When AI moves from "Chat" to "Act," the rules of the game invert. You can retry a bad email. You can't retry a bad insulin dose. What’s required next isn't smarter models; it's runtime architectures that make probabilistic models safe for a deterministic world.

Annoying how QNAP Systems, Inc. uses the same channel for both marketing bs emails and security alert emails so if you want to get the security alerts you have to opt-in to the ads. Not cool.

Great to see a fast/small Codex model. I’ll still rely on full fat Codex 5.3 for most wrk but for straightforward tasks Spark should find its way into my rotation imo this chart shows how we should think about using Spark. Both where it sits vs 5.3 & the great gain over 5.2-Mini

Qwen3 TTS / ASR (custom voices, voice design, voice cloning and qwen3-asr) node-based workflow. All features are available in a single Hugging Face space, built with @gradio × Daggr-UI🔥. Try it here: huggingface.co/spaces/prithiv… 1. custom voices:

A good knowledge drop. Shell + Skills + Compaction: Tips for long-running agents that do real work Practical patterns for building with skills, hosted shell, and server-side compaction in the Responses API.

Amazed there’s publications like BPB that still stick DRM on their books & think it actually helps. It doesn’t. All it does is inconvenience honest readers & in many cases including my own loses their authors a customer. Pity the new KubeEdge book looked really promising.

Can’t help think this feature idea came from Cloudflare’s recently released HTML to Markdown feature. It’s definitely a welcome addition though I can’t believe it’s taken this long. Good bye custom tooling going forward if I use one of the Anthropic models that supports this.

Finally in from the airport. I love travelling, meeting people face-to-face & working in different locations is also awesome but man I hate planes. Now a few hrs of sleep then taking 2 hrs off to watch hockey🏒- Go Canada 🇨🇦 then back to coding. Hope the jet lag doesn’t crush me!

Canada 🍁 🇨🇦🍺 What a game. Man I can barely breathe right now. Props to Czechia. Instant classic.

Impressive paper. Nothing to experiment with sadly but lots to learn from.

Interesting. 1st time I’ve seen ChatGPT respond w/ a detailed “error” when (annoyingly 😡) the web tool breaks. 👍 “Note: I tried to sanity-check a couple tool-name details via web browsing, but the web tool is erroring right now, so I’m sticking to stable concepts + design.”

Super useful for fast iteration and calibration of your coding harness.

An interesting paper and review of feedback loops in autonomous agentic systems but I don’t think this study tells the whole story. The desire and yes need for feedback in a driving scenario is very different than many (most) other autonomous use cases.

Crazy! Nanbeige4-3B-Thinking-2511 was only released a little while ago & they’ve already improved the model. I haven’t seen a better model at this size. Still I need to do more testing. Caveat. Like most SLMs this model generates tons of thinking tokens to make up for its size.

Tomorrow is one of the holiest days on the Canadian calendar (men’s Olympic gold medal hockey game) and the usual docile Canadians get very loud. It’d be interesting to see the token use on an average Sunday vs tomorrow. Go Canada! 🍁 🇨🇦🍺

What I’d give for unlimited tokens to operate with! Everything’s coming along fast & I know I’ll have that luxury sooner than later as I keep at it but it’d sure be nice to have them now! Note: So far Codex 5.3 full fat along w/ Spark has been game changing vs Anthropic models.

The best agents are built on an aggregation of simple tools & processes It’s 1978 all over again and the Unix Philosophy rules: Do One Thing Well Composability/Work Together Text Streams = Universal Interface Prototype Early, Iterate Often Rule of Least Surprise Leverage Tools

They were never exactly a “friendly lab” but Anthropic continues to accelerate down a dark path that’s only going to end badly for them imo.