Louis Guitton

🎓 Function Calling Crash Course Excited to publish a detailed overview of function calling with the OpenAI APIs. Covers: - An Overview - Examples - Customizing Function Calling - Parallel Function Calling - Function Calling for Model Feedback https://lnkd.in/eDTC4xHK

90

1 Kommentar

Hugging Face updated its OpenLLM leaderboard to use harder benchmarks and there are some really interesting findings: 🚫 Models have gotten so good at high-school-level benchmarks they're no longer useful. Some model developers may also have explicitly trained to do well on these benchmarks, hence now do poorly on harder ones. 🔐 Closed models are way better on tough problems. The best GPQA (graduate-level questions) performance of any open model is 17.23. Claude 3.5 Sonnet gets 59.4 0-shot on the toughest set of questions. 🎓 Small models can do very well. Microsoft's tiny Phi-3 medium model—a 3.8B parameter model—performs nearly as well as models over 18X as large, and beats Lllama-3 70B on benchmarks including GPQA. This is quite surprising and suggests that the approach to training Phi models—using high-quality synthetic data generated by larger models for a learning curriculum—is very fruitful. 🇨🇳 Chinese models are doing very well. The idea that China is far behind the US in AI development no longer seems to be true. Leaderboard is here: https://lnkd.in/gv92Z5CR

3

The AI Manifesto According to Josh Today in business speak, AI is synonymous with LLMs. This has us talk about ourselves instead of about our customers. Until 5 minutes ago, AI meant a machine exercising human judgement and implied no particular perspective on how that was accomplished. This phrasing was more powerful because it defines a user experience, not the tech under the hood. Does the thing feel smart like working with a person or does it feel like a computer? Set aside the stochastic parrot people predictably clamoring about stochastic parrots and go play with ChatGPT or Stable Diffusion for a few minutes. These tools are obviously closer to working with a human than any machine has ever been and are free and available to everyone. That's completely insane and unprecedented, no matter how cool the anti-hype crowd tries to play it. But - contra the pro-hype crowd - your customers don't care if you are using machine learning or LLMs or are in the cloud or are in your own data center. They care about your features, SLAs, price, and roadmap. YOU should care about your tech. It's how you deliver those things and LLMs could disrupt a lot of products. But "Now with LLMs!" is not appealing to anyone except your shareholders and bosses. So keep that message for them. The missed opportunity is that LLMs have the potential to advance AI products more than any technology we've seen. Not because of AGI superintelligence or some such, but because they are properly used as an improved interface to reliable systems. Rather than falling in love with the tech and making wrappers for LLMs, recognize that we are now within reach of decent natural language interfaces. We don't want LLMs to schedule a haircut or write longer emails or read shorter emails, but we all hate car touchscreens and there's potential for better navigating websites and e-commerce. Everyone wants access to the business data to get answers to questions even if they don't know how to write SQL. And there's a sea of unstructured data ready to be tamed and ingested. LLMs shouldn't replace the back end, but rather they are the new front end. Companies with unique data and powerful decisioning will continue to have an advantage. Meanwhile, keep the mental model of AI as human-level interactions and judgement, not a particular box from OpenAI. Talk about how great driving the car makes you feel, not the variable valve timing. Of course, you still need the engineers who use variable valve timing to make the car fun to drive. But don't forget why people care about your product. And if you want an AI leader who can deliver the tech and keep the board off your back - but does it while talking about the customer - give me a ring.

12

3 Kommentare

🚀 How much does it costs to Training LLMs and Where Smaller Players Fit In? Training Large Language Models (LLMs) like GPT-4, LLaMA 3, and Gemini is not a means feat. It requires massive amounts of data, compute power, and financial investment. As these models grow, so do the resources needed. Here's a rough breakdown of the costs: 💡 GPT-4: OpenAI spent $100 million to train GPT-4, running tens of thousands of GPUs for weeks. The energy used was comparable to powering entire towns. 💡 LLaMA 3 (Meta): Meta trained LLaMA 3 using huge datasets and a distributed compute setup, likely spending similar amounts as GPT-4. 💡 Gemini (Google DeepMind): Google used its global data center infrastructure to train Gemini, a model with hundreds of billions of parameters, likely exceeding $100 million in costs. 🔋So where does these costs come from?: 1. Data: These models consume terabytes of text data, processed and structured before feeding into training pipelines. 2. Compute: Thousands of GPUs or TPUs power high-performance clusters that run continuously for weeks or months. 3. Energy: GPT-3 used 1.3 GWh of electricity during training, which could power 120 U.S. homes for a year. ⏩ How does the Future of the next Gen LLMs look: The next generation of LLMs could cost $500 million to train, pushing the boundaries of hardware and energy efficiency. Only big players like OpenAI, Google, Meta, and Microsoft will have the resources to lead an effort at this scale. 💡Where does this leave the Smaller Players: We will get into a future where we will start leveraging the best of both worlds. Smaller companies won’t train LLMs from scratch but will: - Fine-tune pre-trained models on specialized datasets. - Build niche applications by customizing LLMs for specific industries. - Focus on optimizing models for efficiency and scalability. - Rely on open-source models to innovate in targeted areas. The future will be a hybrid world where Big tech will continue to drive the training of foundational models, while smaller players will carve out unique applications, making AI accessible across industries. #LLMs #AI #GenerativeAI #MachineLearning #GPT4 #LLaMA3 #Gemini #AIResearch #TechTrends #AIInnovation #AIInfrastructure #FutureOfAI #AIStartups

30

3 Kommentare

pgai Vectorizer. Makes it possible to use one SQL command that will manage your vector embeddings "without" the usual engineering challenges to keep it in sync with your source data!

5

Semantics and Significance: Where Context Really Matters I've had long-running discussions with my fellow semantic interoperability wonks about whether the meaning of a word requires context. My position is this: If we use a recognized string of letters / pictographs in a sentence, the meaning of that word must be available before the creation of the sentence. Often words have multiple meanings, but unless the word in question already has a meaning, the sentence only collapses that list of available meanings into a particular instance from the list. So, the purpose of the sentence is to select a meaning not create one. On the other hand, a perfectly understandable, if somewhat hyperbolic sentence such as the following, is created from a collection of commonly understood terms to create a true statement of fact, a data artifact: "Update: Bedford Metals Jumps 22% as It Plans the Next Exploration Program at the Ubiquity Lake Uranium Project" The significance of this true statement - in context of what has come before - adds the kind of understanding CxOs are looking for in their analytics applications. See the graphic.

13

6 Kommentare

Here's my notes and video for the keynote I gave at PyCon Italia ❤️ 🐍 🇮🇹 yesterday, "Closer to the Metal." The post itself is titled "don't worry about LLMs" and that's really what I was trying to get at with this talk. LLMs are a big topic and there is a lot of demand to build them quickly, but the space is very confusing and inflated with hype. How do we remove the hype and use LLMs to build things we need? We focus our attention on one thing at a time and use the practical engineering and machine learning skills we already have.

217

5 Kommentare