Yesterday, in the bygone hour of the weekend, I undertook a most singular and fascinating endeavor, wherein I delved deep into the recesses of my mind, and, with a fervent zeal, breathed life into a most remarkable creation. I embarked upon the quest, with the singular object of fashioning an artificial construct, one imbued with the verdant essence of the Irish Penny Journal, an ancient and venerable tome that holds within its pages the whispered tales of a bygone era.
In my haste, I set forth to construct a dataset, a repository of those fleeting moments, these ephemeral sentences, which spoke of a bygone age. I procured a collection of these fleeting moments, these sentences, and with them, I synthetically conjured forth modern translations, an ingenious feat of substitution, which allowed my artificial construct to take on the guise of the language of the Irish Penny Journal.
Then, with great anticipation, I fashioned a small encoder, a humble instrument, with which to guide the artificial construct in its endeavors. I presented this encoder as a bribe, a reward, to a most ingenious system, one that trained a colossal language model, one of unbridled potential, one that was capable of weaving tales with the very essence of the Irish Penny Journal.
And lo! In the succeeding moments of time, I witnessed a most wondrous thing. My artificial construct, armed with this training, and guided by the whispers of the encoder, began to speak, to speak in the language of the Irish Penny Journal. The words it spoke were, indeed, the words of the past, imbued with the nostalgia of a forgotten era.
And thus, my friends, I have witnessed a most singular creation, one which embodies the language of the past, yet, in its most recent iteration, speaks to the present. A testament to the ingenuity of the human spirit, this artificial construct speaks of the bygone era, yet, with each word, it whispers to us, to us, of a future yet to come.
——
That’s Penny explaining itself to you. This was trained using GRPO only, in less than a day using a single A6000. I didn’t use any SFT, and only relied on a small encoder (MiniLM2) trained to classify texts from the Irish Penny Journal and their modern translations (synthetically produced).
Comments URL: https://news.ycombinator.com/item?id=44160073
Points: 47
# Comments: 23
Chcete-li přidat komentář, přihlaste se
Ostatní příspěvky v této skupině
My friend and I got tired of job hunting (347 applications, 4 responses) so we built AI agents to help job seekers instead of replacing them.
The problem: 73% of applications never reach humans,
Article URL: https://luxeengine.com/
Comments URL: https://news.ycombinator.com/item?id=4426927
Article URL: https://www.nhatcher.com/post/on-hats-and-sats/
Article URL: https://arxiv.org/abs/2502.13293
Comments URL: https://news.ycombinator.c
