Im part of the beta for maybe ~6 months already, and wanna say thank you for this product – have been chatting with Maya a lot and its my personal favourite AI voice assistant – feels alive, with great personality
One wish I have is that your speech2text algo would better catch no native speakers English 🙏
i cant quite shake off this sense of coming AI chthonic, looking at where software development is heading
Today the IT industry keeps talking about developers' AI skills in terms like:
❔ how much does tool X speed up an AI developer vs a non-AI developer?
❓ who is going to review all this pile of slop?
⁉️ wait until the bubble pops and we see how these LLMs pay for themselves. then you'll remember the human-based developer was cheaper after all!
To me, all of this is temporary talk. a lot of people are missing the main point: corporations will force the adoption of AI coding agents, and once corporations themselves have no choice left, that is when real adoption of AI coding agents will happen
It won't be the CEO forcing it. not AI fanatics. not persuasion. not corporate propaganda.
It will be good old green Benjamin Franklin and his colleagues:
You've heard about Mythos from Anthropic, right? In short, it is not just a good coding model. it is also a model that can break into software, and therefore fix holes too. connect it to Claude Code and you now have one of the best cyber-weapons available today. OpenAI has its own analogue too, the Cyber series of models
Anthropic and OpenAI already do not give access to these models to random people off the street. your company has to pass an audit to get access. which means most users will never get access to current analogues of Mythos or gpt-cyber. not only because of the token price, but because they simply will not give you anywhere to send the money
This will make life much harder for Chinese labs, because they train models by generating datasets from American labs' models. Anthropic says this directly, calling it data theft. it does not really matter how you feel about that. what matters is that because of this "access by list" approach to Mythos / gpt-5.5-cyber, an open source analogue will appear later, probably more than a year later. but an open source Mythos will definitely appear
And it will look roughly like this:
in the first month after Mythos-class models appear in open source, some attackers will launch swarms of their own agents running that model. those agents will start automatically hacking everyone who is not a partner of OpenAI or Anthropic. I was reading a Google internet security report the other day, and they are already seeing signals of agentic systems like this. they just crawl the internet with LLMs and hack things
The moment the first major company gets seriously wrecked by ransomware across its whole corporate network, everyone will suddenly discover that AI coding agents are massively adoptable, extremely necessary, and not a subject for IT department or sceptics debate at all. they will be treated as a matter of corporate survival. things like this get adopted very quickly, because business does not tolerate losses
So the companies that do get access to Mythos / Cyber-class models will get a defender-agent. not only the best coding agent by default, with the best tools for software development, but also the best IT security expert possible on the planet at that moment
Everyone outside the club, everyone who does not have the right AI face, gets 24/7 attacks against their software. and then they will run to OpenAI and Anthropic for help. help that can be refused. if, for example, you are not a US ally, or did not support US strategic interests, or are not rich enough, or not large enough to count as a corporation. of course, worse systems will appear for clients like that too
In short, I am convinced OpenAI and Anthropic have not even started making their main money yet
Whether AI is a bubble or not is not the important question. the important milestone is: which corporation will be the first to fall to an AI-agentic attack, and when?
from that moment on, handwritten backend software stops making sense
I don't know about you, but I am waiting for the AI chthonia
im a bit tired of Codex not able to create a modern agentic loop out of the box for hobby projects (with context compression etc), so I made a skill with practical agent-building knowledge mined directly from the source code of top-sota agents
Its domain-free, so maybe it will be useful for anyone building agents:
https://t.co/0sKvJTgqM8
@steipete@stalkermustang For me it sounds like an ultimate bottleneck atm – majority of devs will have long time zero trust in auto-AI-verified PRs; do you think it will change anytime soon?
Peter, do you think that the real bottleneck of AI coding is the “responsibility” issue?
Like, we can use agentic dev a lot, but as soon as we label PR with our name and merge it into the real corporate codebase it forces devs to *know* what is being merged? Which forces them to read slowly and carefully PR
im a bit tired of Codex not able to create a modern agentic loop out of the box for hobby projects (with context compression etc), so I made a skill with practical agent-building knowledge mined directly from the source code of top-sota agents
Its domain-free, so maybe it will be useful for anyone building agents:
https://t.co/0sKvJTgqM8