Gemini Decisions API launches with 1,000-question multimodal scoring cap

#1 𝕏

Sundar Pichai announced the release of the Gemini Decisions API in Google AI Studio, powered by Gemini Decision 1 to score up to 1,000 closed-set questions across text, images, audio, video, and PDFs in one pass. It supports 3 decision types—including multiple choice with up to 255 options—and costs $0.10 per 1 million input tokens with $0 billed for output tokens, approximately 5x cheaper than Flash-Lite.

Also covered by: @Logan Kilpatrick

#2 𝕏

Google DeepMind announced AlphaGenome 2, a new AI model that focuses on specific cells to help scientists understand how a single-letter DNA change may drive disease in some cells while sparing others.

Also covered by: @Demis Hassabis, @Sundar Pichai

#3 𝕏

NVIDIA AI announced Déjà View, a 117M-parameter model that loops a single transformer block to turn multi-angle photos into 3D geometry and camera positions. NVIDIA AI says it matches or outperforms models up to 10x larger across five benchmarks with comparable or less compute, while letting researchers adjust the number of loops to their compute budget.

#4 𝕏

Anthropic announced its initial public offering and linked to https://ipo.anthropic.com/us/.

#5 𝕏

LlamaIndex 🦙 shared a llamaindex.ai blog post about hardening application testing for sensitive documents, covering how testing was conducted, what was found, and the controls security teams ask about.

Also covered by: @LlamaIndex 🦙

#6 🐙 openai/codex-security

openai/codex-security - **openai/codex-security** is an Apache-2.0 project OpenAI shared for scanning code, validating vulnerabilities, generating patches, and verifying fixes. It includes Codex workflows, a CLI, TypeScript SDK, GitHub Action, and examples; it has 12,124 GitHub stars and a 663 daily trending gain. Worth trying if you need security scanning across repositories, Git changes, CI, or containers.

#7 𝕏

Cognition shared how @trybasis works with Devin, including 20-day agent threads. @trybasis builds AI agents that the source says cut 20-hour accounting tasks down to 30 minutes.

#8 𝕏

Google DeepMind shared that Gemini agents generated over 6M lines of proofs in Lean and audited the code, leaving human experts to review only the critical 1%. It characterized this as a key step toward using AI to verify vital infrastructure, from daily communication to power grids.

#9 𝕏

Ali Ghodsi announced Databricks’ next-generation release of Genie, saying its Agentic Ontology enables him to queue background work via phone—unlike a year ago, when tasks required calls, scheduling, and back-and-forth. Three examples from just this week included finding organizational bottlenecks, analyzing Keanu’s Slack and source code, and interpreting culture-survey results, replacing hours of work and many meetings.

#10 𝕏

claire vo 🖤 reshared @kayvz’s report that safe, automatic PR approvals boost product and engineering velocity, with 10-20% of changes auto-merging in their most active repository.

#11 𝕏

Aravind Srinivas shared research on red-teaming the open-source SPACE sandbox platform, saying CTF red-teaming should explicitly prompt models to break host-guest isolation and reporting that non-cyber models performed better at red-teaming and finding exploits, citing GPT 5.6 Sol versus GPT 5.6 Sol Cyber. He said SPACE was robust to sandbox-isolation exploits, while a DNS-spoofing network bypass was discovered and addressed.

#12 🐙 GoogleChrome/modern-web-guidance-src

GoogleChrome/modern-web-guidance-src - GoogleChrome/modern-web-guidance-src contains code and documentation for building and evaluating reusable skills that bring web-platform best practices and browser-compatibility data into coding agents. It has 2,317 stars and 799 daily trending gains. Worth trying if you need guidance on modern web features or tools for scanning legacy patterns.

#13 ▶️

Tabular Foundation Models for Data Science Agents

Deeplearning.ai

Also covered by: @DeepLearning.AI

#14 in

Marily Nika, Ph.D demonstrated a product-development workflow using 4 tools—Perplexity, Claude, Google Stitch, and Google AI Studio—to move from idea to prototype. She used Reddit for market research and two agents representing pro and opposing sentiments that converse until they recommend what to build to find product-market fit.

#15 ▶️

Claude Code + Obsidian: My family’s AI brain

How I AI Podcast

#16 𝕏

Aravind Srinivas shared that Perplexity Teammates can run a marketing studio handling end-to-end product video creation, an AI use case for product and marketing teams.

#17 🐙 QwenLM/qwen-code

QwenLM/qwen-code - QwenLM/qwen-code is an Apache-2.0 open-source AI coding agent with terminal, editor, desktop, browser, and chat interfaces, plus TypeScript, Python, and Java SDKs. It has 29593 GitHub stars and 151 daily trending gains. Worth trying if you need a coding agent that supports multiple cloud and local model providers.

#18 𝕏

Jason Zhou shared an AI UGC method, quoting an @treg_ai announcement, and characterized it as 25x cheaper—$0.2/s instead of $50 per video. He said the method had been used to create millions of impressions.

#19 𝕏

There's An AI For That recapped Rime’s demonstration of its personal calling agents, saying one agent canceled a hotel booking during a meeting. The post lists 4 setup steps: copy the prompt from the ALT text, paste it into Muse, add an API key, and make a call.

#20 𝕏

Lenny Rachitsky recapped his fifth twice-yearly review of hiring and job-opening trends, highlighting 7 findings for 2026: product managers are gaining in the AI era, engineering demand has stalled, design and remote openings are at record lows, every role requires AI skills, jobs are concentrating in the Bay Area, and growth is occurring mostly outside software.

Also covered by: @Lenny Rachitsky, @Udi Menkes