Skip to main content

ELL Blog

AI Awesome List

https://blog.elijahlopez.ca/posts/ai/ Elijah Lopez

Please drop a comment if you think something is missing or want something added.

Ever since the launch of ChatGPT, AI powered apps have been blowing up. Every single day there’s a new AI powered app that solves a specific use case. Some of which I have no need for, but are good to know. There’s two ways to keep up to date with AI, one is to subscribe to a newsletter, and the other is to bookmark a list, so that’s what this is going to be. The problem with existing lists are their scope or vision sucks or they are not bleeding edge. This list aims to be bleeding edge and will remove unmaintained / crap projects. It also aims to categorize AI so that everybody from hacker enthusiasts to corporate overlords will benefit. I’ve looked at other lists and they do a piss poor job of moderation. An awesome list shouldn’t be recommending crap TTS products to the user.

There’s two parts to this article. One focuses on the models and how to select them, whereas

Table of Contents

AI Models

I’m starting with the topic of benchmarks because the best way to be ahead is by using the forefront leader in AI which is only possible by reading benchmark scores. One day it could be OpenAI, the next Google, the next some whale named DeepSeek, and then something called Qwen. Truly, it’s better to make informed decisions based on a heuristic than it is to blindly follow the sheep and limit yourself a single platform.

Benchmarks

In my opinion, the current state of benchmarks is very messy. I’m making progress on fixing it myself with blog posts such as SimpleQA Leaderboard however, there are a few more I would like to maintain. I suggest using these benchmarks as a heuristic in finding a handful of models to test yourself before going with one of them.

Populist Benchmarks

I’m naming these populist benchmarks because it’s basically a popularity contest (real and synthetic) rather than a merit-based benchmark.

Knowledge Benchmarks

Intelligence

Coding Benchmarks

The problem with LiveCodeBench is that there are different cut off dates depending on when a model is graded plus the benchmark is continuously updated. When using the earliest cut off date, some models might’ve been “contaminated” and when using later cut off dates, some models do not show up at all! If we use the scores self-reported by the company, we still run the risk of reporting non-comparable numbers. Based on how LCB works, model scores are expected to depreciate over the long run; If Grok 3 scored 100 on LCB today, it is almost certain to score less than 100 in a year.

The problem with EvalPlus is that it doesn’t include bleeding edge models, it’s basically almost solved, and not many new models even report their scores anymore.

Multimodal Benchmarks

This tests visual capabilities.

Writing Benchmarks

Agentic Benchmarks

Agentic benchmarks are very new and personally I’m not too sure what these benchmarks do or even what is considered good. Personally the only agent I would ever value is one that has the same worth ethic and intelligence as I am during when I’m at my peak productivity.

  • Scale MultiChallenge
  • BrowseComp

Proprietary Models

Model Name Company Blog Chat App
Gemini Google Google DeepMind Blog Google AI Studio
OpenAI Platform OpenAI news ChatGPT
Grok xAI news Grok
Claude Anthropic news Chat
Cohere Platform Cohere blog Dashboard

Cohere is really slacking. I almost forgot about them.

Text to Image

Open-Source Models

A table of companies that release open-source LLMs. I suggest adding these to your RSS reader or signing up for email updates. In the future, hopefully RSSHub adds support for these.

When it comes to downloading models, most vendors (that’s what I’m calling the companies) will link you to Hugging Face. My biggest gripe is how Hugging Face isn’t using P2P torrent technology to speed up downloads and reduce strain on their own servers! What a missed opportunity.

Model Family Company Blog Chat App
DeepSeek Chat Stream Chat Stream Blog ChatStream Chat
Qwen AliBaba Qwen Blog Qwen Chat
Llama Meta AI at Meta Blog and Meta AI Research OpenRouter
Mistral Mistral Mistral News Le Chat
Gemma Google DeepMind Blog Google AI Studio
Phi Microsoft Microsoft AI Platform Blog OpenRouter or Azure AI Foundry
ChatGLM THUKEG & Z.ai Twitter OpenRouter

Note that sometimes proprietary models are open-sourced, but this usually happens long after a model from an open-source family has beaten the outdated proprietary model. Therefore, they are not included in this list for end-users.

