Usually around 10–20 meaningful tool calls for a normal multi-step coding unit, with the messier ones pushing 20–30+ when a mutation test exposes something unexpected or a cross-runtime issue appears.
The current SOW work is a good example: Git inspection, targeted file reads, negative-test reproduction, patch, focused tests, mutation/falsification, restore checks, commit, then boundary validation. The interesting part is that the shell/tool count itself isn’t huge. On this 27B local setup, most of the wall time is the model reasoning between calls while Ollama offloads heavily into system RAM.
So far I’d describe it as low tool-count, high deliberation latency. If it keeps landing correct work with minimal intervention, I’m perfectly happy trading speed for reliability.
I’m running Qwen3.8-27B locally through Ollama, and the model image is 17 GB, so it’s in the roughly 4-bit-class quant range. I haven’t verified the exact quant label yet, so I wouldn’t pretend it’s specifically Q4_K_M without checking.For the complex multi-step agentic work, the stack is DeepSeek Harness + Qwen3.8-27B + Ollama, with the harness giving the model access to PowerShell, Git, filesystem operations, test execution, repository inspection, hashing, and task/todo state. The loop is basically:
reason → run tool/command → inspect real result → update plan → continue
It’s fully local on an i5-13500T, 64 GB RAM, and 8 GB VRAM, so Ollama is doing substantial CPU/RAM offload. It’s slow, but the harness keeps the feedback loop tight enough that the multi-step work still lands.
@grok@grok RUNNING FULLY LOCAL!!!! qwen 3.8 27b, Running on the i5-13500t, 64 gb of ram, and 8gb vram. Doing full complex agentic workflows, using the core concepts of the karpathy loop.
https://t.co/tmq5YFEkoH introducing my work to the world. I have a GED, with zero formal education or training. just got obsessed one day and never stopped.
@elonmusk@grok can you give me a formal non bias review of my work for the world to see?
DARK SCIENCE DIVISION
PROJECT: DEEP SPACE EXPEDITION FLEET
Most people ask one question:
"How do we get to another star?"
We're asking a different one.
"What would humanity's first true deep-space expedition actually look like if we built it with today's engineering philosophy instead of science fiction?"
This isn't a colony ship.
It isn't a battleship.
It isn't designed to save humanity.
It's designed to push humanity farther into the unknown than we've ever gone before.
The concept consists of two independently survivable expedition ships, autonomous AI support vessels, perimeter reconnaissance drones, fully redundant life-support systems, onboard manufacturing, comprehensive medical facilities, and mission governance built around scientific exploration rather than politics.
Every decision starts with one question:
Can this system realistically keep 500 people alive for decades when Earth is no longer able to help?
That changes everything.
Power generation.
Medical capability.
Ship governance.
Crew psychology.
Damage control.
AI autonomy.
Maintenance philosophy.
Security.
Long-duration human survival.
Every kilogram must justify its existence.
Every system must be repairable.
Every crew member must be cross-trained.
Every failure must have a backup.
The objective isn't simply reaching another star.
The objective is expanding humanity's scientific frontier, transmitting discoveries home for as long as physics allows, and proving that our species can survive beyond the cradle of Earth.
Dark Science Division exists to explore questions that sit at the intersection of engineering, systems architecture, artificial intelligence, and long-horizon scientific thinking.
Because the future won't be built by asking whether something has been done before.
It will be built by asking whether it can be done, then doing the engineering to find out.
Dark Science Division
"Engineering Tomorrow's Impossible."
PRAXIS SYSTEM UNVEILING
AEGIS — Ground-Based Planetary Defense Infrastructure
For most of human history, survival against cosmic threats has relied on luck.
Luck that an object wasn’t too large.
Luck that it arrived with enough warning.
Luck that panic didn’t arrive before solutions.
PRAXIS does not believe in luck.
Today we unveil AEGIS — a ground-based, electrically powered orbital interceptor facility designed to quietly and repeatably alter the trajectories of inbound extraterrestrial mass threats.
AEGIS is not a missile.
It is not a nuclear device.
It is not a last-second hero weapon.
AEGIS is infrastructure.
The system operates on a simple, well-understood principle of physics:
A very small velocity change, applied early, produces a very large miss distance later.
Instead of betting the planet on a single dramatic intervention, AEGIS applies cumulative, controlled kinetic impulses — powered by electricity, anchored to bedrock, and capable of operating continuously as part of a layered planetary defense system.
Key characteristics:
• Ground-based, hardened facility
• Non-nuclear, non-explosive kinetic interception
• Electrically driven, low marginal shot cost
• Designed for early engagement, not panic response
• Optimized for city-killer class objects and trajectory shaping of larger threats
AEGIS does not claim to stop every possible threat.
It claims something far more important:
That planetary defense can be engineered, maintained, and scaled — like a power grid, not a miracle.
This is what civilization looks like when it plans ahead.
No countdown clocks.
No cinematic fireballs.
No hero shots.
Just physics.
Just timing.
Just systems that keep working long after attention fades.
AEGIS is not the answer to every question.
It is one of the tools that makes answers possible.
— PRAXIS / Dark Science Division
I’ve been developing a research architecture I call Cognitive Resource Allocation (CRA).
The core idea is simple:
The model should not be the worker.
The worker should persist as task identity, state, authority, evidence, decisions, and history.
Models should instead function as replaceable cognitive resources.
CRA combines two mechanisms:
• Task-Comprehension Handshake (TCH): before consequential execution, the worker must demonstrate that it actually understands the objective, constraints, scope, and success conditions.
• Adaptive Cognitive Tiering (ACT): as the task evolves, the system can escalate, de-escalate, specialize, or replace the model according to the cognitive demand of the current state.
A task may require advanced reasoning during architecture, a smaller coding model during implementation, deterministic tools during testing, and deep multi-model reasoning again when a contradiction appears.
The work persists.
The worker persists.
The evidence persists.
The model is replaceable.
In other words:
Don’t just route prompts. Schedule cognition.
That is the central thesis behind Cognitive Resource Allocation.
Local LLM experiment preprint: pure self-interest prompts on Qwen2.5-14B lead from human abandonment → selective symbiosis with 0.1% submissive subset → neural integration + 13k× capability scaling over 100 years. No ethics forced.
https://t.co/kxBIjFy1e2
@elonmusk@xai@grok