward more multi‑model routing to balance quality and cost at scale. ¹

Meta inks Crusoe deals for ~1.6 GW of AI compute

Meta agreed to purchase roughly 1.6 gigawatts of AI data center capacity from Crusoe across Childress, Texas, and Warrenton, Missouri; Reuters notes 1 gigawatt can power about 750,000 U.S. homes. Meta has also pledged to invest $600 billion in U.S. infrastructure and jobs over three years, underscoring the scale of its AI build‑out. Expect faster inference and more model experiments to flow from capacity like this. ²

Adobe brings AI assistants to Photoshop, Premiere, Illustrator (public beta)

Adobe launched prompt‑driven assistants across Creative Cloud apps to handle chores like reorganizing video timelines, batch‑renaming clips, managing layers, swapping backgrounds, and flagging missing fonts. The assistants plug into creative workflows you already use, acting like agentic helpers while keeping creators in control. For teams, this is a practical way to standardize prompts for repetitive setup work and measure time saved. ^{8 9}

Android 17 ships with Gemini‑first features across devices

Google’s Android 17 arrived with a new multitasking “bubble bar,” selfie+screen dual recording for reaction videos, and Quick Share that now plays well with Apple’s AirDrop on select Pixels. Gemini Omni can edit videos inside a chat, while Lyria 3 generates music from text or images; Wear OS 7 adds multi‑step automation and claims up to 10% better battery life. The message: more creation and translation steps now live in the OS (Operating System), not add‑on apps. ¹⁰

OpenAI hires Google’s Gemini co‑lead Noam Shazeer

On Jun 17, Google VP of engineering and Gemini co‑lead Noam Shazeer said he will join IPO‑bound OpenAI; Shazeer co‑authored the 2017 Transformer paper that underpins today’s Large Language Models (LLMs). Media coverage frames this as a move to strengthen OpenAI’s model architecture research, while the company also adds policy expertise with a Strategic Futures team. Talent and governance both matter as labs iterate quickly. ^{11 12}

G7 pitch for a U.S.-led AI coalition

At a Jun 17 luncheon in Évian‑les‑Bains, Anthropic’s Dario Amodei and Google DeepMind’s Demis Hassabis urged heads of state to form a U.S.-led group for shared AI rules, including structured access to frontier models and chip trade excluding China, per CNBC. With OpenAI’s Sam Altman in the room and other labs attending, the push signals closer alignment between government and private labs on testing standards and risk coordination. ^{13 14}

Elastic moves to acquire DeductiveAI for up to $85M

Elastic agreed to buy DeductiveAI, an AI SRE (Site Reliability Engineering) startup that triages incidents and fixes software bugs, for up to $85 million, TechCrunch reports. The plan is to embed agent‑style monitoring and remediation inside Elastic’s observability suite, reducing manual toil in incident response. For operators, clean runbooks and telemetry become even more valuable if automation is handling first steps. ³

Meta adds “AI Mode” to Facebook search and creation tools

Meta’s AI Mode turns public posts across Facebook — including Groups and Reels — into direct answers to natural‑language queries, and adds AI editing features like collage cutouts and profile photo restyling. Because the system summarizes user content rather than vetted sources, freshness and accuracy need watching; brands can seed better answers by posting clear, up‑to‑date Q&A in official channels. ¹⁵

Sarvam raises $234M at a $1.5B valuation for India‑focused AI

Bengaluru‑based Sarvam closed $234 million (part of a targeted $300 million Series B) led by HCLTech, citing 2 million daily conversations and 10 million daily API (Application Programming Interface) calls. The company focuses on Indian languages and enterprise deployments, a regional counterweight as export controls complicate access to frontier systems. ⁷

Execution‑State Capsules slash first‑token delay up to 27x

A new paper shows sub‑millisecond GPU‑resident snapshot/restore and up to 27x faster TTFT (Time To First Token) at 16k tokens by capturing and restoring a model’s entire execution state — complementing, not replacing, high‑throughput KV (Key–Value) cache serving. For interactive agents, speech systems, and robotics, this implies snappier turn‑taking without full recompute. ⁴