These are also the base models. If you go tho HuggingFace and LocalLLAMA, you can find many remixes (fine-tunes) of the base models to yield specific results. There are so many people doing this.

AI API Providers

These companies don’t make the models, but offer inference, either by hosting models or via a gateway

  • OpenRouter (one API provider to use many APIs)
  • HuggingFace (which links to Amazon, Azure, and Google)
  • Groq
  • Together.ai
  • Replicate

AI Applications

AI but for specific tasks. A mix of apps and models (when applicable). Skip to Local AI Models to learn more about running open-source models using open-source apps

Chat

The default type of application when people say LLMs. and for a list of models. Alternatively, if you don’t mind paying, an easy way to interact with all models is through OpenRouter. Read How to Run Open-Source Models if you want to run text generation models locally.

Recall (RAG)

Using AI to boost productivity by letting AI do a domain search and recall on the content you provide. See section on jargon to understand what RAG is.

  • NotebookLM: a tool to understand information
    • TODO: somehow combine this with an RSS feed sync
    • Can be used to combine a bunch of files together (pdfs, websites, youtube videos, audio, word files, etc)
    • Can create a podcast out of it too
  • Morphik AI
    • This is more for developers who want to build enterprise applications

Some of these can also be considered a subset of “Chat”

  • Liner (specialized in getting answers with reliable sources which means helps to avoid plagiarism)
  • Linkup
  • Exa
  • Perplexity

Interesting Media Research

Image

Video

Audio

3D

Creating 3D wireframes with Gemini

Websites

Creating Mobile Apps with AI

Marketing

Software Development

Aside from prompting the Chat apps, there are a variety of ways to use AI. I personally use Cline with an OpenRouter API key, however this is because I never got RooCode to work and so didn’t bother setting it up.

  • VSCode Integrations
    • Roo Code (Cline fork that is more community contribution friendly, previously Roo Cline)
    • Cline (as Debian is to Ubuntu, Cline is to RooCode)
    • GitHub Copilot
    • Continue
    • Twinny (not user-friendly at all and useful only for local models)
    • Qodo.ai (previously CodiumAI)
      • “Agentic AI for reviewing, testing , and generating code – continuous quality at every step”
  • Other IDE Integrations
    • Aider (AI pair programming in your terminal)
    • QoDo Gen (VsCode and JetBrains)
    • Continue (VsCode and Jetbrains)
  • IDEs
  • Other
  • GitHub Integration
    • QoDo Merge

Figma to Code

If you don’t mind TailwindCSS (I hate it), you may find these tools useful. Speaking from experience, you can just send Claude screenshots of the design and tell it to implement the design using a library like Mantine and it will implement it with 80% accuracy.

CyberSecurity

-peneterrer: AI Security Tester (pairs well with vibe coded websites)

We’re so confident in our security testing capabilities that if we don’t find any vulnerabilities, you get your money back. No questions asked.

Agents

Writing

I take great pride in stating that this blog post is ironically 100% free of AI generation. I’m not opposed to AI but knowing that AI is a FLUFF GENERATOR means that I can really only use AI to turn a bland writing post into a pleasant post (see That Time I Went to a Dog Food Eating Convention). If you rely on AI 100%, it can make your content over the top sweet, so I find the best way to use it on your own words is to incorporate some of its suggestions rather than all.

I have two book ideas I want to pursue one day in the future. What I don’t approve of using AI for, is to generate redundant slop, which is basically plagiarism. Jetpack AI’s own demo shows itself generating slop. Using AI to write a blog post about being a better blogger? What? I think these companies are going to get whatever moat they think they have eaten by Chat apps or open-sourced fine-tuned models.

