AI has only two stable futures
Two futures become apparent once frontier AI becomes powerful enough to reshape society. One depends on containing its diffusion. The other depends on adapting to it.
ajfisher
The blog and portfolio of AJFisher - technologist
Latest posts
What happens if this is as good as it gets for AI?
For many routine commercial tasks, model capability may no longer be the constraint. The goal is to close the deployment gap and turn existing AI into realised economic value.
Fail fast, fix faster: Why faster models can beat smarter ones
Transcript and slides from my AI Engineer Melbourne 2026 talk about how, once a model is competent enough, the speed of the iteration loop matters more than the headline intelligence of the model.
Mercury 2 won’t outthink frontier models but diffusion might out-iterate them
The release of Mercury 2 suggests diffusion models may change agent design - not by increasing intelligence, but by making iteration dramatically faster. As iteration gets cheaper, verification becomes the bottleneck, reshaping how software engineering systems are built.
No Mac Mini, no worries: running OpenClaw on a Raspberry Pi.
OpenClaw gives us a taste of a personal AI assistant that can actually do things, but you don't need to buy an expensive Mac Mini or give full access to your digital life for it to be useful. This guide walks through running OpenClaw on an inexpensive Raspberry Pi for long-lived research tasks.
AI, incentives and power in 2026
The AI hype is over, and the industrial reality has begun. Here are some predictions for 2026, from the rise of disposable software to potential clashes over worker rights.
5 signs your organisational capabilities are aspirational
A quick checklist to test whether a capability is real or just aspirational, covering ownership, metrics, roadmap, playbooks, and delivery cadence.
The AI infographic wars intensify with ChatGPT release
ChatGPT's new image model is squaring up against Gemini's recent dominance, and the results are strong enough to turn process docs into instant visual explainers.
A locally running LLM advisory board
A local LLM board of advisors that reuses existing Claude, Gemini, and ChatGPT subscriptions, with models debating and converging on answers together.
Chrome finally adds split view tabs
Split View for tabs arrives in Chrome, making side-by-side work and note-taking far easier than juggling windows.
Bring back the weird, fun and joyful web
As web dev becomes codified and KPI-driven, it risks losing the joy. A call to make something weird, wacky, and fun again.
Gemini 3: initial usage thoughts
Gemini 3 feels tighter, more consistent in long responses, and more willing to pull in relevant images and references than earlier versions.
Strategy is still a human activity
A week of deep strategy work with clients highlights the uniquely human capabilities that we need to keep as businesses transition to using more AI within the enterprise and across their workforces.
Atlas and the AI browser strategy divide
Atlas highlights the strategic focus of browser makers versus AI-native companies with the incumbents prioritising inward integration versus more of an ecosystem play.
Scaling coding agents (without breaking your dev team)
Transcript and slides from my Web Directions AI Engineering 2025 talk about scaling coding agents without breaking your development team.
The Qantas breach proves personal data is toxic waste
Another data breach, this time from Qantas, shows again that the capture and proliferation of personal data across businesses needs a rethink. We need to assume data breaches will happen, let go of the illusion of perfect security, and design accordingly.
Hands on with Gemini CLI
Google's new Gemini CLI arrived, so I had a play with some real dev tasks. At a first glance, it brings google on par with the other players in this space but even with a mega 1M-token context window, it doesn't change the status quo that much.
ChatGPT-4o image generation improvements
ChatGPT-4o's new image capabilities add multi-modal context and iterative edits, bringing it into line with recent Google releases.
When does travel planning get seriously upended by AI?
AI is everywhere in the travel industry - just not where most travellers see it. Chatbots can churn out a basic itinerary, but true AI-native trip planning has a much higher bar to clear. Personalised recommendations and seamless bookings may still feel out of reach, but we might be closer to change than it seems.
After the drought: first shoots emerge for Australian retail
After a tough few years, Australian retail is showing the first signs of stability. Challenges remain, but recent data suggests consumers are more confident, spending is increasing, and the drought may finally be coming to an end.
Mistral 3 Small could unlock powerful local AI use cases
Mistral 3 Small is fast, lightweight, and ideal for local workflows, but it exhibits knowledge compression when compared to bigger frontier models.
Deepseek R1 is capable, but shaped by its geography
Deepseek R1 shows strong reasoning, but its safety controls reflect geopolitical constraints and may leak into adjacent topics.
