Google Releases First Multimodal Gemini Embedding 2
Today's top 24 insights for PM Builders, ranked by relevance from X, YouTube, and LinkedIn.
Google Releases First Multimodal Gemini Embedding 2
#1 đť•Ź
Google AI launched Gemini Embedding 2 last week – its first natively multimodal embedding model now publicly available. It turns text, images, video and audio into unified numeric embeddings, powering tools like video analysis and visual shopping assistants.
#2 đť•Ź
OpenAI launched Advanced Account Security for ChatGPT accounts, adding enterprise-grade controls—SAML SSO, SCIM user provisioning, security-key MFA, IP allow lists, and audit logging—to centralize and harden team access management.
#3 đť•Ź
Claude launched Claude Security in public beta for Enterprise customers, scanning codebases for vulnerabilities, validating findings to reduce false positives, and suggesting patches for review and approval.
#4 đť•Ź
clem 🤗 Anthropic has revoked OpenAI’s access to its Claude API, accusing OpenAI of breaching its terms of service by “distilling” the model via the API.
#5 đť•Ź
Google DeepMind is rolling out its AI co-clinician trusted tester program to additional academic and healthcare sites worldwide to capture broader clinician and patient feedback.
#6 ▶️
UPDATE: AI Is Now Closer Than Ever to Automating Content Creation
All About AI
Automates short-form clip creation and upload using FFmpeg, local Whisper, Opus 4.7, YOLO, Light ASD, Remotion and Surf Agent to generate three vertical MP4 clips in under 10 minutes.
- Extracts audio via FFmpeg and transcribes with a local Whisper model (with timestamps), then uses Opus 4.7 to select moments, YOLO for face detection and Light ASD for active speaker detection before reframing to 9:16.
- Processes an 89-minute podcast into three polished MP4 clips in approximately 5–10 minutes using Remotion for captions, zooms, flash effects and meme sound effects.
- Uploads clips through a Surf Agent in the browser, auto-filling title (“A doctor just exposed what’s happening to male fertility”) and setting visibility to Private within seconds.
#7 đť•Ź
Cognition Evinova, AstraZeneca’s health-tech arm, leverages AI agent Devin for regulatory documentation, bug triage, tech-stack migrations, and test automation—cutting regulatory doc prep time ~8× down from the typical 35–40 hours.
#8 đť•Ź
NVIDIA AI: SGLang open-source inference now hits 180 tok/s per GPU on DeepSeek-V4 decoding with ~1 M context on Blackwell hardware. This boost comes from Blackwell-specific hybrid sparse attention optimizations by LMSYS Org.
#9 đť•Ź
claire vo 🖤 lays out a concise glossary of AI agent architecture—defining message, agent (local vs. cloud), sandbox, subagent, tool, hook, connector, MCP, skill, API, and CLI—to clarify how LLMs with tools and connectors interact and execute tasks.
#10 đť•Ź
Santiago argues UI is splitting into headless APIs for AI agents and on-demand generative interfaces for humans, composed at runtime by agents rather than baked into apps.
#11 đť•Ź
Anthropic used its privacy-preserving tool Clio to collect and analyze all data in this study, showcasing its secure data-handling capabilities.
#12 đť•Ź
Andrej Karpathy showed at Sequoia Ascent 2026 that LLMs can build entirely code-free apps like menugen for image-to-image tasks, replace bash scripts with natural-language install.
#13 đť•Ź
Garry Tan sits down with Demis Hassabis to unpack DeepMind’s playbook for turning research breakthroughs (AlphaGo Zero, AlphaFold) into real-world products and charting strategies for safely scaling toward AGI.
#14 đť•Ź
Garry Tan highlights Loudoun County, VA’s AI-driven data-center boom, adding over 14 million sq ft—nearly 25% of U.S. capacity. He notes this surge is generating hundreds of millions in annual tax revenue.
#15 in
Dharmesh Shah suggests reframing FDEs as “Forward Deployed Experts,” deploying deep domain specialists—not just engineers but lawyers, consultants, teachers, etc.—to help customers realize value faster.
#16 in
95e Carl Vellotti tripled individual output with AI—using Claude Code to write 14 PRDs in one week—but daily standups, weekly retros, and monthly governance reviews ballooned meeting time from 6 to 19 hours without reducing the two-week ship cycle.
#17 đť•Ź
Guillermo Rauch says they’re already building organization-level ACLs to support a broad spectrum of use cases.
#18 đť•Ź
bolt.new built an AI agent with the Claude Agent SDK to unify fragmented design systems into a single automated interface for managing components across platforms.
#19 đť•Ź
NVIDIA AI highlights @steipete’s post on how community-driven public audits have strengthened OpenClaw’s security, and asks users to share their real-world Claw use cases.
#20 đť•Ź
Kevin Yien demos an ideal Stripe CLI flow where you “ask Stripe how all my businesses are configured,” apply bulk updates to match a target setup, and get a “done” confirmation. He then asks which console settings or configurations remain inaccessible via the API.
#21 in
Dan Zhang configured a “family manager” agent on Claude Code to scan his inbox and auto-submit his kids’ lunch orders based on past picks. It even debated chicken tenders vs. pizza mid-Zoom, showing how AI agents can seamlessly handle personal chores alongside professional tasks.
#22 đť•Ź
Google DeepMind launched AI Co-Clinician, a research initiative exploring how large language models can partner with clinicians to streamline patient data analysis and support diagnostics and treatment planning.
#23 đť•Ź
Claude launched its new Claude Security suite in public beta for Enterprise customers. Learn more at claude.com/product/claude-security.
#24 đť•Ź
Kevin Yien launched Stripe’s Console AI agent and is asking users to share any feature requests—big or small—to help shape its next set of capabilities.