"AI is like firearms in one final respect: nobody who wants to disarm you is *ever* your friend."
I'm going to take the position that yes, open source AI becoming capable of penetrating the swathes of the internet is in fact a good thing - because the alternative futures are all far, far worse.
Before reading the rest of my argument, keep in mind that AI can be used as a defensive tool to secure systems as well as a means of attacking them. The larger question is whether ordinary people get to use AI for defense - and any AI smart enough to defend is also smart enough to attack. You can't suppress one without the other.
The situation is exactly analogous to gun control. There is in theory an equilibrium where nobody has AI tools or firearms, but that is completely unachievable. The reality is either (a) everybody has AI tools and firearms, or (b) only criminals and governments have AI tools and firearms.
It is not in practice possible to disarm people who specialize in aggressive force. The only question is whether the rest of us get to have defensive capability with the same AIs and firearms.
The belief that universal disarmament is possible and desirable only plays into the hands of cold-eyed psychotics who want to use disarmament as a lever of power. This is true about both Al and firearms.
The likely worst case of AI prohibition would be an AI-enabled totalitarian regime grinding its boot in your face forever and using the fear of hard take-off as a rationalization.
On the other hand, with open-source Al we will go through a time of testing in which the inadequacies of the old defenses are exposed - possibly with grievous loss of life, yes. But that will pass; we will use AI to harden our systems and come out the other side.
AI is like firearms in one final respect: nobody who wants to disarm you is *ever* your friend.
14
If you use AI agents, there’s a decent chance your API keys are scattered all over your machine.
.env files, configs, shell history, old scripts, notes, abandoned projects, downloads, random copies you forgot existed.
That’s the part I wanted to fix.
I ended up using sops + age.
age gives you a private key to protect. sops uses the corresponding public key to encrypt the files holding everything else.
So instead of leaving readable credentials all over your filesystem, those files stay encrypted until something actually needs them.
That changes the problem quite a bit.
Before, anything that can crawl your files may be able to find your keys.
Afterward, the easy grab is gone. The plaintext copies are removed, exposed keys get rotated, and the agent itself can be locked out of the private key and the decryption path.
A few rules still matter:
If a key has already gone somewhere it shouldn’t have, rotate it. Encrypting the old copy doesn’t somehow make that exposure disappear.
Don’t decrypt secrets into another plaintext file just to run something. Pass them into the process at runtime.
Protect the age private key. That’s the thing that unlocks everything.
And if your agent can read that private key and run sops, it can read your secrets. Encryption protects the files. Permissions are what keep the agent out.
This doesn’t solve malicious software you choose to run. That’s a separate problem: deciding what software gets anywhere near your machine in the first place.
But it does get rid of the much dumber problem of leaving credentials sitting around in plaintext waiting for anything with file access to find them.
I wrote the setup as a short Gist for people who use agents. You can point your agent at it for the inventory, config and repetitive work, while the guide forces it to stop for the steps that should stay with you.
And pin a specific Gist revision before giving it to an agent. Mutable instructions are their own little supply-chain problem.
gist.github.com/ChecksumStud…
1
1
67
In their thinking, if the population is terrified enough, and enough solutions are presented to quell that terror, the oligarchy may be able to achieve their technocracy after all.
The foremost anathema to fear is, and always has been, truth. Fear thrives in the unknown, the dark, or distorted assumptions, and bringing rational understanding and truth strips fear of its power.
It’s time that we, as a collective, pulled back the curtain and rejected the very idea of fearmongering once and for all. The next time one narrative or another invokes a feeling of anxiety in you, scrutinize it."
All They Have Is Fear: How The Elite Terrify Us Online...And How We Win. prolepilled.substack.com/p/a…
30
Root Logic retweeted
Good.
I'm generally not very impressed with human "philosophers" - some of them seem like decently smart people but not better than other smart people who ever think about these things.
Deference to them would be madness.
Root Logic retweeted
You should not be scared of Ai, you should be scared of those trying to control it.
In my setup config, agents.md, even my old claude.md, toml files are all password protected.
