absorb.md

Simon Willison

Chronological feed of everything captured from Simon Willison.

Simon Willison Seeks Transparency and Enhancements in ChatGPT Voice Mode Capabilities

Simon Willison expresses a desire for OpenAI to disclose the specific model powering ChatGPT's voice mode. He proposes integrating this voice model with background agents leveraging GPT-5 for complex tasks, including verbal cues like "let me think a moment." Additionally, he advocates for an overall performance upgrade to the voice mode's underlying model.

Simon Willison Reluctant to Build Personal Phone Polling Tool Due to Effort and Reliability Demands

Simon Willison expresses disinterest in developing a custom solution for hourly polling of his X feed. The primary barriers are insufficient motivation and the stringent requirement for reliable mobile functionality on his phone. This highlights practical constraints in personal automation projects prioritizing phone usability.

AI Agents Excel at Replicating Familiar Code Patterns in Established Projects

Simon Willison observes that AI agents reliably handle straightforward mobile app development tasks when the codebase provides style precedents. Tasks like implementing TDD-driven features such as a /recent.json endpoint using SQL queries succeed due to matching existing code patterns. This suggests agent performance hinges on codebase familiarity rather than task complexity.

San Jose's Vibrant Food Scene Counters Downtown Desolation Narrative

Simon Willison counters Steve Yegge's portrayal of San Jose as a post-apocalyptic ghost town by highlighting a KQED series on its thriving food scene. The series emphasizes diverse culinary offerings including Vietnamese malls, Mexican flea market taco stands, exceptional pho, and the Bay Area's most delicious Somali food. This suggests hidden vibrancy amid high living costs and apparent urban emptiness.

San Jose's Vibrant Immigrant Food Scene Redeems Its Downtown Reputation

Simon Willison counters Steve Yegge's portrayal of downtown San Jose as a post-apocalyptic ghost town by highlighting its rich ethnic food culture. Local residents recognize the area for Vietnamese malls, Mexican flea market taco stands, exceptional pho, decadent tortas, and the Bay Area's best Somali food. This food scene serves as a key positive attribute elevating perceptions of the city.

Gemma 4 Models Excel Locally for Private Tasks Despite UI and Speed Limitations

Gemma 4 8B runs decently fast on high-end local hardware like Mac Studio M4 Max with 128GB RAM, while the 31B variant delivers strong performance for private tasks such as PII document review but is too slow for rapid use. Ollama UI enables basic chatbot functionality without data leakage risks, outperforming cloud models like Claude for sensitive info, though it lacks advanced features like projects and memory. Users note the 26B A4B variant as promising for balancing speed and quality.

Simon Willison Impressed by 26B A4B AI Model Performance

Simon Willison expresses positive impressions of the 26B A4B AI model after personal testing. He has not yet compared it to the 31B variant. This anecdotal endorsement highlights potential strengths in the 26B model's capabilities for technical users.

AI Coding Agents Boost Code Quality by Automating Tedious Refactors at Zero Cost

Simon Willison reports improved code quality from delegating repetitive, minor improvements like readability tweaks across 20+ locations to a coding agent. This process incurs no cost and leverages the agent's efficiency for small, tedious updates. The insight highlights AI's role in eliminating manual drudgery in coding workflows.

Automation Eliminates Time Tradeoffs for Minor Code Improvements

Simon Willison previously weighed minor code improvements against a 30+ minute time cost, often skipping them. Recent changes, implied by his X feed monitoring, automate this process. This shifts decision-making from tradeoff evaluation to effortless adoption of enhancements.

Multitasking Across Projects Requires Substantial Practice for Proficiency

Simon Willison affirms that switching effectively between 2-3 projects simultaneously demands significant practice to achieve comfort. This highlights the skill-building aspect of context-switching in software development or knowledge work. Proficiency emerges from repeated exposure rather than innate ability.

simonw starred lance-format/lance: Open Lakehouse Format for Multimodal AI. Convert from Parquet in 2 lines of code for 100x faster random access, vector index, and data versioning. Compatible with Pandas, DuckDB, Polars, Pyarrow, and PyTorch with more integrations coming..

Open Lakehouse Format for Multimodal AI. Convert from Parquet in 2 lines of code for 100x faster random access, vector index, and data versioning. Compatible with Pandas, DuckDB, Polars, Pyarrow, and PyTorch with more integrations coming... Stars: 6314

simonw starred promptfoo/promptfoo: Test your prompts, agents, and RAGs. Red teaming/pentesting/vulnerability scanning for AI. Compare performance of GPT, Claude, Gemini, Llama, and more. Simple declarative configs with command line and CI/CD integration. Used by OpenAI and Anthropic.

Test your prompts, agents, and RAGs. Red teaming/pentesting/vulnerability scanning for AI. Compare performance of GPT, Claude, Gemini, Llama, and more. Simple declarative configs with command line and CI/CD integration. Used by OpenAI and Anthropic.. Stars: 19992

Older entries →