Hi Reader,

Here are three things I found interesting in the world of AI in the last week:

Gemini Diffusion is a new kind of AI, and very, very fast - announcement​

Google has dropped what might be their most significant technical breakthrough of the year with Gemini Diffusion, a text generation model that completely abandons the traditional autoregressive approach. Instead of generating text token by token like every other LLM on the market, it applies diffusion techniques (think Stable Diffusion for images) to generate entire chunks of text simultaneously.

Diffusion is kind of like sculpting, you start with the raw material (random pixels for images, words for text) and then you remove the stuff you don't want until the final form emerges. There's an animation on their announcement page which will give you an intuitive sense of it.

The speed claims are genuinely eye-popping - up to 1,479 tokens per second according to external benchmarks, making it roughly 5x faster than Google’s previous speediest models. For context, that’s nearly 10x faster than OpenAI’s o4-mini (151 tokens/second) and Gemini 2.5 Pro (152 tokens/second). During Google’s internal demos, the model reportedly generated text so quickly they had to slow down the video playback just to make it watchable.

It's noticeable that their announcement compares it to flash-2.0-lite, a weak and old model instead of one of the newer ones, because heaven forbid if a frontier lab ever published benchmarks where the new model wasn't in the lead. It may be fast but it's not particularly strong, yet.

But it is a meaningful contribution to AI research and a reminder that Google has massive capability in this area, as if we needed reminding given all their big announcements this week - $249.99/month Ultra subscription plan, Veo 3 video generation, Deep Think’s enhanced reasoning, Project Mariner release, new AI mode for search.

Claude Code gets an SDK and GitHub integration - documentation​

Anthropic has taken Claude Code out of its terminal-only constraints with a new SDK and GitHub integration that dramatically expands how developers can build with it. It is now possible to do things like "@claude, fix this bug" on a GitHub issue and have the AI whir away in the cloud and make a pull request in GitHub.

The GitHub Action implementation allows Claude to automatically handle code reviews, PR creation, and issue triage within GitHub workflows. It can analyze changes, suggest improvements, implement fixes, and track progress with visual indicators that update as tasks complete.

There is a ton of activity in this space at the moment and I foresee a substantial impact in the way teams build software. "Is this clear and simple enough to give to an autonomous AI" isn't a question many coders have had to grapple with before.

I've been spending a lot of time (and $) on coding assistants and the more autonomous it is, the more experience and skill it takes to get good code for a reasonable price. But ultimately it's making custom software development a lot cheaper and I think we are only at the beginning of seeing what the long term impacts of that are.

As one hypothetical example. Lots of tools have got the ability to work with plugins (browsers, wordpress, xero, github etc.) where anyone can write some code and extend functionality of the tool. Currently it takes an experienced dev and a bunch of time to make them, or a vibe coder driving an AI assistant. With the Claude Code SDK (think programming library) devs can now ship software with "auto plugin generation" features - imagine just telling your browser how you wanted it to work and it writing the code it needs to do it, testing it and then shipping a new feature just for you.

I've no idea if we will see this level of personalisation through in app AI powered plugin development, but the fact that it is a feature that any regular dev can contemplate is definitely new.

OpenAI’s triple play: Codex launch, MCP adoption, and io acquisition

Ok fine, I'm cheating. There were way more than three things that interested me in AI this week. Here are 3 more.

OpenAI launched Codex last week, an agentic coding tool based on their o3 model that runs in isolated cloud environments. You give it a github repo and just start rattling off tasks 'explain this', 'fix this', 'add this' and it will run them all in parallel giving you a stream of pull requests back to the main repo. I've experimented with it a bunch and it's good for simple things, but it
a) speeds up development for professional devs
b) gives a way for non-technical team members to actually ship simple tickets.

So no more being beholden to devs to change the text of a button on your production website. As long as you have solid test suites and CI/CD.

Yesterday OpenAI announced MCP integration for their Responses API - adopting the Model Context Protocol standard originally developed by Anthropic. The alleged usb-c for AI that connects anything with an API to your LLM. This allows OpenAI’s models to connect to external tools and services like Stripe, Shopify, and Twilio with minimal code, creating interoperability with thousands of third-party services. I'm expecting them to release integrations for ChatGPT desktop and web soon and it solidifies MCP as an industry standard.

And then today, OpenAI announced they are acquiring Jony Ive’s design company io for $6.5 billion in an all-stock deal. Ive’s team of 55 (many former Apple designers) will join OpenAI to form a new hardware division with first products expected in 2026. Sam Altman framed it as creating a “new generation of AI-powered computers” that move consumers “beyond screens,” with Jony Ive saying things like he has “a growing sense that everything I have learned over the last 30 years has led me to this moment.” I've heard AI described as Apple's Blackberry moment, which I dunno - but they rarely do anything worth writing about.

Which is interesting and all, and I'm sure the hardware will be fascinating but I just want to celebrate the most weird, bromantic acquisition announcement I have ever seen. Maybe celebrate is an overstatement - I'm not saying it's good, or that they said anything substantial, but wow, I wonder if there is a PR firm somewhere who pitched an idea and walked out of that meeting with a "Oh crap, they went for it".

Cheers,
JV