Hey folks, and welcome to 6770 of you who joined last week. Last week was bonkers, so I took the weekend to play with my nephew and hang with my family.
The weekend was comparatively quieter, except for Sam Altman saying that, yes, AI will take away many jobs from today’s world. More about AI helping teachers, Stable Diffusion, LLaMA and Apalca in today’s edition.
Let’s get to it.
🤌 Our Picks
Highlights if you've only got 2 minutes
Sam Altman (OpenAI CEO) and Mira Murati (CTO) clearly spend more time practising for interviews than training ML models because they absolutely aced this one with ABC News - and it was tough. Hard-hitting and direct questions, a targeted approach of concern, and a stern tone brought this interview straight to the point and got a lot of info out of the top dogs in AI. Altman is “a bit scared” (naturally); he thinks society needs to be given time to adapt; they will adjust the technology as negative things occur; and we’ve gotta learn as much as we can whilst the “stakes are still low”. And Murati is just about the most calm and collected person I’ve seen on screen - check the two of them out in the YouTube vid linked above!
I know there’s a lot of chatter at the moment around scepticism of how much OpenAI actually cares about safe releases and AI alignment, but I get the impression they really do. According to Altman, GPT-4 was finished 7 months ago, and since then all their focus has been on getting it to be as safe as possible. But perhaps I’ve been totally hoodwinked…
Our resident DJ for advocating integrating chatbot AI in education is back on the scene, but this time he’s focussing on how they can help teachers, rather than how to enhance a student’s learning. In his latest post, Ethan Mollick outlines 5 key aspects of teaching where chatbot technologies can be utilised to both boost teacher productivity and improve their approach. Fit with the perfect prompts to guide the outputs, he runs through example/case study generation, AI-powered technical explanations, low-stakes test building, and more. All things that teachers can develop alongside the AIs, helping both themselves and students. Bring on more AI in education I say!
With my bedroom decor starting to look very tired, I could really do with an assistant to help reimagine how it could look so I can spice things up a bit. And it looks like Stable Diffusion’s latest work with Clipdrop could help me do just that. Their new tool, Reimagine, acting as a kind of image-to-image model, takes an input image and ‘reimagines’ how it could look, producing new images that are essentially modifications of the first. You can even select detail enhancements, pushing viewers to focus on an aspect of an image that wasn’t quite capturing their attention before. Looks pretty cool! I can see applications in portrait photos, architecture, design, and a heck of a lot more.
For all you AI hacky types out there, you’ve either already heard of llama.cpp, or you haven’t and need to hear about it ASAP. For those of you that don’t know, this developer in Bulgaria put it together to compress Meta’s LLaMA models into a far more manageable size for your home laptops - a 4-bit quantization approach for your Macs. Big jumps towards democratising AI! And according to him, the llama.cpp repo is buzzing with activity, with double-digit numbers of pull requests in the space of an hour on a Sunday.
It’s also reaching other projects too. These folks have taken the LLaMA foundation model, Stanford’s Alpaca chatbot model, and then messed about a bit with llama.cpp to produce alpaca.cpp, giving us a ChatGPT-style model with just 4GB of weights. (That’s small by the way…)
Normal makes building enterprise-grade GenAI apps hassle-free i.e. normal. (Sponsor)
Customise open and closed-source models (HuggingFace, OpenAI, and more) for:
Personalised and adaptive semantic search apps
Automated prompt engineering over model chains
Multimodal (images, video), and uncertainty-aware apps
Thermodynamic Hardware Advantage
Prototype, A/B test in prod, and scale on-cloud using a visual UI. Sign up for the platform waitlist, and fork from the examples!
Check out their job openings.