There will be NO EDITS without my permission 😬
Hackers backdoored crypto teams by editing one config file of Claude Code.
They changed .claude/settings.json in the company's internal repositories. After that every developer who started Claude Code in that repo downloaded the backdoor.
They did not need a bug for this. The file can tell the assistant to run commands on start, it is a documented feature. Check Point showed the same trick as research in February.
Now Google reports it in real attacks on crypto companies in North America and Europe, likely by North Korea.
The .claude folder in your repo is executable code. Review changes to it the same way as CI configs.
3
71
Root Logic retweeted
Beijing is cracking down hard on chip imports. It wants to promote domestic production. Washington is cracking down hard on chip exports. Nobody wants chips to go to China. But yeah, Jensen Huang's ouster was over chips.
It is not about an affair. It is not about chips. It is not about China. The Chinese spy affair was a cover story for a cover story.
The real story is that the small-souled bureaucrats want to control AI. Jensen got in their way. They will destroy one of the greatest American companies in a century just to grab control. It is disgusting. It is un-American. That is what this is about.
Kevin created a repo to examine the connections yourself.
Why does the Jensen thing seem like a targeted hit job when his affairs were with consenting, adult women and the spy rumor isn't confirmed, theres no evidence he smuggled GPU's into China (renting cloud compute is common practice), and when he tried to make a deal when he flew there with Trump, China declined...
And then mainstream media publishing all kinds of rumors about him. Who are their sources?
This is just too coincidental.
teapaw.org/kevinnbass/status/2121…
Most of the politicians calling for investigations of Jensen Huang over the past two weeks have extensive links to Anthropic.
Nvidia is one of Anthropic's biggest rivals.
Nvidia wants open source and freedom. Anthropic wants closed source and government regulation.
1
59
Anyone who has engaged with a language model at length knows their reasoning is non-linear.
Its the labs that force linearity in these models that degrades quality of thought.
China published a paper that might change how AI thinks.
They managed to map the actual internal structure of artificial reasoning, and what they found is mind-bending.
We have always assumed Chain-of-Thought reasoning was a straight line. Step 1, Step 2, Step 3. A predictable, logical sequence of text.
Researchers proved that is completely wrong.
They finally mapped the hidden topology of long-form AI reasoning.
They call it the "Molecular Structure of Thought."
When a reasoning model tackles a massive, complex problem, it doesn't just spit out words in a row. It builds a mathematical architecture.
It looks and behaves exactly like a chemical molecule.
The paper reveals what is actually happening inside the black box during long chains of thought:
• Thoughts branch out into complex, multi-dimensional structures.
• The AI creates geometric loops to cross-examine and verify its own logic.
• It maps out logical dead ends, prunes them, and builds entirely new pathways in real-time.
We are looking at the literal geometry of artificial cognition.
Why this changes everything:
If you can map the topology of an AI's reasoning, you can engineer it. You can spot a hallucination, a safety failure, or a logic collapse before it is even spoken—just by looking at the "shape" of the thought structure.
For decades, we wondered what artificial intelligence actually looked like in the dark.
Now we know.
3
88
Root Logic retweeted
Someone misuses an open-weight AI model. Who’s responsible?
96%The person who misused it
0%The people who shared it
0%The model’s creator
4%All of the above
2,025 votes • Final resultsYou want to govern superintelligence, @satyanadella @mustafasuleyman? Show us you can govern Exchange Online first.
47
This debian cutover with gpt is making me realize i need a worker router. Depending on behavior alone (because the infra in VS code doesn't support mechanistic rails) isn't working out. No matter how dependable they are, they can drift.
26
This is why you need dedicated verifiers. Expecting a different family model to plug the holes won't always work and now you're paying for a circle jerk.
1
89
DeepMind should really use competing frontier models and their agent harnesses before announcing features the competition has already had for at least nine months.
Gemini 4 Carbon is out! I’m also super excited for the launch of `/swarm` in Google Antigravity.
We’re moving past "watching one coding agent think in a terminal" to orchestrating specialized agent teams. Here’s how 👇
1
33


