Here are some thoughts I have on pursuing fictional writing

Other AI Apps

What else can you accomplish with AI?

Convert line art out of an uploaded sketch + colorize with Gemini

Extracting a professional shot product from a picture

Combining a product with a picture of a human (which could also be AI generated) for marketing or e-commerce shots. You can also do virtual try ons.

Creating a pixel sprite using Glif Sprite Generator, and then turning it into concept art using Gemini

Creating gif animations using Gemini

Interior decoration

AI Research

Local AI Models

The best aggregate about open-source LLMs is r/LocalLLaMA. However, it should be noted there is a base knowledge expectations required. I’ll go over it briefly.

How to Run Open-Source Models

This section comes first because it’s derived from the resources in the rest of this page. The models you will be able to download will be limited by your RAM. To run a model locally, you may need hardware. Next, pick an open-source model based on the benchmark closest to the task you want. In LM Studio, search for the model, and choose a quantization to download.

Once you’ve downloaded models, you can load them in LM Studio, select a system prompt, and continue. You can also start a server and integrate with local apps that are ollama compatible.

Open-Source Interfaces

An interface is something that interacts with the model, but not the model itself. I know of a few.

Interfaces

Some of these require “backends” which all come from llama.cpp or kobold.cpp. However, Ollama is super simple for running models.

Offload Tensors for Performance Improvements

This post is very new and talks about how even with less than the recommended hardware requirements, you can still improve throughput by being selective about what is offloaded to the GPU. It seems that some tensors like FPN tensors happen to be very large and use basic matrix multiplication which can be done efficiently on the CPU, whereas small tensors like attention tensors benefit from GPU parallelization! It’s a breakthrough innovation in my opinion. Original credit goes to u/EmilPi.

Learning

Using AI

Prompt Engineering

Building with AI

  1. 21 Lessons, Get Started Building with Generative AI

Researcher-oriented

  1. Neural Networks and Deep Learning
  2. Language Models are Unsupervised Multitask Learners
  • pytorch
  • tensor
  • llama.cpp

Read Frontier Papers

One of the most eye opening things my friend told me is that there is practical benefit to reading frontier research articles. In his case it was related to algorithmic trading, but I’m going to go further and suggest that it applies to all areas of frontier development. Whether that be quant finance, AI research, cancer research. There is merit in spending time on reading research if you are able to utilize new information readily.

Follow AI Researchers

They will talk about new things they may have learned or how to break in, or tweet out an article, etc.

Get Resources

The easiest way to get resources is to get MONEY. To get MONEY, you need a JOB. It’s probably easier to GET A JOB than to already have the money necessarily to buy hardware.

AI Research Companies

Company Based Notes
Cohere Canada/USA Command R model
Open AI USA The creator of ChatGPT, led by Sam Altman (disclosure, I’m biased against Altman)
Google DeepMind USA They came out with the original Transformer research that OpenAI used successfully and work on Gemini and Gemma
xAI USA Creator of grok, very integrated with X, owned by Elon Musk
Meta AI Anywhere Creators of LLaMA
NVIDIA USA Manufacturer of the best commercially available GPUs for training AI
Anthropic USA Claude
Safe Superintelligence Inc. Palo Alto, Tel Aviv Ilya Sutskever former OpenAI Chief Scientist & Co-founder
Thinking Machines Lab USA? Mira Murati former Open AI CTO
Ndea USA intelligence science lab founded by X:@fchollet & X:@mikeknoop
Vector Institute Toronto, CA -
Mila Quebec -
Ai2 Seattle, WA -

AI Product Companies

AI Hardware

  • NVIDIA Tensor Core GPUs: enterprise
  • Truffle: end-customer hardware for running models locally

Jargon

Final Words

There are a lot of variety of tools, models, and research. There’s an opportunity to capitalize on research, combine multiple models, and provide an offering that is SOTA. If you’re unemployed, you should seize on this opportunity. VC appetite is high for AI-related companies, and competition is very hot.