Cognition announces Devin Apps for private internal tools

Today's top 20 insights for PM Builders from X and GitHub.

Cognition announces Devin Apps for private internal tools

#1 𝕏

Cognition announced Devin Apps, describing them as private, secure, sharable web applications that run directly on Devin’s cloud computer. They are designed for internal tools, preview deployments, and extensions to Devin.

Also covered by: @Cognition

#2 𝕏

Sundar Pichai recapped what Google described as shipping during the week: two new 8th-gen TPUs on Monday; Gemini 4 Carbon, AlphaProof, a code-verified Signal protocol, and Gemini Robotics 2.5 on Tuesday; AlphaGenome 2 for disease research and the Gemini Decisions API in AI Studio on Wednesday; and Gemini 4 Hydrogen and intelligent-eyewear pre-orders on Thursday. He also noted U2’s new album on Friday, explicitly saying it was not from Google.

Also covered by: @Josh Woodward – VP, @Google @GoogleLabs @GeminiApp @GoogleAIStudio

#3 𝕏

Demis Hassabis shared Google DeepMind’s post describing Gemini 4 Hydrogen as built for rapid coding, multimodal synthesis, and everyday workflows. The post mentions “2.5x,” but the metric and comparison are unclear.

#4 𝕏

Boris Cherny shared a post from @claudeai introducing Claude Meetings, which the post says can join team video calls and build documents, whiteboards, and other artifacts with participants in real time.

#5 𝕏

Hugging Face reposted @ggerganov’s statement that llama.cpp v1.0.0 is now available, linking to a Hugging Face blog post.

Also covered by: @Julien Chaumond

#6 𝕏

Thinking Machines shared how it trained Inkling 2 Small with a native 4-bit forward pass from the first pretraining step through the final reinforcement-learning rollout, noting that fewer bits speed training but quantization error can interfere with learning.

Also covered by: @Mira Murati – Now building @thinkymachines. Previously CTO @OpenAI, @NVIDIA AI

#7 𝕏

Google DeepMind is releasing PARDIG and SatWeave, the foundational datasets behind WeatherNext 3, to help researchers and developers build new climate applications. The organization says accurate weather forecasting starts with clearly observing the atmosphere in real time.

Also covered by: @Google DeepMind

#8 𝕏

Google Research shared new breakthroughs from its Antigravity /teamwork multi-agent systems, whose agents are described as solving open problems across mathematics, chemistry, systems software, and ML engineering with academic partners. The collaboration tools are now being made available externally, though access details were not provided.

#10 🐙 google/elephant-goldfish-model

google/elephant-goldfish-model - google/elephant-goldfish-model is an Apache-2.0 open-source plugin shared by Google for designing and reviewing software with AI assistants, with reusable skills for brainstorming, drafting, reviewing, bootstrapping, resolving findings, and detecting document drift. It has 18834 stars and 5560 weekly trending gains. Worth trying if your agent can create the fresh subagent contexts required for reviews.

#11 𝕏

Anthropic shared an Engineering Blog post with 10 recommendations for designing, running, and reporting safety evaluations to assess deployment readiness, identify needed safeguards, and guide training and research priorities.

#12 𝕏

LlamaIndex 🦙 announced that Agentic Plus now costs 25 credits per page, down from 45—a claimed 44% reduction enabling 1.8x more pages with the same credits. The change uses the same API call with no migration, with lower costs attributed to the OCRia-1.5 model family and a tighter agent harness.

#13 𝕏

Aravind Srinivas shared that the browser agent in Computer and teammates now shows users what it’s working on, allowing them to watch live and take over.

#14 🐙 garrytan/g-harness-evals

garrytan/g-harness-evals - **garrytan/g-harness-evals** is Garry Tan’s MIT-licensed evaluation framework for comparing AI agent harnesses with fixed models and tasks; it includes benchmark tasks, harness code, runners, tests, documentation, and recapped findings. It had 81 stars at discovery. Worth trying if you need to compare harnesses by task success, cost per solved task, and completion time.

Also covered by: @Garry Tan

#15 𝕏

Ali Ghodsi says he is 10x better at his job than a year ago and runs the whole company on Genie. He attributes its power to a continually curated Agentic Ontology that helps agents achieve higher quality at significantly lower cost.

#16 𝕏

Dharmesh Shah commented that HubSpot’s MCP connector can return raw CRM records in JSON—including contacts, companies, and object lists—for flexible use. He described it as one of the fastest-growing ways to use HubSpot and said the company is investing in broader system coverage.

#17 𝕏

Guillermo Rauch shared a post from @vercel_dev stating that Vercel Security Review no longer requires a Vercel project to scan GitHub repositories. The source adds that every push to a connected repository triggers a process, but the outcome is unspecified.

#18 𝕏

Santiago highlights a benchmark released by Deepnote that tests agents in a simulated company’s Slack, giving them access to records across 11 services and three minutes to investigate and decide in 50 different situations. It evaluates whether agents know when to help, how to help accurately, and when to stay quiet.

#19 𝕏

Hugging Face shared @SergioPaniego’s post stating that AsyncGRPOTrainer is stable in TRL v1.18.0, with generation and training running concurrently on separate GPUs.

#20 𝕏

Santiago recommended @coderabbitai Secure for open-source security scans, noting that it’s free for eligible users and suggests fixes for detected vulnerabilities. He said $1M in free scans over 10 days found 1,200+ critical vulnerabilities across 1,100 open-source organizations.

Get tomorrow's brief first

Join AI product managers receiving the latest brief before it reaches the public archive.

Subscribe free