OpenAI announced reinforcement finetuning on 2024-12-06, which makes it easy to create expert models in specific domains with little training data.
Mostly true True in substance, wrong in a detail.
Factual: Statements about the present or the past, checked against the record: did they hold up?
Said for an organisation they lead or work for; it counts on both records.
What would count as true or false
Judged as of 2024-12-06. True if OpenAI officially announced reinforcement fine-tuning (RFT) on 2024-12-06 (e.g. day 2 of '12 Days of OpenAI') and presented it as a way to customise models into domain-specific experts using small datasets (e.g. dozens to thousands of examples). Mostly true if it was announced within a day of that date, or a detail of the description differs slightly (e.g. at launch it was available only as a limited research preview). Mostly false if it was announced only as a much vaguer plan, or on a materially different date. False if OpenAI made no such announcement.
Set before the research began.
Evidence (4)
-
E1 OpenAI’s 12 days of ‘ship-mas’: all the new announcements The Verge · 5 December 2024
“In a post after the livestream announcement, CEO Sam Altman said this will make it “really easy to create expert models in specific domains with very little training data.””
-
E2 12 days of OpenAI: The Ars Technica recap Ars Technica · 12 December 2024
“OpenAI plans to release RFT to the public in early 2025 but currently offers limited access through its Reinforcement Fine-Tuning Research Program for researchers, universities, and companies.”
-
E3 That’s a wrap on day 2 TechCrunch · 6 December 2024
“Today’s stream featured demos for “Reinforcement Fine-Tuning,” which OpenAI describes as the ability to fine-tune its o1 model on users’ own data sets.”
-
E4 12 Days of OpenAI OpenAI Own record
“This fine-tuning technique is designed to improve the performance of our reasoning models on verifiable domain-specific tasks.”
Reasoning
TechCrunch's 2024-12-06 day-2 wrap-up says that day's livestream showed demos of "Reinforcement Fine-Tuning", which OpenAI described as fine-tuning o1 on users' own data sets (E3). The Verge reports that after the livestream, Altman described it as making it easy to create expert models in specific domains with very little training data (E1). OpenAI's own page frames it as improving performance on domain-specific tasks (E4). Ars Technica reports that at launch, access was limited to a research program, with a public release planned for early 2025 (E2). The resolution criteria list a limited research preview as a slight difference that makes the claim mostly true rather than true.
Cite this claim
Sam Altman, “OpenAI announced reinforcement finetuning on 2024-12-06, which makes it easy to create expert models in specific domains with little training data.” (said 6 December 2024). Verdict: mostly true. Reliability API, claim c_0390dc446904, method 0.1.0+c876030c, dataset release v1.1.2. https://epistacklabs.com/claims/c_0390dc446904
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Sam Altman
-
Mostly true
“not until 2020 did we train GPT-3, which was the first model that I think actually was over some threshold of usability.”
OpenAI trained GPT-3 in 2020, which Altman considers the first model above a threshold of usability.
-
Resolved true
“that it is so capable that it is discovering new knowledge, doing science, able to add huge economic value, create entire pieces of complex software.”
-
Mixed
“They now hallucinate a little bit.”
-
Mixed
“They don’t do anything unless you ask, right?”
AI models such as ChatGPT do not do anything unless a user asks them to.
-
Mostly true
“But in terms of figuring out what questions to ask or to work on something over a very long period of time, we are definitely not close to human performance.”
Verdicts are made automatically; how verdicts are made.