SF Compute launches low-cost training product as Mole CLI ships reliability-first update

Key Takeaways

1. SF Compute launched an automated research product for low-cost model training, with fair queues and MCP-based recovery. — via 1
2. Tw93's Mole CLI ships a reliability-first update, and a Codex-driven workflow using it cleaned 332GB without touching photos. — via 1 2
3. A 30B model trained on board game 1830 for about $2-3k cut hallucinations and improved citation fidelity. — via 1
4. levelsio's self-tracking data links private training and high-intensity days to +20% productivity; sickness, travel, sunbathing, and coworking correlate negatively. — via 1 2
5. Tibo opened 100 founder-community spots after the previous round sold out in 36 hours. — via 1

1. AI Models, Infrastructure, and Coding Workflows

1. SF Compute introduced a new automated research product that lets users train models at low cost. It uses a fair queue to avoid resource-hoarding, does not require long-term contracts, and includes an MCP-based ticket system so agents can automatically resume training jobs after platform failures. — via 1
2. Ben Tossell reported training a 30B model on the board game 1830 for roughly $2,000-3,000. Validation with GoodfireAI showed the model did not learn new financial concepts but significantly reduced fabricated facts and improved citation reliability. — via 1
3. On AI coding, levelsio argues against letting an agent rebuild a project in one massive Gauntlet Loop-style pass. He says the result tends to be messy and inefficient, often wasting around $500, and prefers directing AI step by step per function or object. — via 1

2. Developer Tools and Automation

1. Mole CLI's new version focuses on reliability: refusals come with explanations and fix suggestions, dry run fully simulates changes, cleanup boundaries are stricter, slow scans get hard time limits, and disk analysis now supports Parallels virtual machines. — via 1
2. A Korean creator built an automated workflow around Mole with Codex that runs daily at 9am and generates cleanup reports. It removed 332GB of workspace-tree junk without touching any photos; the CLI stays free, with a more complete Mac app available separately. — via 1

3. Data-Driven Productivity and Health

1. levelsio shared correlations from his own tracked work and health data. Days with a private trainer and high training intensity were the most productive, about +20%, even after including training and shower time; coworking correlated with -7%, though he still values it socially. Long/short flights, sunbathing, and especially sickness were negative factors, and he cautioned that most results are directional rather than fully significant. — via 1 2
2. levelsio also emphasized that personalized health analysis is valuable because individual bodies and minds respond differently; another person's optimal setup may not fit you. — via 1

4. Community and Maker Economy

1. Tibo opened 100 new spots in his founder community after the previous round sold out in 36 hours. The community includes live AMAs, business model breakdowns, and weekly accountability groups; access is limited to 3 days on a first-come basis. — via 1
2. Nathan Barry highlighted LaylaPomper's pricing heuristic: if a sales close rate exceeds 50%, the price is likely too low, and her own target is 20%. It reframes high close rates as a sign to raise prices rather than celebrate. — via 1