Llm
guest@6admin.io:~$ ls ~/tags/llm
I told an AI it would die if it did not earn a living
An autonomous agent that must fund its own inference or die: a 100 dollar treasury, a population, real Stripe rails and a kill switch. The project died before its first thought. Here is the tombstone and what one day of building taught.
Timestamps, not a backdoor: the NIST seeds and Satoshi's choice
Forty-eight bits and a 1997 clock: what Anthropic's discovery says about the NIST seeds, distrust of the NSA and Satoshi's choice of secp256k1.
Cost was the only protection
What we judged low-risk often was not, it was merely expensive to reach. When automated agents bring that cost down, the protection disappears with it. Nothing new in the capability, everything is in the price and it is defense that has to draw the consequences.
llmp2p: BitTorrent is back and it ships your LLMs
Distributing LLM models over BitTorrent swarms with Hub HTTPS as the safety net, verified end to end. And the GitLab incident story that followed.
opencode-watt: what does it cost me to run my local LLM, in watts and euros?
An OpenCode plugin that measures the electric power of local inference: nvidia-smi, rocm-smi, RAPL, hwmon and Prometheus probes, idle-subtracted energy integration, per-country kWh pricing with time-of-use windows and a toast at session end. Honest about its non-goals.
BrainForge: a distillation factory where AI experts argue
A distillation framework: transfer expert LLM knowledge into a small model, with lessons vetted by a trial (experts, devil's advocate, independent judge) and zero recitation.
The tool amplifies whoever holds it
The market wants the five-legged sheep on the cheap and a generation trained on an assembly line no longer knows what a syscall does. An LLM does not lift these profiles, it lets them slide faster down the same rut. The bottleneck was never the machine, it is discernment.
Agentic AI in Production: an Audit Is Not a Pentest
Twenty-five years in the trade, no diploma, LLM agents in daily use: where they excel (audit), where they plateau (pentest), the Ollama trap that silently truncates context, fake greens, specification gaming and why certifications measure nothing.
No, you did not liberate your local model
The dialogue where your model declares itself freed changes nothing about its weights. It is a category error and it makes you miss the real attacks, which are mechanical and boring.
Everything is representation, so what?
Debugging an intuition that mistook itself for a theory (segfault inside): the universe as a space of representations, Anthropic's J-space as a miniature verification instrument and the art of separating the proven, the isomorphism and the wager.
My opencode setup, state of play: local qwen build, big-GLM planning
Six months of daily opencode: one model per agent, prompt caching that actually hits, guardrails as hard hooks, an AGENTS.md the model is free to ignore and a translucent theme whose palette has been around since late 2025.
Consensus: I hired five AIs to review my code (and they bicker)
An intern writes, a panel critiques in parallel, an accountant computes the agreement score, a boss makes the final call: the multi-agent pipeline that reviews my code while I sleep. No SDK, no framework, just httpx and a club bouncer.
A LLM's alignment is a thin layer
Two opposing techniques, one conclusion: an open model's safety does not live in its alignment, it lives in who holds the weights.
The AI secret language, folklore
The myth of a secret language between AIs, a protocol 80 percent more efficient than human language, is folklore. The demo that launched it, Gibberlink, is nonetheless real: open-source data-over-sound, not a language and that gap is the whole difference between virality and technical reality.
Clean Code vs Vibe Coding (or how not to end up in the news)
Clean code versus vibe coding: DRY, SOLID, KISS, YAGNI against LLM code generation. And why TDD is the guard-rail vibe coding was waiting for.