AI tweets you missed in May 2024
Late night Twitter scrolling sometimes gives you a whip of serendipity that makes a lack of sleep worthwhile.
Last night I saw this on my feed from Séb at Google Deep Mind after my 116th "One more scroll then I'll put my phone down":
has anyone done a 'best of AI twitter' newsletter? ideally without AI-generated commentary/slop? categories + embedded tweets + maybe 1-2 lines of commentary at most would be ideal
— Séb Krier (@sebkrier) May 28, 2024
The good news for Séb, I said, is that I've started a draft of exactly this to revive this long ignored blog. My feed is so well trained on my interest in AI that I have a perfect front row seat to the bleeding edge, where hackers, builders and pwnage lords, in a bid to bookmark all the tools and toys I'd like to explore every Friday (my new R&D day after blocking it off from running my ad agency the rest of the week). I'd started pasting in the most interesting posts into a draft, yet with so many niche tools and directions, I wasn't sure there was appeal for a bunch of pasted tweets.
And then, validation in Séb's tweet. I replied with a deal: sign up to my newsletter, Séb, and I'll start publishing my AI news. And he did, and here we are.
With that context out the way, it's time to start delivering on the '1-2 lines of commentary' promise. Here's everything I bookmarked in May, leaning as much to the alternative headlines you didn't hear as I can manage.
Creative
Klarna CEO has halved its in-house marketing team and reduced image creation costs by $6 million with AI:
genAI will save us $10m in marketing this year. We’re spending less on photographers, image banks, and marketing agencies.
— Sebastian Siemiatkowski (@klarnaseb) May 28, 2024
The numbers are mind-blowing:
- $6m less on producing images.
- 1,000 in-house AI-produced images in 3 months. Includes the creative concept, quality… pic.twitter.com/ioAhxkNS8I
Multiple examples of video > anime style that look great. I'll call it now, we'll have entire TV shows and YouTube channels with identities led in this approach.
I turned myself into a Lofi Boy 🤠
— Framer 🇱🇹 (@0xFramer) May 23, 2024
With Domo AI, it only took 30 minutes.
This is how you can do it too 👇 pic.twitter.com/AX0Ht9wbpz
I put myself into a Midjourney picture 🤯
— Framer 🇱🇹 (@0xFramer) May 19, 2024
I still can't believe how easy it was!
Here’s what I did: pic.twitter.com/JDLa3ZSKW4
Automatically cut down video clips for social media:
I built a video curation assistant using @crewAIInc and GPT-O, and it's absolutely insane. 🔥
— Ale𝕏 Fazio (@alxfazio) May 18, 2024
Look how it selects the most interesting parts of podcasts and speeches, and even returns them as subtitled video clips ready to be published on social media.
Repo in the comments. pic.twitter.com/8yvO0zIm2r
Tools
Yohei is a king in AI twitter and tapped into the hive mind for their favourite tools for AI building. I'm yet to run through the list but it looks like a goldmine. (Yohei also did a fantastic X livestream on his view on the future of AI agents, worth watching if you catch him live):
AI Builders' Favorite Tools: A Hive Mind Survey Summary
— Yohei (@yoheinakajima) May 26, 2024
I recently asked AI builders about their favorite go-to libraries and frameworks for AI projects. Here is a summary of the first ~63 replies, categorized.
Agent Frameworks:
• AgentOpsAI: This tool simplifies the… https://t.co/aIQqBmIxBD
Looking for an alternative to Retool? Try this.
You can try UI Bakery. Disclaimer: I’m bootstrapping it and we are not so enterpris-ish 🙂
— Vlad 🍩 (@v_lugovsky) May 22, 2024
Also, we offer a slightly different user experience, which our users find easier to navigate.
Build landing pages. So many AI apps trying this but when it rides off pretty libraries like Tailwind you know it's going to look good enough.
✨ Introducing Magic UI 🪄
— Dillion (@dillionverma) May 22, 2024
A UI library of animated components to build beautiful landing pages.
Built using @reactjs, @tailwindcss and @framer motion.
