Ben's Bites
Posts
Is multimodal the next step?

Is multimodal the next step?

PLUS: Use LLMs to improve LLMs, IBM Watson's comeback

May 10, 2023

Newsfeed | Search | Advertise | Community | Maker-in-Residence

Hey folks, today we have another banger from Meta. Imagebind, a new model from them has 6 senses - text, audio, visuals, movement, thermal, and depth.

The twin stars, Open AI and Anthropic are both pushing LLMs to their limits to improve LLMs to understand them and make them safer.

And our old AI friends from IBM are trying to revive Watson in the generative AI wave.

So, let’s get to it.

From our sponsor:

Meet Fin. Intercom’s breakthrough AI bot for your support team, powered by GPT-4.

Fin can resolve up to 50% of your support questions instantly, solving complex problems and providing safer, more accurate answers than any other AI bot. Simply pair Fin with your help center to learn your written history and hold natural conversations with your customers.

Join the waitlist today.

🤌 Our Picks

Highlights if you've only got 2 minutes

Meta is taking us into the 6th dimension with ImageBind

Whilst the rest of the world has been focussing on building language, image, video, and audio models, Meta has been focussing on propelling us not into the 5th, but the 6th dimension. Their research teams are publishing and open-sourcing results for ImageBind, an AI machine that binds 6 different modalities. This means that you could describe a rainforest with text and it’d be able to visualise it, create the sound of rain, understand its depth, map thermal imaging, and appreciate motion readings. Amazing research with far-reaching applications if it can develop further. They couldn’t resist connecting it to virtual reality though… Classic Zuck.

OpenAI is using language models to understand their own behaviour

OpenAI’s latest adventure involves using language models to understand… language models. Taking an advanced, large model like GPT-4, they’re using it to use natural language to explain which parts of less powerful models (GPT-2) are responsible for certain outputs. This means that each parameter ends up with an associated explanation for its contribution to a given output. This might seem odd, but it could help us move towards more explainable and understandable models, helping to align them more closely to safe, human values.

TechCrunch coverage here for a brief review.

Anthropic is backing Constitutional AI over RLHF

There seem to be two schools of thought in the world of AI right now: RLHF vs Constitutional. And it’s now officially equivalent to OpenAI vs Anthropic, the firm of ex-OpenAI employees looking to create more ethical AI. Where RLHF relies on humans to read model outputs and rate/align them to human desires, the constitutional approach does away with humans. Instead, it gets given a ‘constitution’ of values and uses a language model to assess how much another language model’s responses align with the constitution. It’s more scalable, and doesn’t require cheap human labour, and Anthropic thinks it’s the way forward…

A detailed write-up on how Constitutional AI can be RLHF on steroids. (link)

Scribe brings the AI tool for documenting processes

Let’s face it, most senior jobs involve lots of documentation. Scribe’s platform uses GPT-4 to automatically create SOPs, help centres, new user guides, and process overviews for any business process.

These step-by-step guides are complete with screenshots and text by capturing your screen while you click and type. In reverse gears, Scribe AI can create full process documentation (including headings, subheadings, and detailed text) with your guides automatically embedded.

No more staring at a blank document thinking, "ok, I have to teach someone how to do this, where do I start?". Scribe AI does it for you.

Ps: I was on the Cognitive Revolution podcast with Nathan and Robert Scoble to discuss scouting AI companies and future developments. (link)

🛠️ Cool Tools

Product launches, updates and demos

Check out BB News for more →

StarChat - An open-source ChatGPT-like model to answer all your coding questions. (link)
Gmail agent toolkit - Enable agents to search Gmail, retrieve, draft, and even send messages. (link)
LLMTown - A cheaper and simpler alternative API for semantic search. (link)
Drumloop AI - Generate an original drum loop with neural audio synthesis. (link)
Embedding Store - Hosted embedding marketplace with public, private, and third party data. (link)
Chatcraft - Developer-focused open source ChatGPT. (link)
Bugasura AI - Unlock the impact of a bug on your software and your customers. (link)
Briefly AI - Transform your meeting transcripts into polished documents and post-meeting deliverables. (link)
Peridot - Unique AR pets that stay by your side. (link)
Total crap - A magazine written by AI. (link)
AI background changer - Realistic AI backgrounds generated for your product photos. (link)
Where to AI - Discover new destinations, create unforgettable memories, and find the best places to stay. (link)
Check out BB News for more →

🤓 Miscellaneous

News, podcasts, videos, blogs etc

Check out BB News for more →

Bessemer Venture Partners is committing $1 billion to invest in AI startups. (link)
Google Cloud’s generative AI partnership with popular workplace apps. (link)
Accelerating tech adoption with AI. (link)
AI is not good software. It is pretty good people. (link)
Palantir stock soars 21%, says demand for AI unprecedented. (link, without paywall here)
IBM intros a slew of new AI services, including generative models. (link)
With seed funding secured, AI detection tool GPTZero launches a new browser plugin. (link)
Ascend raises $25 million for pre-seed AI startups in the Pacific Northwest. (link)
Eventbrite integrates GPT capabilities into the platform to aid the event planning process. (link)
What made Hinton into an AI doomer. (link)
TidyBot - Personalized robot assistance with large language models. (link)
FrugalGPT - How to use LLMs while reducing cost by 98% and improving performance. (link)
Check out BB News for more →