Uploaded October 2025 | Updated September 2026, 31 minutes ago
Fine-tuning may have just leveled up. đ
OpenAIâs former CTO and VP of Applied Research launched a new company, Thinking Machines, and their first product is a game-changer: Tinker.
Instead of the old âupload your data and weâll fine-tune itâ approach, Tinker gives you Python-level control over algorithms, data, and training loops, while handling the messy GPU infra in the background. That means you keep ~90% of the creative control that actually matters (data curation, loss functions, optimization tweaks) while offloading the complexity youâd rather not touch (distributed training, infra headaches).
This shift could make fine-tuning practical for many more teams. Instead of writing endless few-shot prompts for giant models, we might see faster, cheaper workflows by finetuning smaller LLMs with Tinker. Itâs not about âstylingâ modelsâitâs about narrowing scope, boosting robustness, and owning your data pipeline.
Of course, the hardest part remains: building a high-quality dataset and running proper evaluations. But the entry barrier to real fine-tuning? Just dropped.
Iâm Louis-François, PhD dropout, now CTO & co-founder at Towards AI. Follow me for tomorrowâs no-BS AI roundup đ
#tinker #MachineLearning #LLMs #short
Fine-tuning may have just leveled up. đ
OpenAIâs former CTO and VP of Applied Research launched a new company, Thinking Machines, and their first product is a game-changer: Tinker.
Instead of the old âupload your data and weâll fine-tune itâ approach, Tinker gives you Python-level control over algorithms, data, and training loops, while handling the messy GPU infra in the background. That means you keep ~90% of the creative control that actually matters (data curation, loss functions, optimization tweaks) while offloading the complexity youâd rather not touch (distributed training, infra headaches).
This shift could make fine-tuning practical for many more teams. Instead of writing endless few-shot prompts for giant models, we might see faster, cheaper workflows by finetuning smaller LLMs with Tinker. Itâs not about âstylingâ modelsâitâs about narrowing scope, boosting robustness, and owning your data pipeline.
Of course, the hardest part remains: building a high-quality dataset and running proper evaluations. But the entry barrier to real fine-tuning? Just dropped.
Iâm Louis-François, PhD dropout, now CTO & co-founder at Towards AI. Follow me for tomorrowâs no-BS AI roundup đ
#tinker #MachineLearning #LLMs #short