🛠️ Cool Tools
Product launches, updates and demos
Stable Diffusion Reimagine - New Clipdrop tool to generate multiple variations of a single image. (link)
ViperGPT - A python framework that combines vision and language models using code generation to achieve state-of-the-art results. (link)
Bloop - Understand your codebase with GPT-4 and semantic code search. (link)
If you don't have access to GPT-4, you can try it for free with Replit. (link)
Wolverine - Run your python scripts, and when they crash, GPT-4 edits them until it is fixed. (link)
Alpaca CPP - Locally run an instruction-tuned chat style LLM. (link)
GPT repository loader - Convert code repos into an LLM prompt-friendly format. (link)
Free AI-powered subtitle generator from Animaker. (link)
Praktika - Immersive language learning app with generative AI Avatars. (link)
NPM picker - Find the best npm packages for your project. (link)
Gradio GPT-4 demo for anyone to access without needing an API key, thanks to Hugging Face. (link)
Demo of a ChatGPT-style chatbot using LLaMA deployed on Baseten. (link)
Gradio demo of CoAdaptor - Use style, colour, depth and other inputs at the same time to create customised images. (link)
News, podcasts, videos, blogs etc
GPT-4 is already being used for work in startups, replacing manual coding. (link)
I asked GPT-4 to write a biography of me. (link)
This week in AI doublespeak. (link)
Lessons learned from conducting 100+ ChatGPT user interviews. (link)
Deepfake ‘news’ videos ramp up misinformation in Venezuela. (link, without paywall here)
GPT-4 from OpenAI shows advances—and money-making potential. (link, without paywall here)
Strangeworks might be the first startup to rely on AI to create everything it brought to SXSW. (link)
Women in AI - Join this list of women interested in and building in AI. (link)
A discussion with the creators of Vesuvius challenge - Using AI for understanding old and damaged texts. (link)
ChatGPT is the Myspace of AI - It's fun, but it's far from taking your white-collar job. (link)
We are underestimating the effect LLMs will have on the nature of software distribution and end-user programming. (link)
How-to’s and resources
How to create an OpenAI Q&A bot with ChatGPT API + embeddings. (link)
👋 Too many links?! I created a database for all links mentioned in these emails. Refer 1 friend using this link and I'll send over the link database.
Published research papers
Vid2Seq - A visual language model for dense video captioning. (link)
1.7 billion parameter text to video generation diffusion model from Alibaba. (link)
DS-Fusion - Artistic typography via discriminated and stylized diffusion. (link)
HIVE - Harnessing human feedback for instructional visual editing. (link)
FreeDoM - Training-free energy-guided conditional diffusion model. (link)
GlueGen from Salesforce - Plug and play multi-modal encoders for X-to-image generation. (link)
A recipe for watermarking diffusion models. (link)
CoLT5 from Google - Faster long-range transformers with conditional computation. (link)
CHAMPAGNE - Learning real-world conversation from large-scale web videos. (link)
GPTs are GPTs - An early look at the labour market impact potential of large language models. TLDR - Approximately 80% of the U.S. workforce could have at least 10% of their work tasks affected by the introduction of GPTs, while around 19% of workers may see at least 50% of their tasks impacted. (link)
ElasticViT from Microsoft - Conflict-aware supernet training for deploying fast vision transformers on diverse mobile devices. (link)
Dual-path adaptation from image to video transformers. (link)
LOCATE from Google - Localize and transfer object parts for weakly supervised affordance grounding. (link)
αSurf - Implicit surface reconstruction for semi-transparent and thin objects with decoupled geometry and opacity. (link)
Short, sponsored links
Deepgram is hiring a research scientist. (link)
Replit is hiring a technical writer. (link)
Potion AI is hiring a senior frontend engineer. (link)
FoundationCap is hosting an exclusive hackathon for a select batch of founders building in AI. (link)
Clem Delangue, Co-founder and CEO of Hugging Face will be talking with Elad Gil on March 27th in SF. Register here to attend live.
GPT-4 hackathon to be hosted by Cerebral Valley at Shack 15 in SF on March 25th. (link)
Have a product, service, job, event, newsletter, app, book, movie, tool, or anything you'd like to share with over 50k subscribers?
Advertise | Jobs | Community |
Investor + Founders | Maker-in-Residence
🖼 AI Images of the day
Funny memes and pics from around the web
Send this to 1 AI-curious friend and receive my AI project tracker database! Use this link.
How was today's email?