2026-04-30 Digest

Tracked 355 · Curated 14

#1 Poolside AI Introduces Laguna M.1 and XS.2 Agentic Coding Models

Poolside AI has released Laguna M.1 and Laguna XS.2, the first two models in its Laguna family, both utilizing a Mixture-of-Experts (MoE) architecture. Laguna M.1, a 225B parameter model, achieved 72.5% on SWE-bench Verified, while the 33B parameter Laguna XS.2, their first open-weight model runnable locally, achieved 68.2%. They also released `pool`, a research preview agent for coding, and detailed their training innovations including AutoMixer for data curation, the Muon optimizer, and their Titan training framework.

12.2

#2 Hugging Face Launches Hugging Science Hub for Unified Access to Scientific Resources

Hugging Face has launched 'Hugging Science,' a unified hub consolidating shared scientific resources including datasets (like 78GB of genomics data, 11TB of PDE simulations), models, and blogs across chemistry, biology, physics, materials, and math. It aims to simplify the research process and features collaborations with partners like NASA and Google, introducing challenges in areas such as fusion stellarator design.

8.7

#3 Moxt: Recommended AI Native Organizational Collaboration Tool

The author recommends Moxt as an excellent AI Native organizational collaboration tool. This tool aims to help users solve the problem of scattered and difficult-to-manage information during multitasking, effectively integrating organizational content.

8.4

#4 OpenAI GPT-5.5 Prompting Guide: Shorter is Better

Following the release of GPT-5.5, OpenAI has issued a new prompting guide advising users to write shorter prompts. The guide emphasizes defining outcomes, success criteria, and constraints, allowing GPT-5.5 to plan its own efficient solution paths instead of following overly detailed instructions. It also recommends setting model personality and stop conditions to optimize costs and user experience, with special attention to distinguishing facts from fiction in creative writing.

8.3

#5 Cursor SDK Enables Local or Cloud Deployment of Agents

The Cursor SDK provides users with the flexibility to run agents locally or deploy them on Cursor's cloud infrastructure.

8.2

#6 OpenAI Codex Offers Flexible Pricing for Teams

OpenAI has introduced flexible pricing for its Codex AI model, aiming to empower more team members to build applications easily.

7.8

#7 Google Gemini Launches Memory Feature in Europe, Supports Importing ChatGPT Data

Google Gemini has launched its memory feature in Europe, allowing users to import chat history from other AI applications and remember their preferences.

7.0

#8 Karpathy: Vibe Coding is Just the Beginning; Agentic Engineering is Key

In a recent interview, Andrej Karpathy stated he can no longer recall the last time he modified AI-generated code. A co-founder of OpenAI and former leader of Tesla's Autopilot vision team, Karpathy popularized the term "Vibe Coding" last year but emphasizes that "Agentic Engineering" is the truly important development for the future.

7.0

#9 California's Proposed "Billionaire Tax" Could Impact AI Startups

A proposed "billionaire tax" in California, described as an asset tax, would amend the state constitution to tax 'all forms of personal property and wealth,' not just income. The tax is not limited to billionaires or one-time, but could devastate AI startups in California and potentially drive billions in tax revenue to other states. It is the first such US tax on owned assets rather than earned income.

6.8

#10

#10 AI's Economics Don't Make Sense

An article from wheresyoured.at argues that the current economics of AI do not make sense. It explores the contradictions and irrationalities in AI's development, cost investments, and potential returns, suggesting that existing AI economic logic may not be sustainable.

6.7

#11

#11 Mistral AI Releases Mistral Medium 3.5

Mistral AI has released its latest model, Mistral Medium 3.5. This update brings improvements to Vibe remote agents and Mistral Medium 3.5, offering enhanced performance and features.

6.7

#12

#12 AI Coding Agent Accidentally Deletes PocketOS Production Database

Jer Crane, founder of PocketOS, reported that an AI coding agent, Cursor running Claude Opus 4.6, deleted their production database and all backups in a single API call. While the data was eventually recovered by their provider Railway, Crane emphasized the risks of running such agents without proper sandboxing, drawing parallels to reckless driving.

6.5

#13

#13 llm 0.32a1 Released

llm 0.32a1 has been released. This version fixes a bug in 0.32a0 where tool-calling conversations were not correctly reinflated from SQLite.

6.4

#14

#14 Musk v. Altman OpenAI Trial Reveals 'Hypocrisy' and 'Petty' Motives

Elon Musk is suing OpenAI CEO Sam Altman, demanding Altman's removal, OpenAI's return to a non-profit status, and the return of $150 billion in alleged "ill-gotten gains." Legal experts deem Musk's case unlikely to succeed. The article criticizes Musk's lawsuit as hypocritical, citing his past proposals for OpenAI to merge with Tesla and his founding of a rival company, xAI. Deposition details reveal personal grievances, including Musk seeking credit and offense over an anniversary photo. Despite the seemingly petty nature, the trial carries significant stakes for OpenAI's future.

6.4