Meta Releases SAM 3.1 with Object Multiplexing

If you want to master building enterprise-level AI products from scratch and future-proof your career with the most important AI skills, this #1 AI PM Certification is for you.

Run by real AI operators: Rohan Varma (ex-cofounder, first ever PM at Cursor, now at Codex) & Henry Shi (former co-founder of Super.com, a $200M+ ARR company, now technical staff at Anthropic Labs, working directly with the founders).

3,000+ AI PMs graduated  ·  1,000+ reviews on Maven — highest in any category

#1 𝕏

AI at Meta released SAM 3.1, a drop-in upgrade to SAM 3 that adds object multiplexing for much more efficient, high-accuracy video processing on smaller hardware—model checkpoint and codebase now available.

#2 𝕏

Logan Kilpatrick shares that Google AI Studio’s latest release lets you go from plain-English prompts to a fully deployed app (with auth, database, and backend) in a single browser tab, and he and @ammaar will demo it live on April 1.

#3 𝕏

LlamaIndex 🦙 unveiled intelligent table extraction that goes beyond basic OCR by reconstructing spatial relationships, preserving header hierarchies, and validating data across complex PDFs.

#4 𝕏

Harrison Chase points to Vic’s LangChain Agent Evaluation Readiness Checklist as a go-to, step-by-step guide for taking AI agents into production.

#5 📝 Eleanor Berger & Isaac Plath

I’ve configured instructions in AGENTS.md, but the agent isn’t following them. What should I do? - A troubleshooting post about why an AI agent might ignore instructions stored in AGENTS.md and what to check to get the agent to follow them. It offers practical checks such as verifying the file is loaded, confirming formatting and precedence, and testing changes with simple prompts.

#6 𝕏

Philipp Schmid: Starting April 1, the Gemini API billing tiers get monthly spending caps that pause the API once reached (resuming next month or upon upgrade), with faster automated tier upgrades. You can also set per-project spend limits directly in AI Studio.

#7 𝕏

clem 🤗 pushes enabling 50K inference-provider models, 3M Hugging Face models, llama.cpp local inference and BYO training to deliver real model choice over costly, biased cloud lock-in.

#8 in

Peter Yang built an /exec-review AI skill that models a Meta VP’s decision-making framework for on-demand product feedback. It includes core principles, feedback patterns, communication style, a review checklist, and example comments.

#9 𝕏

DeepLearning.AI introduced AToken, a single tokenizer-and-encoder model that processes images, videos and 3D objects, matching or surpassing specialized models. It enables cross-media knowledge transfer to slash training data requirements.

#10 𝕏

Aravind Srinivas says Perplexity is now powering AI in Samsung’s browser pre-installed on 1B+ devices (100M+ active users). This deepens their Samsung tie-up beyond Bixby to Galaxy S26 preloads alongside Gemini.

#11 𝕏

Demis Hassabis launched a desktop import feature in Gemini that lets users transfer their preferences and chat history from other AI apps in just a few clicks.

#12 𝕏

Santiago says Dreamina Seedance 2.0 with Dreamina is the best video model yet, offering control via @-referenced images, camera-motion transfer, style/audio blending, and element-level editing—like editing photos for video.

#13 𝕏

Guillermo Rauch says they don’t routinely ask for extra verification but will prompt users when risk signals—like creating multiple accounts—arise.

#14 𝕏

Mustafa Suleyman announced a partnership with Crusoe to build a 900 MW AI factory in Abilene, Texas, significantly expanding their AI fleet capacity.

#15 𝕏

Josh Woodward unveiled Gemini’s new Import Memory feature, letting users seamlessly transfer chats and personal data from other chatbots. Access it via Settings > Import memory to Gemini, with “importing memory surprisingly smooth.”

#16 𝕏

AI at Meta launched SAM 3.1 with object multiplexing, enabling the model to track up to 16 objects in one forward pass and doubling video throughput from 16 to 32 FPS on a single H100 GPU by cutting redundant computation.