Top 10 LLMs you should know
Welcome back, fellow geeks. Let’s check out what’s new in the field of AI this week.
What I have for you today:
- Top 10 LLMs you should know, and how they’re used.
- A curated list of free AI courses from Google and Amazon.
- 5 best performing ChatGPT prompting frameworks.
- Weekly Pulse on freshest updates in AI world. Edition №9
Top 10 LLMs you should know, and how they’re used
An LLM, or large language model, is a general-purpose AI text generator. It’s what’s behind the scenes of all AI chatbots.
To produce correct and coherent output, LLMs are trained on huge corpuses of data.
There are dozens of LLMs, and you are most certainly lost in all these “gpts” and “lamas”.
Here’s the simple guide how to differentiate top 10 most significant, interesting, and popular LLMs, and where are they used.
1. GPT by OpenAI — 175B+ parameters, used by ChatGPT, Microsoft, Duolingo, Stripe, Zapier, Dropbox
2. Gemini by Google — parameter data is largely unknown, used by Google in some Android apps and Google AI portal, previously known as Bard
3. PaLM 2 by Google — 340B parameters, used in Google Docs, Gmail and other Google Suite apps
4. Llama by Meta — 7/8/13/70B parameters, used in Meta AI assistant, across Meta applications (Facebook, Instagram, Threads), open source
5. Claude by Anthropic — parameter data is unknown, used in Slack, Notion, Zoom
6. Grok by xAI — parameter data is unknown, used in X(Twitter), Grok Chatbot
7. Mixtral by Mistral AI — 39B active out of 141B parameters (SMoE), academic and commercial use, partnership with IBM, ChatGPT-like “Le Chat”, open source
8. Command, Rerank and Embed by Cohere — parameter data is unknown, used in Notion, Jasper, HyperWrite
9. StableLM by Stability AI — 3/7B parameters available, 15/30/65/175B in progress, part of StableDiffusion, open source
10. Vicuna by LMSYS Org — 7/13/33B parameters, used in LMSYS Chatbot Arena, open source
But why so many of them?
↳ OpenAI’s research and effort showed the practical side of LLMs, so others started doing the same.
↳ Many models could be retrained or adapted so companies could tailor those to their needs.
↳ Despite the difficulty to train an LLM, it becomes more affordable and takes less time (weeks or months).
↳ Investment landscape is plentiful in AI field.
Which other prominent LLMs are worth mentioning?
A curated list of free AI courses from Google and Amazon
Last week I wrote about Microsoft, today adding Google and Amazon to this list.
1. Google courses today (intro to advanced level, links in attached pdf):
Beginner level:
↳ Introduction to Large Language Models — 30 minutes
↳ Introduction to Generative AI — 45 minutes
↳ Introduction to Image Generation — 8 hours
↳ Transformer Models and BERT Model — 8 hours
Intermediate level:
↳ Attention Mechanism — 8 hours
↳ Encoder-Decoder Architecture — 8 hours
↳ Create Image Captioning Models — 8 hours
↳ Gemini for Google Cloud Learning Path — 8 activities, 45 minutes to 6 hours each
Advanced level:
↳ Generative AI for Developers Learning Path — 11 activities, 30 minutes to 8 hours each
2. Amazon courses (again, intro to advanced level, links in attached pdf):
Beginner level:
↳ Introduction to Machine Learning on AWS — 6 hours, beginner
↳ Generative AI Learning Plan for Decision Makers — 3 hours
↳ Amazon Transcribe Getting Started — 1.5 hours
Intermediate level:
↳ Low-Code Machine Learning on AWS — 4 hours
↳ Generative AI Learning Plan for Developers — 11 hours, Python-specific
↳ Machine Learning Learning Plan — 11 hours
↳ Foundations of Prompt Engineering — 4 hours
Advanced level:
↳ Building Language Models on AWS — 5.5 hours
Bookmark it!
Every week I’m going to add to this pack!
5 best performing ChatGPT prompting frameworks
Using ChatGPT correctly is simply about communicating well and managing your expectations. If you excel at both in the real world, you can manage AI tools to generate wealth in the 21st century.
Here are 5 of the most effective prompting schemes for ChatGPT:
1. R-T-F — Role, Task, Format
↳ “Act like a [insert the ROLE you want AI to take]. Give me a [insert TASK] in [insert FORMAT] format.”
