Google Launches Gemini 3.1 Flash-Lite and Introduces New CLI for Humans and Agents
Today's top 18 insights for PM Builders, ranked by relevance from X, YouTube, Blogs, and LinkedIn.
Google Launches Gemini 3.1 Flash-Lite and Introduces New CLI for Humans and Agents
#1 đť•Ź
Demis Hassabis launched Gemini 3.1 Flash-Lite, a compact but powerful model delivering lightning-fast inference and optimized cost efficiency. Google introduced a new CLI for humans and agents.
#2 đť•Ź
Google Research introduced a training technique that teaches LLMs to perform Bayesian inference optimally, significantly improving their ability to update predictions and generalize across new domains.
#3 đť•Ź
Josh Woodward launched Cinematic Video Overviews in NotebookLM, automatically handling scriptwriting, visual direction, and critique loops for polished video summaries—now available to Ultra users in English.
#4 ▶️
Build and Train an LLM with JAX
Deeplearning.ai
Build and train a 20 million parameter GPT-2 style LLM from scratch using JAX’s automatic differentiation, just-in-time compilation, and distributed compute features, then run inference via a graphical chat interface.
- Implements a GPT-2 style model with exactly 20 million parameters using JAX’s automatic gradient computation and compilation for distribution across CPUs, GPUs, or TPUs.
- Preprocesses a dataset of stories using JAX’s data loading tools, trains the model while saving checkpoints, and scales training over tens of thousands of chips.
- Loads the saved checkpoint of the mini GPT model to support interactive chatting through a graphical user interface.
#5 đť•Ź
Sebastian Raschka notes that Gated DeltaNet modules don’t increase KV cache size, so Qwen3.5’s 3:1 ratio makes it significantly more memory-friendly than earlier Qwen3 models.
#6 đť•Ź
Philipp Schmid launched a new Gemini Interactions API skill for building advanced agentic apps with Gemini models, installable globally via the Vercel or Context7AI CLIs.
#7 đť•Ź
Harrison Chase rolled out a new LangSmith CLI with built-in tracing and evals, and introduced LangSmith Skills to train agents on using the CLI.
#8 đť•Ź
LlamaIndex 🦙 launched LlamaSplit, a UI tool for defining custom configs to split complex docs into structured categories and extract specific sections.
#9 đť•Ź
Santiago built a Redis agent skill that lets AI autonomously generate Redis code, highlighting 2026’s surge in product-specific agent skills.
#10 đť•Ź
Peter Yang unveils how three AI-native companies—Linear assigns tasks to AI “team members” via natural language, Ramp drives performance by mandating Claude Code usage, and Factory AI packages product management, UI, and data analysis into reusable AI skills—offering concrete...
#11 đź“ť Simon Willison
Anti-patterns: things to avoid - A guide section describing behaviors that are anti-patterns in agentic engineering, with emphasis on avoiding unreviewed code being pushed to collaborators.
#12 đť•Ź
NVIDIA AI shows how micro data centers—compact, distributed facilities tapping underutilized power substations—can deliver low-latency AI inference compute at scale without overloading the electric grid.
#13 đť•Ź
Aravind Srinivas is building a JARVIS-style assistant by integrating Perplexity Computer’s new Voice Mode, so users can simply speak commands to perform tasks.
#14 đť•Ź
Santiago prompts his assistant to reflect on any deviation from its defined skills and then updates those skills to prevent the same mistake from happening again, iterating this process regularly.
#15 ▶️
SaaS is minting millionaires again (here's how)
Greg Isenberg
Greg Isenberg outlines a 30-step AI-powered SaaS startup playbook that uses ideabrowser.com, Manis, Claude Code, ChatGPT for niche validation, content automation and agent workflows, and transitions from per-seat subscriptions to $200-per-task outcome pricing.
- Uses ideabrowser.com to uncover subniche markets such as the FIRE (Financial Independence, Retire Early) movement within Gen Z finance.
- Automates daily scroll-stopping content creation by querying AI tools—Manis, Claude Code, and ChatGPT—for non-obvious viral ideas, scripts, and publishing calendars.
- Replaces per-seat subscription models with per-task pricing, charging approximately $200 for each completed workflow and moving towards outcome-based billing.
#16 đź“ť Simon Willison
Something is afoot in the land of Qwen - Simon comments on Qwen 3.5, calling it a remarkable family of open-weight models from Alibaba, while noting concern about the team's recent high-profile departures.
#17 đť•Ź
Harrison Chase announced LangChain OSS Skills—installable modules for LangChain, langgraph, and DeepAgents that provide prebuilt agent functionalities to accelerate workflow development.
#18 in
Peter Yang: In the AI era, PMs must onboard and manage AI agents — Linear lets you assign tasks to AI teammates conversationally, Ramp insists on Claude Code for peak performance, and Factory offers reusable AI-driven skills for PM, UI, and data analysis.