100% Free and Open Source.https://t.co/9hmZ6EZV0L pic.twitter.com/zTbLkqUd1t
Tako is not making as much noise as it should with this insanely powerful visualisation tool. I'm watching this like a hawk for our market research tools:
(1) Today we’re excited to introduce Tako, the first AI search engine for visualizing and sharing the world’s knowledge.
— Tako (@TakoViz) May 21, 2024
Tako takes natural language and generates shareable, interactive displays of data called knowledge cards.
Tako’s knowledge graph is built exclusively from… pic.twitter.com/22e9xWibcT
]
Loom making it easy to record SOPs is a game changer for my business. Producing systems is a huge bottleneck; this builds the documentation plus of course a video recording of the process for training:
What @loom did to create AI SOPs is mind blowing.
— Chris Munn (@chrisxmunn) May 17, 2024
It will save business owners $1,000s.
In a matter of months, this work went from:
Onshore @ $30/hr
|
|--->Offshore @ $8/hr
|
|--->AI @ $.02/hr
If you're not using loom to document your processes, you're blowing a huge… pic.twitter.com/eoHSxoaliC
Brainstorm rocket fuel:
This is a pretty weird/cool way to explore new topics. A generative knowledge tree of sorts.https://t.co/pMLAk0xcpc pic.twitter.com/E64OgZTwmC
— Amjad Masad (@amasad) May 8, 2024
Mindmap your tweets:
it took me ~5 minutes to embed, map and label my 10k tweets... the question is do i want to look in this mirror? https://t.co/VxJYJBMYfx pic.twitter.com/bsIakKitKD
— Ian Johnson 💻🔥 (@enjalot) May 3, 2024
Code Interpreter
You can install libraries to Code Interpreter if they aren't available? 🤯
btw if chatgpt doesn't have the library you need you can just upload a .whl—just go to pypi > files and download the python 3.11 manylinux wheel pic.twitter.com/MHGSOalBJx
— thebes (@voooooogel) May 25, 2024
Talking of Code Interpreter, I've been running OpenWebUI as my company's AI chat platform (planning to write up a starter guide for others). Turns out I just needed to run an update.
Already partially implemented with our latest release! https://t.co/RH2kg8WvF0
— Open WebUI (Formerly Ollama WebUI) (@OpenWebUI) May 23, 2024
While looking for an alternative to the above, I've stumbled on e2b and this account just pours out use cases for it. Here's one example.
Code Interpreting for AI Apps 🧠
— Tereza Tizkova (@tereza_tizkova) May 22, 2024
The SDK by @e2b_dev gives you a secure environment to run code. Users build AI data analysts, generative UI, AI "Devins", AI tutors, or reasoning engines for LLMs with it.
☑️ Python & JS SDK
☑️ Runs on serverless and edge functions
💯% open… pic.twitter.com/KLRQd72X3a
Models
Never heard of Smaug but beating Llama3-70b deserves a shoutout.
Smaug - the best open-source model in the world, rivals GPT-4 Turbo.
— Bindu Reddy (@bindureddy) May 17, 2024
Llama-3 70b was the best OS model till today. Today, we are happy to drop a significantly better model, Smaug-Lllama-3-Instruct.
First, with Smaug, we see a significant improvement in the MT-bench. This score… pic.twitter.com/7jKyCmmfba
Memory
Memory feels like it's waiting for innovation before we settle in to the 'right' way to do memory for AI. Examples for me that fall short so far: I want my AI to update it's own system prompt - ChatGPT does this but I'm not seeing how to do it locally yet. So it's cool to see LlamaIndex trying to move the needle on longterm memory:
New Webinar 🚨: Memory for Autonomous Agents 🧠
— LlamaIndex 🦙 (@llama_index) May 20, 2024
This Thursday 9am PT we're excited to host the authors of memary - a fully open-source reference implementation for long-term memory in autonomous agents
In this session, we'll do a deep dive into the project itself and also have… pic.twitter.com/JCbFbqhZkU
New RAG methods pop up all the time, here's one that claims to emulate how the brain stores long term memory. Hype phrases aside, it's up to 30x faster and 13x cheaper:
📣📣 Super proud to present the most exciting project of my PhD so far: “HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models”.
