The AI landscape is shifting again. QWQ-32B from Qwen has just thrown down the gauntlet, demonstrating performance gains that not only surpass distilled DeepSeek R1 32B models but also challenge DeepSeek-R 1-671B on key benchmarks. This isn't just incremental progress; we're witnessing a leap.

My initial experience with QWQ has …

How is AI making you more effective, efficient or creative?

Recent advances in coding models and pair-programming tools like Aider have completely changed how I build tools. What used to take hours—or remain unfinished due to lack of time—is now possible in 30 minutes or less.

In the …

Think multimodal AI needs massive models? Think again!

Microsoft's Phi-4-multimodal is here to challenge that, packing speech, vision and language processing into a lean 5.6B parameters. What's the secret? A "mixture-of-LoRAs" architecture that allows it to handle different data types in parallel, creating a truly unified experience. This means …

A few hours ago, Pliny obtained and shared the new system prompt for Claude v3.7, noting only a "few small differences" from the one Anthropic had previously published. Interestingly, the prompt even includes a strawberry-flavoured Easter egg. This level of openness fosters trust—without which there is no foundation …

"Perplexity.ai open-sourced R1 1776, a version of the DeepSeek-R1 model that has been post-trained to provide unbiased, accurate, and factual information."

DeepSeek’s release of R1 was a generous contribution to the world, showcasing innovations in efficient pre-training, attention mechanisms and the sparse use of mixture-of-experts (SMoE). This new …

Thinking Machines – the latest AI research and product company. They plan to:

• Frequently publish technical blog posts, papers, and code.
• Build things correctly for the long haul, maximising both productivity and security rather than taking shortcuts.
• Contribute to AI safety by: -- Maintaining a high safety bar—preventing misuse of released …

How comfortable would you be attending a meeting with delegate agents or sending your own delegate? Is this a cultural step too far? While I was not surprised by the topic of this paper, it raises some important questions on transparency, authenticity, cultural acceptance and data ownership.

We've all heard the adage: "You only get one chance to make a first impression." While that first impression certainly carries weight, it's not the be-all and end-all. There's hope for recovery, thanks to a psychological principle known as the peak-end rule.

Pioneered by Nobel laureate Daniel Kahneman and Barbara …

Can AI help a child prepare for a spelling test?

That’s the question I asked this week while helping my daughter study for hers. Instead of just drilling words the old-fashioned way, I wondered—how hard would it be to build an AI-powered spelling assistant?

With Open WebUI, it …

Ever wondered how a transformer (LLM) works? How input text is tokenised, embedded, normalised and passes through multi-head self-attention layers? Well, someone decided to make it into an interactive visualisation. Wow.