Vercel Labs’ Agent Desktop uses accessibility trees, not pixel matching

#1 🐙 vercel-labs/agent-desktop

vercel-labs/agent-desktop - **vercel-labs/agent-desktop**, shared by vercel-labs, is an Apache-2.0 CLI and C library that lets AI agents operate macOS, Windows and Linux apps through accessibility trees, using snapshot-specific JSON references instead of pixel matching. It includes reusable agent skills and has 8,127 stars, with 2,243 daily trending gains. Worth trying if you need verifiable, reference-based desktop automation.

#2 𝕏

Mistral AI demonstrated Mistral Large 4 generating a quantum chemistry SCF solver from scratch, validating its output, and building a 3D view from a short prompt. According to the post, the model independently chose how to break the molecule and which orbitals to show without further input.

Also covered by: @Mistral AI

#3 𝕏

LlamaIndex 🦙 retweeted @jerryjliu0’s introduction of Extract Max Effort, described in the post as an agent for document extraction and claimed to be the world’s best frontier agent for the task.

Also covered by: @LlamaIndex 🦙

#4 𝕏

Andrej Karpathy created HackerTube as a weekend hack, using Claude to filter 43M comments across 20 years of HN into 6,454 repeatedly recommended videos in ~2 hours with Opus 5.5. After he created a Google Cloud API key for YouTube metadata, started a VPS, and bought the domain, Claude updated, secured, hardened, deployed, and optimized the no-login app.

Also covered by: @Andrej Karpathy

#5 🐙 garrytan/gbrain-evals

garrytan/gbrain-evals - **garrytan/gbrain-evals** is the MIT-licensed public test suite for gbrain, containing benchmark data, comparison systems, scoring code, and published results. With 449 GitHub stars, it recaps a 500-question LongMemEval test where Opus 5.5 beat particle-1 by 3.4 percentage points but cost about 23 times more per correct answer. Worth trying if you need repeatable evaluations for an AI memory system.

Also covered by: @Garry Tan

#6 𝕏

Aravind Srinivas shared an instruction asking “Computer” to visualize Earth across 400 million years.

#7 in

Peter Yang recapped his conversation with Jaclyn, a PM Director at Google Labs, on how Gemini 4 changed Google’s product-building process. For AI builders, her key advice is to prioritize what should exist, design chore-like tasks for agents first, solve root user problems, and avoid requiring elaborate prompts.

#8 🐙 JustVugg/colibri

JustVugg/colibri - **JustVugg/colibri**, shared by JustVugg, is an Apache-2.0 pure-C engine that runs large open mixture-of-experts models by keeping active parts in RAM and streaming experts from disk; the repo includes 13 documented engines, cross-platform setup, a browser dashboard, API endpoints and MCP tools. It drew 655 daily trending gains in the supplied snapshot. Worth trying if you need local model serving and can meet the 8 GB RAM and 22 GB disk minimums.

#9 in

Peter Yang recapped his Grok Bot tutorials and interviews, including an interview with Peng and Lauren about how the Grok Bot team uses Grok Bot, a setup guide for five useful bots, and the 11 Grok Bots he still uses after 100+ hours. He also said another video featuring one of his favorite new AI features was coming that week.

#10 𝕏

Udi Menkes 🚢 shared insights on AI PM hiring, onboarding, ramp-up, and success profiles at Anthropic, noting that the first 2 months inside one of its PM teams are supposed to feel like some of the hardest of a person’s career. The post says 5,000+ product leaders read GenAI PM.

#11 in

Udi Menkes shared Ramp data showing monthly AI spending per employee of $9,568 for the top 1%, $813 for the top 10%, and $15.70 for the median company—a roughly 609x gap between the top 1% and median. He advises AI product builders to design pricing, limits, and support for high-spending accounts beyond per-seat plans, noting the figures reflect card and invoice spending, not usage or outcomes.

#12 𝕏

Jason Zhou commented that Treg charges per answer with 0% markup on provider rates, no subscription, and $0 for misses. Users do not pay when a provider finds nothing.

#13 𝕏

Santiago announced he is using Grok to identify accounts that consistently post AI-generated slop, adding them to a list for mass muting and blocking while limiting false positives from accounts flagged only once or twice. He plans to share the list later so others can ask their bots to mute those accounts.

#14 𝕏

Lenny Rachitsky shared how WisprFlow built a UGC growth engine where creators are making $100k+/year.

#15 𝕏

Yann LeCun commented that JEPA world models can require considerably less memory and training data than LLMs by emphasizing understanding over accumulated declarative knowledge. He said an unnamed paper shows that data’s similarity structure uniquely determines representation geometry when encoders use anti-collapse criteria such as SIGReg or VCReg rather than reconstruction.

#16 𝕏

Sundar Pichai praised Carbon’s high-quality benchmarks, benchmark headroom, and cost line.

#17 𝕏

Aravind Srinivas shared a 3D map of Leonida paired with real Florida.

#18 𝕏

Guillermo Rauch advocated owning your intelligence, harness, data, connections, runtime, and agents. He urged AI builders to keep their options open, support proliferating local models, and embrace open source, noting that massive models may remain out of reach for local use.

Also covered by: @Guillermo Rauch