Nous Research pioneers distributed LLM training
Nous Research is training a 15B model with distributed techniques, pointing to a future where more organisations can train their own LLMs.
Software engineering reimagined for the AI-Native era
AI is redefining the software development process. By viewing development through a lens of problem-solving rather than code creation, engineers can focus on outcomes and value generation - not just lines of code.
Flux, LoRAs, and comicbook experiments
Flux makes image generation easier to prompt, and the community's new LoRAs are already pushing it into playful new directions.
Follow me. I know what I'm doing... (I think)
Stepping suddenly from practice to enterprise leader can be daunting. This talk, given at WDC Leaders 24 explores how to navigate one of the most significant career transitions you'll ever make.
Bias in generative image models is still prevalent
Image models still skew heavily in who they depict, forcing users to over-prompt for diversity and reinforcing bias by default.
Now or never: harnessing growth opportunities in retail
Amidst a wave of layoffs in the tech sector, retail and consumer businesses have a unique chance to leverage cutting-edge technology and talent, creating an opportunity for growth and innovation.
Mixtral 8x7B and model anxiety
Mixtral 8x7B can spiral into anxious feedback loops, a reminder of the quirks baked into models trained on the internet.
Retail numbers are "fine" (but margins less so)
Retail revenues are stabilising, but heavy discounting is pulling sales forward and eroding margins heading into 2024. Now is the time to analyse and adapt.
AI in 2023 - it works well enough now
After attending the Web Directions AI conference, it's clear 2023 brings significant advancements in AI. Yet, as its potential grows, challenges in interaction models and ethical considerations persist.
Dating, AI and digital twins
Snack’s AI dating gimmick is another marker in a broader shift toward digital twins and AI assistants handling everyday tasks.
Generative AI meets VFX workflows for outstanding results
A breakthrough VFX workflow shows how generative AI can amplify human creativity when paired with skilled teams and good tools.
ChatGPT and Generative AI in the enterprise
There is a lot of opportunity for Generative AI from a business technology perspective. Talking with other technology leaders, we discuss what risks and rewards may exist.
ChatGPT and style transfer experiments
ChatGPT style-transfer prompts are a surprisingly fun way to explain complex topics within tight constraints.
Building embodied bots with JavaScript
Explorations of how to build bots with JavaScript that blend virtual and physical worlds, using chat interfaces like Slack to create embodied agents.
Making the move to metalsmith
I moved this site from WP to metalsmith.io in order to get a more pure authoring experience. The result is good, though there were a few gotchas along the way.
The meandering journey of NodeBots Interchange
What started as a plan to simply make components work a bit better with NodeBots turned into a way of managing any firmware using JavaScript.
Datatium - data as material for contextually responsive design
Responsive design is still very screen size focussed where as we should be considering context much more than we do. This talk explains how Datatium helps resolve this.
JavaScript: not just the language of the web
JS is moving out of the browser and into all manner of places. My notes and slides from my talk on this topic at JSFoo 15.
Applying the lessons of mobile dev to IoT
IoT product development has the opportunity to draw on some of the lessons hard won over the years of developing for mobile.
DDD JavaScript IoT workshop
A summary of the various projects completed as part of the JS IoT workshop given at DDD Melbourne 2015.
Building a portable minecraft server
Using a raspberry pi and some batteries to create a portable minecraft server that can be put in a lunch box and taken to the park.
Controlling networked LEDs using a smartwatch
The ESP8266 is a capable device in its own right and can run a simple web server. Couple this with a smartwatch for control and you have some interesting remote control options.
Book launch of Make: JavaScript Robotics
This new book on JavaScript robotics launches today, co-authored by 15 of core NodeBots community members from around the world.
The role of context on ambient technologies
Ambient technologies are very hard to design well. A blinking light for a new email is novel for a few minutes before quickly becoming a distraction. This essay provides a model for thinking about ambient technologies and how a designer may maximise their use.
Building an information radiator
"Information Radiators" - devices that use physical means to relay encoded information are great projects if you want to explore IoT. This post provides a how to example.
A Device API safari
There are many interesting things to be found in the Device API. This talk went on a journey to see what we could find worth playing with.
Should JavaScript devs build real things?
JavaScript is leaving the browser and entering the real world. Is this just a bunch of web devs playing with hardware or are we starting to see a fundamental shift in the direction web technologies are taking?
Book launch of Jump Start Responsive Design
Working on my first book was quite good fun, though hard work, and today after all of that effort it's finally released.