Karl WAN NAN WO

Today is the first OpenAI DevDay for this autumn. No new models are expected, but OpenAI has promised to introduce some long waited features to APIs. This is the schedule (in San Francisco timezone): - 10:00 - Welcome kickoff by Olivier Godement & Romain Huet - 11:00-11:45 - Topic to be revealed during keynote - 12:00-12:45 - Topic to be revealed during keynote - 12:45-12:00 - Lunch - 2:00-2:45 - Topic to be revealed during keynote - 3:00-3:45 - Topic to be revealed during keynote - 4:00-4:45 - Fireside chat with OpenAI leaders - 4:45-7:00 - Closing reception So reading this, there could be up to four new announcements coming. Here is my wishlist in priority order: - Audio/voice multimodal API for gpt-4o (so advanced voice mode API) - Function call webhooks, so that AI could call app function in mids of voice-call or o1 reasoning sessions. - New gpt-4o image generation API - Prompt caching - Web search enabled api (ala Perplexity or Command-r) But I hope they leave the best for London in end of month, where I'm going 😂

12

1 commentaire

I comment I made on a thread of Benjamin White's received some love: "If you're looking for a freelancer to write shoehorned history/LegalTech analogies, Benjamin White, I daresay I could have a go! Eleanor of Aquitaine: M&A deal where the acquired party brought a hell of a lot to the table but was left languishing alone by the new owner Roman Republic: a plucky startup takes on and swallows up all-comers until it becomes a juggernaut that loses momentum, disintegrates and is picked apart by ravaging outsiders. PE asset stripping? Visigothic and Ostrogothic kingdoms East India Company: Alternative Government Service Provider. Does the same work as government, operates factories, and works solely for corporate profit." Once I've finished the draft of my next novel and have sent it out to beta-readers I'll knew a new creative project to keep me sane. Who'd read a tongue-in-cheek look at history through the eyes of a Tech Bro?

4

I faced this situation when moving same code from one GPU to a more powerful one —————— Summary from the article: In software engineering, dependency versions are constrained and containers are used to ensure reproducibility. However, this may not be enough with LLMs if the GPU is not identical. All of this can introduce challenges, especially if the hardware used in development differs from that used in production, or if you want to opt for more powerful GPUs at some point in order to scale up in production. In this article, I first present the protocol of the experiment, then compare the results between two environments from which only the GPU differs, one with an Nvidia T4 and the other with an Nvidia A10G. I start with simple cases and then longer prompts, like for a RAG. Each GPU has its own architecture, optimizations and parallelization capacity, which can lead to small differences in the accuracy of the calculations. With LLMs, these small differences repeated over millions of calculations can lead to a discrepancy influencing word choice when generating sequences. Since LLMs are autoregressive, i.e. they predict the next word based on previous ones, a different word in the middle of the generation can cause a cascading reaction, resulting in a discrepancy in the response. For RAGs, longer prompts require more computation, increasing the possibility of inaccuracies and promoting potential discrepancies between two GPUs. Compared to my usual posts on LinkedIn, this article allowed me to flesh out my thoughts and experiences in a more in-depth way, while also including some code.

4

3 commentaires

Want to get started in automated anomaly detection but lack data science skills ? 👉Discover Diagfit ! Turn now data into actionable insights : try Diagfit ! 🔎Discover how : https://lnkd.in/dw27ZvSq #predictivemaintenance #AI #MachineLearning #DiagFit #Diagnostic #timeseries

5