Maybe this is part of why I find GPT-5 on ChatGPT so annoying — apparently its system prompt is explicitly set to *not* ask clarifying questions?!? I find it really annoying the way it just goes off and tries to solve the world in one shot. I really want to iterate!
LLMS
-
Questionable Analogies for LLMs: Why They Matter
By
–
Hence why "I love collecting questionable analogies for LLMs" In my experience it's rare to encounter an analogy that isn't easily argued against I still like them if they highlight any interesting characteristic of the thing they are describing…or if they make me smile
-

Grok Code Achieves Top Leaderboard Ranking Beating Claude Sonnet
By
–
Grok Code just hit #1 on the OpenRouter leaderboard, beating Claude Sonnet
-
ChatGPT Outperforms in List Comparison Task
By
–
Another victory for AI: I asked ChatGPT, and its list was much better. x.com/TIME/status/19…
-
Context Ontology: Mapping Sources and Prompt Usage
By
–
been noodling on a concept i’d call a “context ontology” – mostly because I wish I had a single place where I could see where context comes from, and which prompts they are used in. helpful context for an LLM to generate a response or take action might include things like (1)
-
Microsoft Unveils MAI-Voice-1 and MAI-1 AI Models
By
–
Microsoft is leveling up its AI game unveiling two new AI models! MAI-Voice-1: a lightning-fast, natural speech model already powering Copilot Daily, Podcasts & Labs. Think storytelling, guided meditations, even “choose your adventure” audio. MAI-1-Preview: their
-

Local RAG: Index Millions of Documents with 97% Less Storage
By
–
RAG everything locally! Index millions of documents on your laptop using 97% less storage than traditional vector databases. Also work as a drop-in semantic search search MCP for claude Code. 100% Opensource and free with Ollama.
-
Claude Code Sub-Agents Architecture with Haiku Model
By
–
Claude Code uses sub-agents, some of which use Haiku for operations that benefit from a cheaper, faster model. That's not new, they've done that since the start. Here's a trace I captured on June 2nd https://
static.simonwillison.net/static/2025/lo
g-2025-06-02-17-10-25.html
… – see also -
GPT-5 Performance Gains Across Benchmarks Not Gaming
By
–
The papers I have seen argue that this is true for smaller models but not as much major releases. But the gains are broadly similar against a wide range of benchmarks. And GPT-5 is very similar to other models released around the same time. I don’t think it is mostly gaming.
-
Cerebras OSS 120B: 25 cents per million tokens, 3000 TPS
By
–
"But Cerebras hardware costs millions" My man, it's 25c per M tokens for OSS 120B. Let us worry about the hardware bill, you enjoy the 3,000 TPS tokens.