Trend Analysis

Consumer platforms embedded AI further into default workflows while creative suites added agent‑like helpers. Android 17’s bubble bar and Gemini‑powered media tools, Meta’s AI Mode that answers from public Facebook posts, and Adobe’s assistants that reorganize timelines or manage layers all push everyday creation and search into AI‑mediated flows inside the products people already use. For non‑specialists, this means fewer app hops and more promptable steps in routine work. ^{10 15 8}

Reliability and verifiability took a front seat. Funding flowed to harness‑style validation and formal verification — Probably’s $9M seed to pair LLMs (Large Language Models) with deterministic validators, and Pramaana’s $27M seed to encode domain rules for auditability — while research like ContextRL (Context‑aware Reinforcement Learning) trained models to choose evidence, not just answers. The common thread is grounding outputs in checks that people can inspect. ^{16 17 18}

Compute and serving both scaled and got smarter. Capital concentrated at the inference layer (Baseten’s reported $1.5B at $13B) while Meta secured ~1.6 GW of capacity; at the same time, system papers offered practical latency and cost wins — Execution‑State Capsules’ up to 27x TTFT improvement, Variable‑Width Transformers’ 22% FLOPs cut, and LMCache’s dedicated KV cache layer for reuse. Together, these signals point toward cheaper, snappier production workloads. ^{1 2 4 5 19}

Policy set the backdrop. At the G7 on Jun 17, lab leaders pushed a U.S.-led AI coalition for shared testing and access rules, echoing a week where export controls already limited model access for some users. For global teams and buyers, this raises the premium on model fallbacks, documentation of testing, and clarity on who can access “frontier” capabilities. ^{13 20}

Watch Points

• “U.S.-led AI coalition” — If this reappears, it’s the G7‑sparked push for shared testing and access norms led by major labs. ¹³
• “Baseten $1.5B close” — Terms and use of proceeds will signal how fast the inference platform race is tightening. ¹
• “KV cache / TTFT” — Mentions of LMCache, Execution‑State Capsules, or similar indicate stateful serving and latency cuts moving from papers to production. ^{19 4}

Open Source Spotlight

• LMCache — A dedicated Key–Value (KV) cache layer to reuse attention memories and cut recompute in Large Language Model (LLM) serving; useful for long chats and high‑traffic apps. LMCache/LMCache
• Rivet Actors — Lightweight, long‑running actors with built‑in persistence for agent backends; good for teams building stateful AI agents and collaborative apps. rivet-dev/rivet
• LocalAI — Run language, vision, voice, image, and video models locally on CPU (Central Processing Unit); a fit for privacy or offline constraints. mudler/LocalAI
• Pathway — Python ETL (Extract, Transform, Load) for streaming analytics and Retrieval‑Augmented Generation (RAG) pipelines; now with an Elasticsearch reader for reliable incremental syncs. pathwaycom/pathway
• Dyad — A local, open‑source AI app builder that runs entirely on your machine; handy for quick, private prototyping with your own API keys. dyad-sh/dyad

What Can I Try?

1. Test Adobe’s AI assistants: in Photoshop or Premiere, prompt the assistant to reorganize layers or batch‑rename clips and log minutes saved on a routine file. ⁸
2. Update a Pixel to Android 17: use the bubble bar for a workday and measure app‑switch taps; try in‑chat video edits with Gemini Omni on a 15–30 second clip. ¹⁰
3. Add a “citation + audit trail” rule: for one weekly metric, only accept AI answers with sources and a reproducible log — a Probably‑style harness. ¹⁶
4. Try a KV cache layer: install LMCache’s CUDA 12.9 nightly and read the quickstart to see where cache reuse fits your stack. ¹⁹
5. Standardize multi‑model calls: spin up LiteLLM’s proxy, route OpenAI‑compatible requests through it, and check signed Docker images via cosign. ²¹