2. T-A-G — Task, Action, Goal
↳ “I need to accomplish [insert the TASK description]. In the process, I will perform [insert ACTIONs or processes to be undertaken]. As a result, I must achieve [insert the desired GOAL].”
3. C-A-R-E — Context, Action, Result, Example
↳ “Here’s the background of the task: [insert CONTEXT information]. I want you to perform [insert ACTIONs for AI to undertake] and achieve [insert description of desired RESULT]. Here’s an example: [insert EXAMPLE clue as a concrete illustration].”
4. R-I-S-E — Role, Input, Steps, Expectation
↳ “Act as [insert ROLE AI should take]. Here’s what you need to know: [insert INPUT describing situation or resources available]. Perform these steps: [insert STEPS instructions], and give me a [insert EXPECTATION of the desired outcome].”
5. R-A-C-E — Role, Action, Context, Expectation
↳ “Act as a [insert ROLE AI should take]. I want you to perform [insert ACTIONs for AI to undertake]. Here’s the background of the task: [insert CONTEXT information], give me a [insert EXPECTATION of the desired outcome].”
There are dozens more, obviously. Which are working for you?
Weekly Pulse on freshest updates in AI world. Edition №9
Curious of latest AI? I read dozens of article, so you don’t have to.
In full on the latest advances in AI and its implications.
Edition #9: 8 compact models for the masses from Apple, Meta unlocks Metaverse’ open OS, Microsoft’s tiny Phi-3 model, Biotech Profluent firm edits human genome with AI, Meta’s smart Ray-Ban major upgrade, Llama 3 topples ChatGPT, and more.
8 compact models for the masses from Apple:
- submits 8 new AI models to Hugging Face
- models are called OpenELMs, openly available for use
- smaller models, efficient for devices, not cloud
- also better privacy — no cloud upload
- unusual move for Apple’s typically closed software
Meta unlocks Metaverse’ open OS:
- opening Meta Quest OS to 3rd-party hardware
- introduces Meta Horizon OS for mixed reality
- combines tracking, blending digital-physical
- leverages Android AOSP for mobile foundations
- includes app store, social features, companion app
- aims to bridge virtual worlds across devices
Microsoft’s tiny Phi-3 model:
- smaller, affordable
- comes in Mini (3.8B), Small (7B), and Medium (14B)
- Mini 3.8B beats Llama 3’s 7B model on benchmarks
- very compact: some users could run it on phones
- smaller size usually means lower costs for users
- good option for budget-conscious companies/users
Biotech firm Profluent edits human genome with AI:
- GenAI aiding deep scientific research
- reveals open-source AI-generated gene editor
- enables precise editing of human genome
- aims to create cures for currently untreatable diseases
- technology available to license for ethical research
- major step forward in using AI for scientific research
Meta’s smart Ray-Ban major upgrade:
- smart glasses collection expanded
- adding video calling with WhatsApp and Messenger
- Meta AI with Vision for hands-free info
- users can ask glasses about what they’re seeing
- providing information without using hands
Llama 3 topples ChatGPT, experts claim:
- Wharton prof says 8B Llama 3 beats free ChatGPT-3.5
- Llama 3’s 70B model ranks 2nd only to GPT-4 Turbo on Lmsys
- Llama 3 models are open-source, cheaper to run
One-liners:
- Nvidia is acquiring the Israeli startup Run:ai, which built a Kubernetes-based GPU orchestrator, for approx. $700 million, highlighting the growing importance of Kubernetes in the generative AI era.
- Inspired by AI-powered coding tools, former Microsoft engineer Igor Ostrovsky built Augment, an AI-powered coding platform that emerged from stealth with $252 million in funding from incl. former Google CEO Eric Schmidt.
- The Rabbit R1, a new AI-powered mobile device from Humane, is impressing reviewers despite a rocky start for the company’s previous AI gadget, though it still has some limitations.
- Elon Musk’s xAI, a dreaded competitor to OpenAI, is reportedly close to raising $6bn in funding.
Hey there!
With 15 years of experience in software development, system architecture, and team leadership, I’m passionate now about helping business leaders realise the power of AI.
I usually write for my LinkedIn. Follow me there for actionable guidance on AI, leadership and strategy!
Cheers!
