RAG with LLMs: Easy to POC, Hard to Iterate On.
The paradox of LLMs is that they have made POCs (Proofs of Concept) incredibly easy to create, but at the same time, they have made iterating on the product ...
Generative AI
Is There a Future for On-Device Open Source LLMs?
One obstacle to using open source LLMs is the cost of hosting them on GPU instances in the cloud. But perhaps there is a future for on-device open source? ...
Why bigger context windows won’t make RAG disappear.
We sometimes hear that RAG (Retrieval Augmented Generation) is destined to disappear because LLMs have an increasingly large context window and can therefore...
Padding strategy matters when Fine tuning LLMs
A poor padding strategy can disrupt an LLM after fine-tuning, and the reasons are quite subtle. This is especially true since there is no pad_token in the Ll...
Navigating the Jungle of LLM Quantization Formats, GGUF, GPTQ, AWQ and EXL2 - Which One To Pick?
Quantizing LLMs reduces calculation precision and thus the required GPU resources, but it can sometimes be a real jungle trying to find your way among all th...
RAG vs Finetuning - Complementary not Opposing
RAG and finetuning are often opposed: Should we fine-tune our LLM or do RAG? In reality, the approaches do not oppose each other. On the contrary, they compl...
Assessing the Criticality of AI Failures
What is the criticality of a bad prediction from your algorithm? In my opinion this is one of the first questions to ask before considering a generative AI o...
The Non-Determinism of OpenAI and Anthropic Models
OpenAI and Anthropic's models are not deterministic, even at temperature zero and with a fixed seed, and this can be a real problem. How is it that some...
From Classic ML to LLMs: Similar Development Trajectories?
It's interesting to see how certain patterns in Machine Learning seem to recur, including with LLMs: Classic ML: </p> Step 1: The nature of the mo...
4 Simple Tips for Improving RAG Performance
Here are four simple tips for improving the performance of your RAG (Retrieval Augmented Generation): </p> Understand the RAG: The RAG combines a "R...
Most LLMs don’t succeed my “coco pops’ test”
Ah, really? Few good results for my Coco Pops test... The new Mistral model: surprising. Claude 2's response: legendary. Bard (PaLM2): Anthropic Cl...
Speed Learning Machine Learning concepts in 1 Hour
If I only had one hour to familiarize myself with an unknown machine learning subject, I would try to choose some steps from the following list (ranked by pr...
Questioning the AGI Dream
I don't really know what to think about these AGI (Artificial General Intelligence) stories. Many players like OpenAI, Anthropic or Deepmind are pursuing thi...
How an Average LLM Can Fool Benchmarks and Match GPT-4
How can an average LLM fool benchmarks and achieve performance comparable to GPT-4? In this scientific article, researchers show how a 13 billion parameter...
The Fleeting Value of LLM Benchmarks
You can't improve what you don't measure. However, the relevance of measurement can degrade over time. Regarding LLMs or machine learning in general, we of...
Riding the LLM Hype Wave
I'd be lying if I said there wasn't hype around LLMs. Of course, the emergence of ChatGPT has propelled LLMs into the spotlight. Today, even people outside o...
Open source LLMs are sometimes cheaper than APIs
There are cases where using an open source LLM is cheaper than going through API calls. The main obstacle to using open source models is the inference cost a...
Is prompt engineering a scalable long-term approach ?
The Scalability Concerns of Relying on Prompt Engineering. I'm wondering if prompt engineering is a scalable long-term approach. I feel like today a large ...
Did Microsoft leak the number of parameters in ChatGPT?
A Microsoft publication is causing quite a stir because it appears to have leaked the number of parameters in GPT-3.5 Turbo, which would be 20 billion ...
Create a website with ChatGPT and Github Pages
The aim of this tutorial is to learn GitHub Pages and enhance your skills in HTML, CSS and JavaScript by building a website in a fun, engaging way....
Test Mistral-7b and Compare LLMs Easily Without Installing Anything
Want to Test the New Mistral-7b Model and Compare it to other LLMs Without Installing Anything? For quite some time now, I've been using this very handy ...
Are LLM Benchmarks Reliable? Not Such a Simple Question…
Are LLM Benchmarks Reliable? Not Such a Simple Question... "When a measure becomes a target, it ceases to be a good measure” - Charles Goodhart. Let's tak...
Can AI-Generated Art be Considered Real Art?
Can AI-Generated Art be Considered Real Art? I was struck yesterday by the comments I found on a LinkedIn post (post in comments). The post in question h...
The Era of Fast and Inexpensive LLM is Underway and Microsoft is Joining the Dance!
The Era of Fast and Inexpensive LLM is Underway and Microsoft is Joining the Dance! This current trend of focusing on developing capable but smaller LMs ...
Large Language Models Are Not One-Size-Fits-All for NLP
Using a jackhammer to crack a nut... That's kind of what we're doing when we use a large language model for a basic NLP task... What about good old TF-ID...
The Fuzzy Notion of Overfitting with GPT4 - What Does it Mean to Overfit on All Knowledge?
What could it possibly mean to overfit on 'all available knowledge'? The notion of overfitting for GPT4 is fuzzy. Does it still make sense? With the emerg...
LinkedIn’s AI “Experts”
(This article is a transcript in English of one of my french linkedIn posts) AI experts are sprouting up on LinkedIn even as it becomes increasingly diff...
From Data Scientist to Data Engineer… ultimately to AI Engineer.
(This article is a transcript in English of one of my french LinkedIn posts) Lately, a clear trend of many Data Scientists moving towards Data Engineer ro...
Does no one care about Data Drift in LLMOps ?
(This article is a transcript in English of one of my french LinkedIn posts) By the way, there is indeed a term to describe the life cycle management of l...
Should we redefine our conception of intelligence ?
(This article is a transcript in English of one of my french LinkedIn posts) We will undoubtedly have to redefine our conception of intelligence. Wi...
The Data Scientist made ChatGPT and ChatGPT undid the Data Scientist.
(This article is a transcript in English of one of my french LinkedIn posts) The headline is deliberately provocative... Nevertheless, I observe that in...