I suspect that you didn’t mean it this way, but the minute Coda starts using my personal documents to train its public LLM/AI is the minute that I am no longer a Coda user.
2 Likes
If that was your interpretation, you are correct; that’s not what I as intimating. The idea is that Coda document(s) possess vast amounts of knowledge and owners (of those documents) should be able to blend LLMs with their documents for their own benefit.
LLMs alone are pretty dumb; most queries are south of 40% accuracy. When blended with your domain of knowledge, accuracy rises to north of 86% and costs drop to 1/4th that of the poorly performing LLM alone.
Everyone - and by that, I mean even my slightly annoying very non-technical brother-in-law, are able to use GPT to write stuff. There is no defensive moat; the drunken stupor will not last forever. And when sober times arrive, only the vendors who have …
- Baked workflow and productivity into the ML solutions;
- Baked user content into the ML solutions;
- Lowered costs in the ML solutions… will be standing on solid AI benefits.
2 Likes
Let me say, I’ll be happy to use it where I’ll need the data that looks convincing — but NOT necessarily correct 
That’s my experience with AIs so far anyways
Not belittling the value it will bring though.
2 Likes
Indeed. And given that Coda struggles with full-text search and a number of other findability challenges, the combination of LLMs, embeddings, and fine-tuned domain-specific variant models would solve for x-squared; AI that negates the need for search while making intelligent discoveries in reach through natural language.
Everyone has GPT; only you have your data. This is the deeper value proposition that LLMs make possible.
4 Likes
Will it be possible to use it in other languages as it can be done in other AI?
Just bit of feedback. I’d really like a Sync Table ‘ChatBot’ to work through a list of task dependencies, to chase and be proactive with other users updating records through NLP. This would be a lot more practical ‘AI’ use-case then most of this videos. The table creation thing does look cool though.
I want to be in the alpha!
2 Likes
This has to be one of the top AI demos I’ve seen recently, both from a utility and from a UX perspective. Well done!
4 Likes
My goodness the potential is beyond overwhelming. Amazing work Coda Team!
Hi Sarah – thrilled to hear of your interest! We’re in early stages, and we will be starting with a small group, but rolling out access over time, so we can best test and iterate based on feedback. One of our goals is to better understand how Coda AI can unlock new workflows or to help makers work faster, so we appreciate any thoughts you have!
Great to hear it! You can learn more and sign up for the alpha waitlist here. The form is about halfway through the page – if there’s any scenarios you have in mind where AI could be valuable, feel free to share those.
Hi Laura, wanted to point you towards our terms of service addendum, specifically:
We do not permit our third-party providers of the AI Functions to use your User Content to train their AI models without your consent.
Hope that helps, and thanks for all your feedback!
3 Likes
Loved it! looking forward to start testing it!
1 Like
There were some really interesting things in the demo, for example the email interrogation. When I worked on SAP, my unread emails would increase by fifty a day during busy periods. And finding the right email was a schlepp.
But I got cold shivers when I saw the AI generated blog posts… The world is already drowning in information, can you imagine 10,000 people cranking up our blog posts at the speed of white light?
I do think that the best use cases for AI is still to be determined. Both in the corporate and the personal area. In certain industries, for example medical care, it is going to be a game changer. AI is already better than many doctors in diagnosing breast cancer. Robotic surgery, even across continents, is a growing concept. AI can augment also the knowledge throughout the hierarchy, from the least trained carer, to the highest trained specialist.
And as far as information worker productivity goes, Coda is going to be right there, allowing us to build even more amazing tools.
It’s just a ramble…
1 Like
I can’t speak for the vast world of AI use cases, but at Stream It, we already use AI to generate questions to anticipate what customers will likely ask about new products and services. We use it to summarize email messages, and tag them with keywords and entities present in the message. This in turn creates more data we use for training with embeddings.
For us, many vectors into the benefits of AI arose years ago and now we rely on them. The advent of GPT is a bit of old news to us as we’ve been on these early models since OpenAI launched in 2020. You don’t need to use ChatGPT to derive great benefits of the underlying models. And we don’t regard the models themselves as the limiting factors. Instead, we create fine-tuned instances and also use embeddings to create AI services that meet very specific domains of work.
It’s really hard to make a festering cesspool smell any worse.
Let’s be clear; the world was long down the pathway to crappy content before GPT made it possible to do stupid stuff that might be misinterpreted as less stupid. Bots were already responsible for 90% of the crap and false information we see on the Interwebs. A few billion more bad artifacts will not make it worse; it’s already a very smelly cesspool.
2 Likes
Correct me if I’m wrong about this, but I get the sense that eveything Coda has built into this AI demo can be built by anyone using GPT (models and APIs) as well as Packs, and formulas. Is that correct?
I ask this because I’m trying to understand what’s the difference between the big bag of Legos that represent GPT, Packs, formulas, and other Coda features and this new Coda AI product except, you packaged some of the common ideas into a working template?
2 Likes
Hi @Katy_Turner !
I have a question! Will this AI tool understand tables made by users in the same doc or can generate just new content based on external data’s?
Off course i would hope it is able to integrate with itself (i mean the doc in which is in)!
This will make doc final client able to generate extra tables or interconnect them based on new needs without the original maker help! THIS will be a game changer! (no dev/maker needed anymore for new features!)
Or, also create button that does edits in already in-place tables!
Please don’t tell me it cannot 
This is a great question that dovetails with this comment and this one.
I think the answer is the latter; it creates new data from external information. However, there are some exceptions. I think I saw AI processes running against Coda data to produce more Coda data. We have been doing that in my company for over a year. Packs made this possible in 2022 and webhooks can provide some intersting ways to use the Coda API to churn on Coda tables to embellish documents and tables with additional data.
What I do not see in the demo is the ability to use Coda-based data to train models such that they are more able to intelligently know how to use your own data to formulate useful outputs.
2 Likes
Thanks for the question! Stay tuned — we’re going to be experimenting and iterating over time, and will share updates as we learn more!
1 Like
Any updates on this?