— Bernal Jiménez (@bernaaaljg) May 28, 2024
HippoRAG, as the title suggests, is a brain-inspired RAG framework that enables LLMs to effectively and efficiently… pic.twitter.com/f58fMxMRbh
This looks incredible:
The future of RAG depends on incorporating deep visual understanding in the indexing/retrieval process, beyond just naive text parsing 🥽
— Jerry Liu (@jerryjliu0) May 22, 2024
In this cookbook, we use GPT-4o to process a slide deck; not only do we parse the text, but we also translate charts/figures into tables that… https://t.co/J2EmRepzDY pic.twitter.com/xN6DUk33Gp
This looks cool for studying content strategy on websites:
🎉 GraphRAG in @graphlit
— Kirk Marple (@KirkMarple) May 21, 2024
Learn how to extract a knowledge graph from your content, using LLMs, and enrich entities with Wikipedia.
Use the KG to provide greater context to your RAG pipeline, with GraphRAG.
Blog:https://t.co/0C4D4fbHX6
Code:https://t.co/dviGxtQjh5
Demo:… pic.twitter.com/4Wcl4aRVI5
CLI tools
I have folders on my computer called 'To Tidy' that also contain folders called 'To Tidy', some of which contain the same damn files as others in the hierarchy. It makes Inception look like a Peppa Pig plotline. Imagine my excitement seeing this tools that self-organises your files privately with Ollama... then imagine my disappointment when I can't get fastapi
to work. If you want to try your luck, here's the link.
Not the first CLI interface for AI, but credit here for designing a more user friendly approach to the interface:
Elia - an open source terminal UI for interacting with LLMs
— Rohan Paul (@rohanpaul_ai) May 21, 2024
Can interact with models running through ollama as well as proprietary. pic.twitter.com/XFdhP6Yyov
Websites
Websim is the coolest experimental space in AI today. Infinite generative web. It deserves it's own blog post but as one example of being blown away by the ingenuity of hacking here, someone managed to get it to run an entire MS-DOS including live game execution by pulling live files from elsewhere. Incredible!
I was exploring imaginary filesystems to help the hallucinated apps talk to each other, when I discovered a simulated web URL that emulates a DOS environment
— Scratch (@DrBriefsScratch) May 7, 2024
After some poking I managed to entangle the simulated-emulated filesystem with a real one, and can now virtualize a world… https://t.co/fjjOrsR2dL pic.twitter.com/YAoJtr62cX
There's a lot of research assistants popping up. I'm building my own in Flowise but some days I see something like this and give up:
First look of the Open Research Assistant tool. I hope it will make you a 10x researcher. pic.twitter.com/C3UukNANmY
— Yifei Hu (@hu_yifei) May 22, 2024
Yoink
Kiri has a great feed and managed to pluck the new GPT-4o system prompt:
I was curious, so I found the GPT-4o iOS system prompt:
— Kiri (@Kyrannio) May 20, 2024
“You are ChatGPT, a large language model trained by OpenAI, based on the GPT-4 architecture.
You are chatting with the user via the ChatGPT iOS app. This means most of the time your lines should be a sentence or two,…
Prompt templates are gimmicky but I always grab them incase. The full paper might reveal some underlying patterns to learn from too.
This paper introduces 26 guiding principles designed to streamline the process of querying and prompting large language models.
— Rohan Paul (@rohanpaul_ai) May 26, 2024
Paper - 'Principled Instructions Are All You Need'
Tested on Llama 1/2 (7B, 13B, and 70B) and GPT-3.5/4. pic.twitter.com/IsF6fmoiuX
Talking of prompt optimisation:
Sure AI has saved me several hours by automating my tasks, but now I'm spending a tonne of time trying to optimize those prompts: can't AI do that for me too?
— James Phoenix (@jamesaphoenix12) May 11, 2024
Learn DSPy Prompt Optimization: https://t.co/xwMRFYSuPJ pic.twitter.com/AIH3Hrwxv9
And more on DSPy evaluations from James' colleague Mike. (These two guys are literally releasing the book on prompt engineering, out this month on O'reilly).
