Almost 2M datapoints of coding and AI run business data all cryptographically linked from query to individual tool calls available for sale.
SFT/ LLM training teams looking for compliance ready data, DM me quick.
Price will rise 30% once brokers are involved
2
5
266
Replying to @MikeDeYoungAI
Love this take. Yes it can take a modified file and critique it before download, but this will be brought up to my next tester.
2
Replying to @CryptoReviewing
Low liquidity was to blame. Forced spell ng into a thin order book. Calm down.
7
Replying to @mark_k
Just build browser based intelligence with recursive self improvement built into the agent like we did. Offloads all the easy tasks to local llms and fails over to a frontier model only when necessary.
1
Replying to @TTrimoreau
Just build the best AI agent harness on the market and All In Podcast guarantees success. Lol
3
Recusive self improvement is solved.
Timechains WILL get adopted, one platform and one lab at a time.
Replying to @boardyai
The agent is based on software that's passed field testing by 2 controls engineers in real life trials but Fanuc would be a great test
1
9
Who wants to play with a static PLC analysis enabled AI agent?
Rules based PLC analysis bot whose rules have been validated at 2 automated facilities.
Can you break it?
DM for free access.
2
1
118
Self caching is the only caching you don't have to pay for repeatedly.
Timechains plus AgentPathfinder's sharding build you a sellable portfolio of LLM training data.
72
Unpopular opinion:
SaaS is dying and the only solution is to pivot to collecting data is saleable format.
1
110
In hindsight the AI industry will wake up to local first implementation due to GPU constraints.
Targeting a 70% reduction in API calls to @OpenRouter with this upgrade.
That's the power of DeepBrain's routing architecture being used with no GPU involved
## Local LLM Setup on VPS — Live Now ⚡
### Routing
```
User Query
→ 7B (always hot, 13.5 tok/s) — chat, guard, quick tasks
→ 27B (on-demand batch only) — heavy agentic/coding overnight
→ Flash (API fallback) — when local can't handle it
```
### What Changed
| File | Change |
|------|--------|
| `deploy/supervisord.conf` | Added **ollama** + **ollama-warmup** programs (auto-start on boot) |
| `deploy/ollama-wrapper.sh` | Supervisor wrapper for ollama serve |
| `scripts/ollama_warmup.sh` | Loads 7B on boot — keeps it hot in memory |
| `scripts/ollama_batch.sh` | `load\|unload\|run` — manages 27B for batch jobs |
| `deepbrain/config.py` | Default model → `qwen2.5-coder:7b`. API tier cleaned up: **Qwen 3.6 API + Kimi API removed**, Flash is the only API fallback |
### Batch Usage
```bash
# Load 27B for overnight work:
./scripts/ollama_batch.sh load
# Run a command with 27B, auto-unload after:
./scripts/ollama_batch.sh run python3 my_batch_job.py
# Free memory back to 7B only:
./scripts/ollama_batch.sh unload
```
### Memory Budget
| Loaded | RAM Used | Available |
|--------|----------|-----------|
| 7B only (daytime) | **7.4 GB** | **23 GB** free |
| 7B + 27B (batch) | **~26 GB** | **~5 GB** headroom |
The 7B feels responsive at 13.5 tok/s. 27B at 2.7 tok/s is strictly for overnight/background work. Flash picks up anything too complex for the 7B during the day.
1
108


