GLM-5.2 Open-Source Model Impresses, Markdown Emerges as Programming Language, and AI Agent Disrupts Sales SaaS

Key Takeaways

1. GLM-5.2, an open-source model with 1M token context and MIT license, rivals GPT-5.5/Opus 4.8 and runs locally, suggesting a 2027 local AI trend.
2. Guillermo Rauch declares Markdown as the next programming language, with agent structures deployable via Vercel, making programming accessible.
3. Elon Musk reveals that machine learning work is 90%+ evaluation and data cleaning, with training only 2%, highlighting practical resource allocation.
4. AI agents are disrupting sales SaaS: Mutiny CEO rebuilt the company around AI, as salespeople spend only 21% time selling.
5. A dataset of 2.2 million US municipal laws is now public, challenging government opacity and enabling transparency-driven change.
6. Replit now integrates with LinkedIn, allowing direct portfolio display on profiles.

1. AI Models and Development Paradigms

1. GLM-5.2, a breakthrough open-source model, features a 1M token context window, MIT license, and can run locally on personal machines. It exhibits performance comparable to GPT-5.5 and Opus 4.8, with some preferring its outputs. This has led to predictions that 2027 will be the year of local AI. — via 1 2
2. Guillermo Rauch posits that the next hot programming language is Markdown, paired with a minimal agent structure of agent/instructions.md and skills/ files, deployable in one command via Vercel. This represents the most accessible programming for software fully defined and controlled by humans. — via 1
3. Elon Musk shares that in machine learning projects, the actual work distribution is heavily skewed: evaluation and data cleaning dominate, with training accounting for only about 2% of effort. This indicates that bottlenecks are not in training compute but in data quality and assessment. — via 1

2. Startup Strategy and Product Moves

1. Garry Tan highlights the Mutiny CEO's decision to restructure the company around AI agents, noting that salespeople spend only 21% of time on actual selling, which could undermine the entire sales SaaS model. He also advises putting bad news first on board slides as a forced error-correction mechanism, emphasizing that only good boards will work through problems together. — via 1
2. Replit partners with LinkedIn to allow users to showcase their built projects directly on their LinkedIn profiles, strengthening developer portfolios and bridging self-hosting and social networks. — via 1
3. Clifton Sellers points to Electra's NASA-next-gen aircraft concept, which integrates propulsion and advanced aerodynamics to reduce commercial flight fuel consumption by 17%, representing a gradual path to aviation decarbonization rather than a silver bullet. — via 1
4. Garry Tan recounts how Medium changed its guiding principles from "free and open" to "safe, respectful, and inclusive," burning $2.6M/month, which led to Substack's rise. This serves as a cautionary tale about platform governance shifts that can erode trust and open doors for competitors. — via 1

3. Data Transparency and Governance

1. A new public dataset compiles 2.2 million municipal and county laws from across the United States, making legal opacity hard to maintain. Garry Tan notes that transparency often leads to radical change, as governments work to keep information hidden; this dataset could be a historical catalyst for reform. — via 1 2