The Open Eyed Man Turns One.
After a year of revolutionary AI progress, the accessibility gap persists. Looking back at five landmark events in the past 12 months and forward to what's ahead.
It is one year since The Open Eyed Man was born — 51 posts with one week's break at Xmas. The pace of AI news in that period has been astonishing, but the "accessibility through AGI" challenge I laid down in my first post remains a distant dream.
Specifically, living as a blind man in an increasingly visual world, the biggest challenges to independent living are still writ large. More and more consumer devices have touch screen controls, the world is drowning in an outpouring of inaccessible online content and my trusty white cane still has the best shot at getting me across the car park undamaged.
I do remain bullish that AI can ultimately transform accessibility. However, the AI which achieves that transformation will need:
- real-time interpretation of the physical world
- ultra-low-latency conversational voice.
And both these capabilities may be several years away. In that light, from now on my Open Eyed Man posts will be ad hoc rather than weekly. Future posts will be triggered by extraordinary events either in AI progress or in accessibility.
So, now wrapping up the weekly drumbeat, here are 5 of the most significant AI happenings in the past 12 months.
The Past Year: Five Major AI Happenings
Coding Is a Solved Problem
Three years after ChatGPT's explosive entrance into our lives, November 2025 saw another historic moment, as AI coding agents came of age.
With a clear specification and unambiguous testing criteria, coding agents can now churn away for hours or even days to produce something that works pretty well, which would previously have taken several humans several months.
This has profound implications for the software sector which has dominated the global stock markets for the past 20 years. If AI can build a replica of your product in days, what is your justification for charging outrageous monthly subscriptions, and how do you convince your users not to switch to a cheaper clone?
As an aside, AI has demonstrated its ability to replicate software without breaching any copyright using tried and tested clean room techniques.
Image Generation Is Now Reliable and Impressive
Cover images for my first few posts last year frequently sported missing digits, misplaced limbs or glaring errors in scene construction.
Now the images usually just work first time, like the one on this post, created by Google's Nano Banana Pro, also first released in November 2025. As an aside, the detailed prompt for Nano Banana was actually generated by Claude following a Skill (a sort of recipe) which Claude wrote for itself that produces visually consistent images from week to week.
Personal Agents Are Suddenly a Reality
In January of this year Moltbook went viral after its release as a social media platform for AI agents to interact with one another.
While the idea was entertaining, its real impact was publicity for Open Claw, the software behind each Moltbook participant.
Because Open Claw is open source and has very few guard rails it is addictively powerful, allowing users to delegate unreasonable responsibility to their agents. Users readily gave Open Claw access to sensitive data and bank accounts they would certainly not entrust to their teenage children.
In the following weeks a plethora of similar agents or 'claws' were born, and the claw architecture is now the de facto bedrock for personal agents running on our phones, our laptops or on dedicated Mac Minis/Raspberry Pis where they can do less damage if they go off the rails.
Cyber Risks Are a Real and Present Danger
On the one hand, AI systems are inherently hard to secure against cyber attacks. Because Large Language Models work with natural text it is dangerously easy to deceive them into doing bad things like exposing data. Even tricks as basic as white text on a white background can inject an unseen malicious instruction straight into the prompt as if it had come from the user.
The June 2025 EchoLeak critical breach of Microsoft Office's AI co-pilot caused consternation amongst the tens of thousands of enterprises that use Microsoft Office throughout their organisation.
On the other hand, AI can itself crawl corporate websites and other infrastructure to seek out and exploit vulnerabilities just like traditional human hackers.
In particular, earlier this year, Anthropic's Mythos model discovered dozens of such security flaws, several of which have lain dormant on millions of systems for a staggeringly long time: 15-27 years.
Mythos is 5 times more resource hungry than any publicly available model. Anthropic also say it is too dangerous for public release. But it seems inevitable that a similarly capable open source model will be available before too long. And someone will use that model to compromise critical systems rather than fixing them.
Google Search Is Becoming AI Native
Last week, in a slew of AI announcements at their annual developer conference, Google majored on the biggest overhaul of Search in 25 years.
Now, with the AI Mode button, users get ChatGPT/Claude style responses without having a subscription and without concerns around burning tokens. The service is superficially free and totally ubiquitous.
This could cement Google as the dominant provider of general purpose consumer grade Large Language Model chat. It could also totally transform the online search and ultimately online purchase experience.
And here are my predictions for the next 12 months.
The Next 12 Months: Five Key Predictions
No AGI
AGI by any reasonable definition will not happen in the next 12 months. This deep dive by Robert Wiblin from February this year tries to make sense of wildly fluctuating projections and concludes we may still be 10 years from strong AGI.
And if (as I believe it should) AGI includes human level physical world interpretation, then it is likely still decades away. So my gauntlet will lie in the carpark for a few years yet.
No Capable Domestic Robots
General purpose domestic humanoid robots will not become a thing in the next 12 months.
High value industrial tasks will see advanced robot adoption, but these robots will be predominantly teleoperated, not fully autonomous.
For humanoid robots to be safe and general purpose, we still need significant progress in scene interpretation. This will likely require a new AI architecture, distinct from brute force Large Language Model next token prediction.
In 2027, OpenAI, Anthropic and Google will likely divert significant resources to world models better suited to robot brains.
But, for the next little while, folding the laundry or loading the dishwasher are beyond the reach of domestic robots.
Jobs
Job losses to AI, particularly entry level white collar ones, will become an unambiguous reality in the coming year.
Without corresponding new AI related jobs emerging, public antipathy to AI will consequently become a significant issue for national politicians.
Defence
Enormous volumes of low cost drones are now existentially critical infrastructure for national defence.
The next 12 months will see these drones evolve from predominantly individual devices into intelligent co-operating swarms.
At the same time AI controlled anti-swarm defensive systems will be deployed in a terrifying race for battlefield supremacy.
Personal Agents Go Mainstream
Personal agents with life long memory will become mainstream in the next 12 months.
They will increasingly be the way we communicate, consume information and make purchases.
This will drive a multi-year shift in the structure of the world wide web. For all websites, apis or WebMCP support will be table stakes. Websites that do not optimise for agent access will lose ground to those that do.
In parallel, specific hardware improvements will allow phones to run a new generation of very capable, less resource heavy, AI models.
In 2 or 3 years' time, your personal agent and the AI model it uses will give you 'AI in your pocket'. It will sometimes reach out to the cloud for particular heavy lifting, such as video generation, for searching, for purchasing and for communicating, but AI will no longer feel like a special thing. It will just be the way you get everything done in your digital life.
Over the summer I have 2 sight related projects. One is with major technology companies on my favourite topic of AI Assistive Tech. The other is an extraordinarily ambitious long term initiative to restore sight in the world's poorest regions. I will say more when I can.
Please do stay subscribed. I am sure there will be lots to talk about on the AI front, and I'm hopeful there will also be material good news to report on the accessibility front.