Google DeepMind Launches Nano Banana 2

Today's top 25 insights for PM Builders, ranked by relevance from X, Blogs, LinkedIn, and YouTube.

Google DeepMind Launches Nano Banana 2

#1 𝕏

Google DeepMind launched Nano Banana 2, a state-of-the-art model for creating and editing images.

Also covered by: @Philipp Schmid, @Jeff Dean, @Demis Hassabis, @Logan Kilpatrick

#2 𝕏

Sundar Pichai introduced Nano Banana 2, Google’s best image model yet, powered by Gemini’s world understanding plus real-time web search images and live local weather for ultra high-fidelity 2K/4K views.

Also covered by: @Philipp Schmid, @Jeff Dean, @Demis Hassabis, @Logan Kilpatrick

#3 𝕏

Qwen launched Qwen3.5 with day-0 support on MLX-VLM, enabling immediate visual-language model integration.

#4 𝕏

LlamaIndex 🦙 built a private equity deal-sourcing agent with LlamaAgents Builder that classifies opportunities into buyout, growth, or minority strategies and extracts key metrics (revenue, EBITDA, growth rates, debt levels).

#5 𝕏

Aravind Srinivas released the new pplx-embed models along with an API to power search and vector DB workflows. Get started via the quickstart guide at docs.perplexity.ai/docs/embeddings/quickstart.

#6 𝕏

Cursor launched Bugbot Autofix, a feature that automatically detects and fixes issues in pull requests.

#7 𝕏

There's An AI For That launched an Anthropic Claude + Ahrefs connector that lets you research keywords, analyze backlinks, compare competitors, track rankings, and monitor brand visibility—all in a single Claude conversation.

#8 📝 PromptLayer Blog

Prompt Repetition Improves LLM Accuracy - Google researchers found that repeating a prompt (copying and pasting it twice) can dramatically improve LLM accuracy on non-reasoning tasks, producing large gains without performance degradation. This technique offers a simple way to boost reliability for many common tasks.

#9 📝 PromptLayer Blog

Benchmarking Gemini 3.1 Pro: Latency, Cost, and Reasoning Trade-offs - Google's Gemini 3.1 Pro, announced in February 2026, advances reasoning capabilities while aiming to avoid higher costs for users. PromptLayer evaluates its latency, cost, and reasoning trade-offs for practical developer usage.

#10 𝕏

Santiago launched Code Review Bench v0—the first independent benchmark for AI coding—testing multiple models on pull-request review tasks with metrics on bug detection accuracy, suggestion relevance, and review consistency.

#11 📝 Simon Willison

Google API Keys Weren’t Secrets. But then Gemini Changed the Rules. - Coverage of a Truffle Security finding that Gemini and Google Maps APIs shared the same API keys, which became dangerous once Gemini privileges were enabled; many public Map keys could now access sensitive Gemini endpoints. The post warns to check for affected keys while Google works to revoke them.

#12 📝 Simon Willison

Hoard things you know how to do - Advice from the Agentic Engineering Patterns guide encouraging developers to accumulate and retain knowledge of techniques and capabilities, which helps when working with coding agents. The piece argues that understanding what's possible and how to accomplish tasks is a crucial part of productive agent-assisted development.

#13 in

Dharmesh Shah argues that the composability and rich documentation of CLIs make them a perfect fit for agentic coding, and suggests platforms like HubSpot should offer a CLI interface alongside APIs and MCP to unlock more powerful agentic workflows.

#14 ▶️

What is Perplexity Computer?

Greg Isenberg

Greg Isenberg demonstrates using Perplexity Computer’s $200/month Max plan with Sonnet 4.6 agents to automate Gmail-connected hyperpersonalized cold emails, daily 8 a.m. EST competitor monitoring alerts, and parallel investor pipeline research into structured spreadsheets.

Perplexity Computer’s Max plan costs $200 per month and grants access to parallel agent tasks powered by the Sonnet 4.6 code model.
In the warm-outbound use case, Perplexity Computer connected to Gmail, researched 30 target companies simultaneously, found founders’ or partnership contacts, drafted hyperpersonalized emails referencing news/funding, and scheduled follow-ups at days 3 and 7.
The automated competitive intelligence agent was scheduled to run daily at 8 a.m. EST to check five competitor podcasts’ websites and X for pricing changes, new feature announcements, blog posts, and episode drops, then send a push notification or email summary only if changes occurred.

