bigsk1

Code Chemist 🛸

[bigsk1 @bigsk1_com]

https://t.co/syFbEbqh5d

[bigsk1 @bigsk1_com]

Personal self-hosted AI assistant. Voice control, 80+ tools, and autonomous coding. Cloud providers or fully local with Ollama. UI, CLI, TUI. Deterministic (code-driven) orchestration! https://t.co/8iNXjr9D9F

[bigsk1 @bigsk1_com]

@xai grok 4.20 (grok-4.20-non-reasoning-latest) is way to tool happy and even provided with my native tool results keeps trying to call the same tool again. Sticking with 4.1 for now.

Add a user-controllable flag (e.g. pipeline_mode: true, single_agent: true, or tool_trust_level: high) that tells the system "this is an external agent orchestrator with its own RAG/tools — minimize internal re-verification and agent debate. Prefer early exit after one successful tool call on direct queries."

[bigsk1 @bigsk1_com]

The LLM needs to SEE the actual text to provide specific feedback like:
- "The system prompt says X but this conflicts with Y"
- "Tool description for `get_time` says 'returns current time' but doesn't mention timezone"
- "Intelligence insight #24 was misleading because..."

### Tool Description Detection

The system automatically includes descriptions for:
1. Tools that were actually used
2. Tools that SHOULD have been used (based on query keywords)

For example, a query with "time" will include the `get_time` description even if it wasn't used, so the LLM can say "This tool should have been used because..."

### The Feedback Prompt

```
You just completed a task as a voice assistant. Now provide HONEST FEEDBACK...

=== YOUR TASK WAS ===
User Query: What time is it?

=== RESULT ===
Success: Yes
Response: It's 2:15 PM on Sunday
Tools Used: get_time

=== WHAT YOU WERE GIVEN ===
System Prompt Summary:
- Memory-first rules requiring semantic_recall before external tools
- 25-word voice output limit
- Tool descriptions for 50 tools
- Intelligence insights (learned patterns, known failures)
- Auto-context with recent conversation history

Tools Available: 50
Intelligence Insights: Enabled
Auto-Context: Enabled (window=3, minutes=10)
Response Style: auto

=== PROVIDE FEEDBACK ===
Rate your experience (1-5)...
```
then

EVOLUTION PIPELINE

[Feedback Logs]
- ratings + suggestions
- logs/feedback/feedback-*.jsonl
|
| Trigger: 5+ low ratings (<6) on same component
v
[Analyze Patterns]
- group by component (system_prompt, tool:xyz)
- extract repeated complaints/suggestions
|
v
[Generate Variants]
- FEEDBACK_MODEL proposes improvements
- produce 2–3 candidate versions
|
v
[Verify Candidates]
1) syntax validation (JSON parseable / Python runs)
2) semantic checks (required fields present)
3) synthetic query tests
|
v
[A/B Test] (optional)
- random 50/50 split for N interactions
- compare avg ratings
|
v
[Promote + Archive]
- winner becomes active
- old version archived w/ audit trail

[bigsk1 @bigsk1_com]

every miss step is a learning event for self improvement, I built in reflection step that uses a feedback llm that produces an insight. The most relevant insights top 3 are presented back to the LLM when the right conditions are met , query or tool being used. So based on previous experiences it helps next time around. I have also built in a tool rag + ghost tools ( tools which are always present no matter the similarity ) to not flood context window. I also have a evolution process based on bad feedback thresholds, over certain time frame evolution is triggered and improved tool descriptions and system prompt is improved version controlled.

[bigsk1 @bigsk1_com]

@cb_doge @grok I want to wash my car, and the car wash is 50m away. Should I call an uber, drive or call a coworker because I also work at the car wash?

[bigsk1 @bigsk1_com]

no worries easy fix, yes it is custom.

{
"$schema": "https://t.co/1sf1aaQtEL",
"provider": {
"ollama": {
"npm": "@ai-sdk/openai-compatible",
"name": "Ollama (local)",
"options": {
"baseURL": "http://192.168.11.111:11434/v1"
},
"models": {
"mistral-nemo": {
"name": "Mistral Nemo"
}
}
},
"openai": {
"npm": "@ai-sdk/openai",
"name": "OpenAI",
"options": {
"apiKey": "{env:OPENAI_API_KEY}"
},
"models": {
"gpt-4o": {
"name": "GPT-4o"
}
}
},
"anthropic": {
"npm": "@ai-sdk/anthropic",
"name": "Anthropic",
"options": {
"apiKey": "{env:ANTHROPIC_API_KEY}"
},
"models": {
"claude-sonnet-4-20250514": {
"name": "Claude Sonnet 4"
},
"claude-sonnet-4-5-20250929": {
"name": "Claude Sonnet 4.5"
}
}
}
},
"disabled_providers": ["cloudflare-workers", "amazon-bedrock", "nvidia", "workers-ai", "bedrock", "huggingface"],
"autoupdate": true,
"permission": {
"edit": "ask",
"bash": "ask"
}
}

[bigsk1 @bigsk1_com]

@parthsareen @ollama ollama --version
ollama version is 0.16.0

[bigsk1 @bigsk1_com]

grok imagine

[bigsk1 @bigsk1_com]

@elonmusk @grok yep using api works great

[bigsk1 @bigsk1_com]

plus I use:

memoryFlush.softThresholdTokens: 5000
Triggers flush earlier (more safety margin)

memoryFlush.systemPrompt
Custom prompt for better memory saves

hybrid.enabled: true
Vector + keyword search combined

cache.maxEntries: 75000
Larger embedding cache

batch.enabled: true
Faster/cheaper bulk indexing

https://t.co/ucBHIGcNHz: true
Real-time file watching

[Marius Avram @securityshell]

Holy shit… the exploitation of CVE-2025-55182 has reached a new level. There’s now a publicly available Chrome extension on GitHub that automatically scans for and exploits vulnerable sites as you browse. Absolutely wild. 🤦‍♂️

[Bitcoin Archive @BitcoinArchive]

JUST IN: 🇺🇸 U.S. government seeks maximum 5-year sentence for Samourai Wallet developers

This isn’t right. Code is speech.

Chilling precedent for open-source devs

[bigsk1 @bigsk1_com]

@AutismCapital oh ya newsrover was fun also

[Angry Tom @AngryTomtweets]

This is wild...

The first-ever text-to-film AI agent is here.

It can automatically generate an entire film, from script and storyboard to consistent characters, video, voice, lip-sync, LoRA, and music.

Here's how it works: (step-by-step tutorial)

[bigsk1 @bigsk1_com]

@ItsTheEnforcer Never seen someone that just wants war so bad, maybe job security?

[bigsk1 @bigsk1_com]

@ItsTheEnforcer Job security