Tower Language - Safe by Construction, Expressive by Nature More info coming soon. Talks a LOT about AI in the mean time.
Joined June 2026
- Tweets59
- Following60
- Followers7
- Likes26
This is what I mean by “question engineering”. The best results come from the best framed questions.
While traveling, I asked my agents: what would a modern Hiram Bingham (the explorer who found Machu Picchu) do with today's tools? Their answer: "LiDAR". They found a dataset and spent one night working on it. They had flagged tons of potential structures across Mexico.
12
I let Astra Medium play RuneScape and he's been doing quests!
Here's a Timelapse of Prince Ali Rescue.
Time was about 34 minutes at 1x speed.
What should Astra try next?
95
Stop scrolling.
The AI-assisted hacks making headlines are the tip of the iceberg. Watch what happens as the cost of attacking keeps falling.
I was talking about this three years ago, watching a friend with no software engineering background use AI to build tools for ethical security work and bug bounties. The barrier was already coming down.
Now an attacker using AI tools has extracted data from South Korean financial institutions while leaving their own AI session logs exposed. Sloppy operational security. Real damage.
Sam Altman: “I no longer think we’re going to be able to just patch all the bugs.” He warned that smarter models could uncover hardware-level vulnerabilities beyond our usual threat assumptions.
Anthropic’s deputy CISO, Jason Clinton: “we need to hold on for ~18 months.” That’s his outlook for defenders catching up.
Now think about scale.
Better models mean more capable agents. Cheaper hardware means more copies of today’s models. You don’t need another intelligence breakthrough to run more attacks at once.
Imagine the same attacker going from 10 agents to 1,000. Then 10,000. Now give that operation a nation-state budget.
The damage already happening is serious. Treating it as the ceiling would be a huge mistake.
Same attacker. Vastly greater reach. Buckle up.
If this scares you, use that feeling. Take it as a warning to prepare, not another headline to scroll past.
Sources behind the post:
- South Korean financial-sector breaches and exposed AI session logs: crowdstrike.com/en-us/blog/u…
- Sam Altman’s cybersecurity remarks at Stanford, starting around 19:16: youtube.com/watch?v=anAmTGce…
- Jason Clinton, Anthropic deputy CISO, on the roughly 18-month outlook: teapawt.net/JasonDClinton/status/2…
as bad as things have gotten in the past 4 days, they are going to get a lot worse before they get better. security leaders, never let a crisis go to waste: please go to your boards and ask for help right now. not for ai, for the basics. (we need to hold on for ~18 months.)
89
Stop thinking human->agent.
Start thinking human<->agent.
You prompt it to perform better. Ask it to “preprompt” you before your next interview, meetup or difficult conversation.
The goal isn’t just a more capable agent.
It’s a more capable human-agent team.
67
10+ years in enterprise software, deep into AI, and hiring feels broken.
I wonder how many interviewers have job-hunted lately. Do they know what it’s like now?
I answer for myself, not with AI feeding me answers in real time. But that’s an option now. What are these interviews actually measuring?
54
The skill is shifting.
As models get better, we spend less effort telling them exactly what to do and more effort figuring out the right question to ask
66
Prompt engineering → context engineering → loop engineering → graph engineering → question engineering
Let your app choose the right model, tool, or action in near real-time with Decisions API, now available to all developers in public beta.
The Decisions API makes decisions up to 10x faster than GPT-6 Luna through the Responses API.
89
AI Thought Experiment:
What if sufficiently capable AI red teamers eventually force software to converge?
One web framework. One database. One auth system. One hardened stack that nearly all new software is built on.
Not because it's the best developer experience, but because maintaining thousands of frameworks, libraries, runtimes and dependency combinations becomes impossible to secure against superhuman attackers.
Maybe extreme offensive capability eventually makes software diversity itself an attack surface.
1
74
Codex starts a long-running job. How does it know when to continue?
I made wake-codex to solve that problem.
With wake-codex, your program signals the right chat when it finishes, pauses, or needs attention. Codex picks up from there.
No model polling. No manual “continue.”
github.com/rotcev/wake-codex
68
Being smart is losing its value.
Thinking has become an engineering problem. AI is already eating through it.
Being the smartest human in the room starts to feel like being the best human at chess.
The blackpill: AI does the thinking. Humans don’t have to. Idiocracy is a documentary.
56
AI just searched biological code at a scale humans never could and surfaced what might be an undocumented biological API.
Now we get to figure out what it does.
And more importantly, whether we can program it.
BREAKING: Anthropic’s new wet lab produces its first discovery, with nearly 1,000 Claude agents autonomously uncovering a novel enzyme system after 21 hours of work.
116
Codex's missing feature is a "Full web access" mode - a happy medium between "Approve for me" and "Full access".
62
This is kind of insane.
You can now give Codex an entire Linux desktop to operate.
Not a browser sandbox.
Not a purpose-built API.
Your actual remote desktop.
I built xRDP Agent Harness, an MCP server that gives Codex screenshots, mouse, keyboard, and control of your existing desktop session.
The point is simple: you can now automate software and workflows that were previously difficult or impossible to automate because they had no API.
If you can do it through the GUI, Codex can do it too.
The GUI is now an API.
2
2
139
I hope this helps someone. It solved a pretty big problem for me.
github.com/rotcev/xrdp-agent…
45
🚨 OpenAI may have just quietly released one of the most important APIs of the AI era.
Their Navier–Stokes breakthrough came from ~10,000 coordinated agents exploring approaches in parallel and feeding useful discoveries back into the swarm.
Now the Agents API gives developers durable agents, sandboxes, tools, persistent sessions, orchestration, recovery, and subagent delegation.
OpenAI isn't just selling access to intelligence anymore.
They're exposing the primitive for scaling intelligence horizontally.
Those willing to spend for compute are likely to have massive breakthroughs across the board.
Buckle up.
36
Recurrent attention and recurrent workspaces offer two complementary ways to extend transformers.
Attention revisits the evidence.
The workspace works through what it finds.
The interesting experiment is teaching the model when to stop gathering information, and when to stop refining its internal state and produce an answer.
Rather than taking a fixed number of passes, the model could learn how much time to spend gathering evidence and how much to spend working through it before answering, depending on the question.
35