Petr Brzek’s Post

Fine Tuning the GPT-3.5 Turbo

"GPT-4's parameter and training details leaked GPT-4 is reported to have over 1.8 trillion parameters across 120 layers and utilizes a mixture of experts (MoE) model. It is trained on approximately 13 trillion tokens and has a 32k sequence length. The training cost for GPT-4 is estimated to be around $63 million, and inference costs are higher compared to previous models. Other details include parallelism strategies, dataset mixture, and speculation about its performance." https://lnkd.in/efNk3QcS

GPT4 got significantly faster with the 9/25 update and I'm so here for it.

There's a common belief that #opensource LLMs aren't as performant as #GPT4. Well, oftentimes they're not... unless you #finetune! You can achieve GPT-4 accuracy or better for your use case with the power of fine-tuning. Want to learn how you can fine-tune your own #LLMs that outperform GPT-4? Join our upcoming virtual workshop to learn more: https://pbase.ai/43Yuvfb

GPT 3.5 turbo is fine tunable now. Slightly expensive to consume fine tuned model but apparently it can match or exceed the accuracy of GPT 4 on narrow set of use cases https://lnkd.in/gvzRyf8n

After GPT-4 is released, the cost for API calls to GPT-3.5 is dramatically reduced. It is so miniscule that thousands of small requests only cost $1 or $2. You can actually use GPT-3.5 on a large scale.

In most of your projections for GAI Costs across the next 3 years, make some assumptions in proper allocation of budget. Remember how the cost of storage dropping so low may not have decreased your budgets, but it changed every one of your enterprise's core systems and how you treat data... more capability and more data than in all of history. The same thing will happen with forms of AI as prices drop. However... remember the most expensive things, that work and have the most value will have a high price- Just look at how the pricing worked for IBM Mainframes, Oracle Databases, Microsoft Office and Sales Force - AI Software and Hardware will follow the same trajectory. This change over will cost most Global Fortune 1,000 companies hundreds of millions of dollars to adopt and adapt. It is also very likely that we will enter the software paradox problem again... (notes on the sofware paradox in the comments). #Pricing #economy #change