#15 𝕏

Boris Cherny highlights that they’ve been building Claude Code’s OSS infrastructure for a while, with core components like Bun and the Agent sandbox already open-source, and discusses the trade-offs around standards and openness.

#16 𝕏

Peter Yang argues that building an MCP server should be the final step—first expose each capability as single-purpose, well-documented APIs with agent-readable docs and ensure your product works outside a human UI so AI agents can seamlessly chain tasks.

#17 𝕏

NVIDIA AI published its 2026 State of AI in Telecom Report, sharing insights on how AI-driven network automation, predictive maintenance, and customer-experience enhancements are accelerating transformation across the telecom industry.

#18 𝕏

Andrej Karpathy designed nanogpt and nanochat to be ultra-forkable repositories. He loves seeing the diverse directions the community takes them.

#19 in

Greg Isenberg reports that Block cut 4,000 jobs (nearly half its workforce) as AI-driven tools and leaner teams redefine how companies operate.

#20 ▶️

Generative Adversarial Networks (GANs) Specialization

Deeplearning.ai

The DeepLearning.AI GANs Specialization teaches how to build and apply Generative Adversarial Networks—including implementing a basic GAN with separate generator and discriminator networks, adding convolutional layers for stability, and conditioning the model to generate specific outputs—using Python and deep learning frameworks like TensorFlow, Keras, or PyTorch.

Week 1: implement a basic GAN comprising two neural networks (generator as “art forger” and discriminator as “art inspector”) in Python.
Prerequisites: understanding of neural networks including convolutional neural networks, ability to code in Python, and experience with TensorFlow, Keras, or PyTorch.
Week 4: extend the GAN to allow conditional generation, e.g., producing specific dog breeds (golden retriever) or modifying face attributes like age.

#21 ▶️

TensorFlow: Advanced Techniques Specialization

Deeplearning.ai

The DeepLearning.AI TensorFlow: Advanced Techniques Specialization uses TensorFlow's Functional API and custom training loops with distribution strategies to build complex multi-input/output and looped neural networks, advanced computer vision models like a "zombie detector," and generative models including style transfer, autoencoders, and GANs.

Course 1 teaches creation of custom models, layers, and loss functions using TensorFlow's Functional API rather than the Sequential API.
Course 2 cracks open the model.fit training loop and applies TensorFlow distribution strategies to synchronize loss reduction across multiple GPUs or TPU cores.
Course 3 covers advanced computer vision tasks such as image segmentation, object detection with a "zombie detector," and model interpretation.

#22 ▶️

Awakening a giant: Cisco President on why AI is critical for humanity’s survival

Lennys Podcast

Jeetu Patel explains how Cisco’s networking, optics, security, and data platform technologies interconnect GPUs—from servers and racks to clusters up to 800 km apart—to deliver critical AI infrastructure for synchronous large-model training.

Cisco has 90,000 employees and streamed a 12-hour AI summit from 9:00 a.m. to 9:00 p.m. with a two-hour break, watched by 43,000 of its workforce.
Cisco networks GPUs from Nvidia and AMD across servers, racks, and clusters, synchronizing data centers 800 km apart for AI model training.
Cisco transformed from 251 siloed products into a platform of loosely coupled but tightly integrated offerings focused on reliability, trust, elegance, and design.

#23 ▶️

When open-sourcing your code goes wrong...

Fireship

The video examines five open-source project failures—including OpenClaw’s 200,000 GitHub stars and acquisition by OpenAI, Faker.js’s v6.6.6 deletion on npm, and Pars’s $85 million Facebook acquisition and 2016 shutdown—highlighting precise metrics and outcomes.

OpenClaw went from a failed side project to over 200,000 GitHub stars in weeks and was acquired by OpenAI.
In 2022, Marac Squires deleted the Faker.js source, published version 6.6.6 with only the text “endgame” on npm, breaking thousands of JavaScript apps with millions of weekly downloads.
Facebook acquired the backend-as-a-service Pars for $85 million in 2013 and shut it down in 2016, after which its server code was open-sourced as Pars Server.

#24 𝕏

Aravind Srinivas shares a new technical report on PPLX-Embed, unveiling state-of-the-art embedding models for web-scale retrieval that outperform existing competitors in accuracy and efficiency.

#25 𝕏

v0 now supports Nano Banana 2 via the Vercel AI Gateway, with automatic linking and zero setup required.

Google DeepMind Launches Nano Banana 2

Get tomorrow's brief first