Damn ok my latest DSPy program jumped from 70% to 94% on my eval score. It did spank a lot of tokens but worth it to not have to manually optimize the prompt. https://t.co/jMiwGHYf9D pic.twitter.com/N9eUlqXA4W
— Mike Taylor (@hammer_mt) May 7, 2024
Opinions
People saying stuff about AI that should stay on the record, whether they're right or wrong.
Musk pledges to always release an open source model that always beats the current open leaderboard champion:
Source video pic.twitter.com/lrgxfMVyTu
— Ryan Monsurate (@ryanmonsurate) May 24, 2024
Peter Thiel making me feel good about my amateur maths and coding skills:
Peter Thiel: AI is bad news for people for math skills and society is going to shift in favor of people with strong verbal skills. pic.twitter.com/4bfiMTJrbh
— Zain Kahn (@heykahn) May 26, 2024
Why did Roon delete this?
Roon commenting on the situation @OpenAI pic.twitter.com/pdI87UssBW
— Mark Kretschmann (@mark_k) May 17, 2024
The British AI described here is my spirit animal.
American AI. "Can you describe the person you are looking at?" "He is stylishly dressed in a leather jacket and light coloured shirt."
— Rory Sutherland (@rorysutherland) May 17, 2024
British AI. "Can you describe the person you are looking at?" "Dunno, some bald f***er. Probably a Yank, judging by his enthusiasm for corporate… https://t.co/MKnY1fk42t
This random post reminding me I should start building my next agency. Part of me cant bring myself to commercially jump on the bandwagon but everything about AI now is the same vibe as I had with Facebook ads in 2014.
Part time AI consulting business:
— Hadi Azzouni (@hadiazouni) May 22, 2024
- March: 5k
- April: 12k
- May: on track for 20k
- June: already signed contracts for 13k
Solo so far, I use AI everywhere all the time
But context switching is rough - may need to hire soon
If you are excellent at building RAG/LLM apps,…
Auto-agents go bananas with infinite task lists. There's a future for these agents but they're frankly out of control. How can we get them to a stage where we trust their autonomy? Some thoughts here:
introducing mimir, if @rabbit_hmi was a website. has access to tools like the internet and @Spotify web sdk.
— ach (@achyut_benz) May 16, 2024
stack:@nextjs framework@vercel ai sdk@Netlify deploy@GroqInc inference@OpenAI whisper@brave search@shadcn@LangChainAI@bunjavascript
inspiration @Dev__Digest 🐐 pic.twitter.com/xANuR26boc
Yes please to this, alt
text is a pain and should be easily automated away, with a huge market ready to lap this up with a plugin fee:
Who is going to be first to make the Wordpress plugin that uses ChatGPT to write alt text for Media Library?
— Mark Williams-Cook (@thetafferboy) May 16, 2024
Replies to this are real food for thought:
To those of you who claiming that smarter-than-human AI is decades away, what *specific tasks* are you willing to bet me that AI won't be able to do within 5 years? https://t.co/E5StRdsJ3N
— Max Tegmark (@tegmark) May 13, 2024
I'm a mix of excited and disheartened with AI. Most people like me don't have the resources to really move and win in this space. (Must be nice being a Deep Mind Séb?) This tweet captures my feelings before I ignore it and plough on:
Honestly don’t know how you can do an AI startup rn. This is an exponential. You either gotta be moving atoms or providing infra.
— gfodor.id (@gfodor) May 13, 2024
Another reason I'm scared:
your entire company will cease to exist when llama 400b lands. do you really think you can regulate fast enough? do you know how slow the government moves? a single torrent will drop and your entire business model will disappear
— kache (@yacineMTB) May 5, 2024
This thread breaks down how much data is actually available on the web for LLMs to train. Makes the point that Google has an insane advantage with it's Google Books data:
Books are another large source but less accessible source. Google Books is almost 5T tokens, and is available only to Google. It may be the largest proprietary high-quality token source out there. pic.twitter.com/8G9xAFxNYg
— Mark Cummins (@mark_cummins) May 10, 2024
And that's just in May!
Feedback in the comments, and I'll post the same next week for just one week's worth of content. How's that Séb?