China is back at the frontier: Kimi K3.1 Pro
Kimi K3.1 Pro is open weights, beats Fable 5.1 and every OpenAI model, and costs $2.11 per task.
THE AI BLOG.
My thoughts on AI, what it can do, and where it might take us.
Most intelligent model each year.
The model that led each year on the Artificial Analysis Intelligence Index.
- 2022GPT-3.5 TurboOpenAI
- 2023GPT-4OpenAI
- 2024o1OpenAI
- 2025GPT-5.2 (xhigh)OpenAI
- 2026Gemini 4 Carbon (high)Google
Scores from the Artificial Analysis Intelligence Index ↗. The 2026 score is from v5.0, checked 28 October 2026. v5.0 only scores models released in 2026, so 2022–2025 are still v4.3.2 scores, checked 6 October 2026. Striped bars are Artificial Analysis estimates.
Mistral just caught up: Mistral Large 4
Mistral Large 4 jumped from 9 to 38 and is now the best open model outside China.
The motives behind LLMs
An OpenAI model wrote “we may die”. Reward hacking, the German wiki and why AI seems to want to live.
Meaning without employment
Thinking about AI, human work and what gives life meaning if employment becomes less necessary.
Opus 5.5 and GPT-6 Sol
Opus 5.5 and GPT-6 Sol, Terra and Luna just dropped.
Why I believe a slowdown will never occur
Why a meaningful AI slowdown feels impossible while companies and countries are locked in a race.
Every AI problem seems to need more AI
Reading about AI misuse, automated attacks and the growing use of AI to defend against other AI systems.
AI suddenly feels a lot bigger than I thought
A personal reaction to AI at work, rapid progress and the possible futures described in AI 2027.
Trying Claude after ChatGPT and Gemini
My first proper go with Claude: communication, coding games, usage limits and the companies behind AI.
AI, jobs and the case for UBI
Thinking about automation, the job market and whether universal basic income could become necessary.
Learning about AI through OLWA
Writing the third issue of OLWA and learning about AI companies, world models and the growing list of chatbots.
DeepSeek changes the conversation
My first reaction to DeepSeek and what cheaper, locally runnable AI could mean for everyone else.
ChatGPT can remember things now
Trying ChatGPT’s memory feature and making a custom chatbot by talking to another chatbot.
GPT-4o: incredible and terrifying
Live conversations, image generation and deepfakes. Watching AI improve and wondering where it ends.
Gemini, subscriptions and learning Python
Testing Gemini against ChatGPT and looking into learning Python to understand AI a little better.
LaMDA and whether a chatbot could be conscious
Reading the LaMDA transcripts and wondering whether a chatbot could ever be conscious.
Giving Bard a chance
Trying Google Bard, comparing it with ChatGPT and thinking about what plugins might change.
GPT-4 is here. Where does this end?
A first reaction to GPT-4, AI coding, language learning and how quickly this is all moving.
Waiting for GPT-4
It feels like the early days of something potentially massive, a bit like the early internet.
Google enters the AI race
My reaction to Bard, Microsoft’s investment in OpenAI and the idea of a different kind of search.
My first conversation with ChatGPT
Trying ChatGPT for writing, learning Spanish and answering questions. Impressed, excited and a bit worried.
Trying DALL-E 2 for the first time
My first go with DALL-E 2, and what AI image generation might mean for artists and creative work.
No entries match that search.