Vercel's AI Accelerator
vercel.com
vercel.com
How about you guys fix telemetry, or configs, or finish next 13, or fix the auth layer (its a joke).
That's like 1 (£850k in credits in total for 40 engs) among the 4 things (community, coaching, demo day) Vercel highlight in the post?
We'll also have a community of folks we're providing educational content for, connecting them with other builders, and having speakers come to show new AI tooling.
I'm not sure what specifically you are referring to with finishing Next.js, but possibly the App Router, which is now "finished" API wise with v13.4 (stable core APIs). We don't have a built-in auth layer in Next.js, what do you mean here?
Serverless because it can't work in a serverfull setting?
"Looking at the next.js repository, there are a lot of memory leak issues, often without resolution / response": https://community.fly.io/t/high-memory-usage-in-deployed-nex...
https://nextjs.org/docs/app/building-your-application/deploy...
> RIGHTS IN ENTRY MATERIALS. As between Entrant, on the one hand, and Sponsor and Partners, on the other hand, Entrant retains all ownership of the Entry Materials. However by submitting an Entry, Entrant grants Sponsor and Partners and their agents and representatives a non-exclusive, worldwide, royalty-free, sublicensable, perpetual and irrevocable right and license to use, reproduce, modify (solely for the purposes of transmission), distribute, publicly display, publicly perform, and transmit all or a portion of the Entry Materials in connection with the Contest, for promotional purposes and otherwise, including on and through third-party distribution channels selected by, but not affiliated with, Sponsor or Partner.
It's not like "startup accelerator" has a scientific, stable meaning. And yeah, it's marketing - they're a for-profit firm. It's marketing for all the companies involved.
And what they themselves struggle with is finding use cases to sell to even larger enterprises.
These types of half baked AI accelerators are just ways of crowd-sourcing content ideas.
The killer app I can see is a “personal copilot” where you train a model on some set of documents or code or both and it can give you a ChatGPT-like interface to your domain. But as far as I know, that’s going to be insanely expensive or time consuming if done at a personal level. The UX of selecting the code and documents doesn’t seem straightforward either and seems like something that would need to be done at the enterprise level.
But after that, my imagination sort of dries up.
I think it's going to be really cheap.
Don't fall for the common trap of assuming that you need to fine-tune a model in order to implement Q&A etc against a large corpus of documents.
It turns out it's much more effective (and much easier / less expensive) to do the Bing/Bard/etc trick, where you use an off-the-shelf LLM and run searches (including fancy embeddings vector-database searches) against your content, then feed the top matches into the LLM as part of a prompt that includes the user's question.
I wrote about one way of doing that back in January: https://simonwillison.net/2023/Jan/13/semantic-search-answer... - it's got even easier to implement since then.
ChatGPT functions and ChatGPT 3.5 getting cheaper (and gaining a 16k token context) makes this even more accessible today than it was yesterday.
I’ve also found that GPT-4 tends to hallucinate when I ask a question but the search process doesn’t result in a good match. Like making up an author or title for a document I’m asking about.
Hallucination continues to be a problem, though GPT-4 is massively less prone to it than 3.5. My instinct right now is that you can resolve a lot of that problem for the Q&A case using additional careful prompt engineering, but I don't yet have a recipe I'm 100% confident in myself.
If it could fit within the context length then you wouldn’t need to search it to find relevant portions at all. I’m talking about situations where you want to ask questions about a very long document, using the search approach, and one of those questions might be “please summarize chapters 2 through 8”.
The search method will never accommodate that, but a model fine tuned on your data might plausibly.
please summarize chapters 2 through 8
- Iterative: Read and summarize one `MAX_INPUT_TOKENS - SUMMARY_LENGTH` section at a time, adding to your summary as you go along. The sections could be meaningful like chapters, or just arbitrary sets of paragraphs, depending on the context.
- Intentful: Intuit a context-appropriate set of questions to ask about the text, use semantic search to find and summarize the answers to those, then clean up all the separate answers into one cohesive one.
If your summary is getting too long to consider at once for meta-summaries, just add more layers of the above algos. I.e. combine summaries in a bracket of sorts rather than adding on to a single heap, or search for the relevant parts of the summary to tweak based on a given block of new content.
Ultimately, yeah, one inference will only be able to understand as much information as it can fit in its context window. So if you want it to write a novel while holding every word in the novel up to that point in its head at once with equal weight, that's going to be hard. But
a) context windows are getting longer
b) I have never once seen fine-tuning give that level of context-specific accuracy - it is a language model after all, not a knowledge model! and
c) that's kinda a lot to ask. Could a human even do that?
Hope that rant has some interesting parts :)
For the latest example I've been excited about:
the Doordash app sucks, and we spent 10 minutes the other day trying to convince it to let us order from the Boba franchise in the next town over rather than the one close by (ahh, the wonders of Silicon Valley, a true Eden).
If the app was powered by an LLM with access to a series of API endpoints rather-than/in-addition-to the traditional visual interface, one can easily imagine how such a problem could be solved in seconds. Suddenly, it becomes possible to design interfaces that fit niche/edge/unexpected usecases much more reliably.
rewind.ai is already doing this. Also I think we can personal LoRA(low parameter diff to model) that keeps evolving as we use computer.
I am also really curious to see the projects which will eventually be selected. It's an interesting time we live in and I would argue that before we know it we will use highly AI augmented or enabled services in the future which will provide value we don't even know of yet. We had maps before we had google maps. And we had coaches before we had cars. And yeah, a lot of the AI stuff we see these days is mostly cool toys and lots of buzz to ride the wave but I am convinced we will see some really impactful stuff come out of it.
https://chrome.google.com/webstore/detail/coldmessageai-for-...