# Daily Summary for 2026-01-19

## 2026-01-19 16:01:15

# AI Digest - January 19, 2026

## Tips & Techniques
- **Claude Code /clear vs /compact**: /clear creates new sessions (keeping old ones), while /compact replaces conversation with summary. Use /export before /compact to preserve full context. [link](https://x.com/agenteer/status/2013276215281385741)

- **Vibe Coding for Long-Form Content**: LLMs excel at orchestrating narratives in markdown. Edit by modifying files directly—iterate on storyboards, visual sequences, or complex workflows without regenerating everything. Tested with Claude Code + visual agents. [link](https://x.com/spillai/status/2013270403679215951)

- **Use Socratic Method Prompts for Deep Learning**: Structure LLM tutoring with first-principles reasoning, pause for validation, and avoid dumping formulas. Works for complex domains from algorithms to constitutional law. [link](https://x.com/darin_gordon/status/2013275889753108535)

- **Skills Iteration > Perfection**: Don't obsess over getting skills exactly right on first try. Deploy, observe failures, tell AI to fix weaknesses in the skill definition, repeat. Converges faster than expected. [link](https://x.com/dazuck/status/2013273813639692700)

## New Tools & Releases
- **GLM-4.7-Flash**: 30B open-source model with MLA architecture, 4B active parameters, competitive with top closed-source for coding. Runs locally on M3 Ultra at 81 tokens/s (4-bit) via MLX. [link](https://x.com/Zai_org/status/2013272967199773029)

- **FunctionGemma Tuning Lab**: No-code browser tool to fine-tune Gemma directly without coding knowledge. TRL backend handles the heavy lifting. [link](https://x.com/SergioPaniego/status/2013269871547617681)

- **VT Code**: Anthropic Tool Search support + natural language BM25 tool search for Claude Code workflows. Solves the "too many tools" problem. [link](https://x.com/vinhnx/status/2013274499769155663)

- **OctoCodingBench (MiniMax)**: Measures whether coding agents follow process constraints, not just produce working code. Reveals that models struggle combining multiple rules simultaneously. [link](https://x.com/joelniklaus/status/2013280337930150361)

- **Customer Context Graph**: High-signal opportunity—aggregating all customer threads (emails, transcripts, Slack, contracts, configs) into explorable context for agents. Segment for SaaS but for AI-native stacks. [link](https://x.com/aikarjal/status/2012975439275294758)

## Research & Papers
- **Rectified Flows**: Straightens flow model trajectories to reduce sampling steps. Interactive article explains why models learn curved paths and geometric intuition behind the fix. [link](https://x.com/alec_helbling/status/2013251186019754460)

- **AI & Mathematics: Erdős Problems Forum**: Humans + GPT-5.2 + Aristotle reasoning models now solving hard problems at scale. Confirms transition from "Amplified Mathematician" to "Superhuman Mathematician" stage. [link](https://x.com/nasqret/status/2013277192491299139)

- **Deep Delta Learning**: Replaces fixed residual shortcuts with learnable rank-1 transforms. Per-layer control + stability + minimal overhead (one scalar + one vector per block). [link](https://x.com/guitchounts/status/2013274327781663083)

## Industry News
- **Salesforce's Agentforce Struggles Signal Paradigm Shift**: Three pricing pivots in 12 months indicate weak product-market fit. Opportunity for "anti-Salesforce"—outcome-based pricing, agents-first design, no click-button bloat. [link](https://x.com/thedealdirector/status/2013273189653127659)

- **Open-Source Models Dominating Developer Adoption**: Qwen crossed 700M downloads on HuggingFace (Dec alone > next 8 combined). GLM-4.7, DeepSeek surging. US lead narrative needs asterisk: *closed-source only. [link](https://x.com/DumbEinstein/status/2013278283429794172)

---
*Curated from 900+ tweets across AI/startup/research feeds*