Parker Rex
All videos
Parker Rex

Did AI Just Get Perfect Memory? (Huge if True - Copy This)

November 22, 2024

Watch on YouTube@parkerrex

Pushing the envelope on AI tooling, this update covers new SQL copilots, frontend-cloud IDEs, code search acquisitions, memory-focused models, multi-agent workflows, and more. Here are the key takeaways and what to try this week.

SQL copilots for Postgres in Supabase#

  • Acknowingged feature: a SQL copilots trained specifically on Postgres that works well even if you’re not using Supabase as your backend.
  • Benefits:
  • Schema-aware suggestions if you’re using Supabase (great for targeted queries).
  • Quick, practical helper for daily SQL tasks and exploration.
  • Actionable takeaway:
  • If you write a lot of SQL, test this Postgres-specific copilot in your environment to speed up query writing and data exploration.

VZO (frontend-cloud style tool) features and implications#

  • Multi-page file editor: you can create multiple “pages”/files that the tool writes (SVGs included), giving you a dashboard-like workspace.
  • Deploy integration: five-shot a website, then deploy directly; the tool ties prompts, edits, and deployment together.
  • Live preview and navigation: URL bar, development preview, and a “select and inline edit” workflow (open in new tab, fullscreen, refresh).
  • Standalone and deployable: it’s positioned as the frontend cloud, aiming to let you build, test, and ship by prompting.
  • Grep acquisition: they bought Grep for fast code search, planning a standalone product with API and an integrated search engine.
  • Practical use-cases:
  • Rapid prototyping: generate and deploy UI pieces from prompts without leaving the tool.
  • Quick code discovery: search across repos, prep for forking or contributing more efficiently.
  • Actionable takeaway:
  • If you’re doing front-end work, experiment with VZO’s multi-page workflow and the new code search capabilities to speed up iteration loops.

Code search and repo discovery: Grep and the broader search play#

  • Grep integration: fast code search across your stack; faster than traditional GitHub search in some contexts for specific queries.
  • standalone tool and API: designed to let you locate, fetch, and start editing repos quickly.
  • Practical use-cases:
  • Find a repo to fork or contribute to without manually digging through dozens of results.
  • Jump straight into an existing project with a quick search that brings the repo into your workspace.
  • Actionable takeaway:
  • If you regularly explore new codebases, add this search capability to your toolkit to speed up starting new side projects.

Bolt.new and inline editing patterns#

  • Bolt.new mentioned as a popular entry point for folks new to coding; somewhat standalone, not deeply integrated for developers yet.
  • Notable detail: code editor components (e.g., CodeMirror) are used in some examples, so you can learn how inline code editing is wired in these tools.
  • Opinion: useful for exploration and getting people excited about code, but less of a day-to-day dev workflow for seasoned builders.
  • Actionable takeaway:
  • If you’re curious how inline editing works in practice, check out Bolt.new’s approach and the CodeMirror integration to understand the building blocks.

Gemini memories and the memory arms race#

  • Google Gemini launched “memories” with a larger context window, enabling more persistent context across interactions.
  • Microsoft reportedly claiming near-infinite memory capabilities (in their announcements).
  • Why it matters:
  • Larger context windows and persistent memory can reduce repetition in multi-turn tasks and improve long-running workflows.
  • Actionable takeaway:
  • For long, multi-step tasks (like building a video description pipeline or multi-page prompts), explore memory-enabled models to see if context retention reduces prompt churn and improves consistency.

Other AI copilots and conversational tools#

  • Mistral’s LayChat: a competitive model with citations, designed to be strong in a chat/QA flow but still aiming for profitability in deployment.
  • Perplexity added one-click checkout and expanded shopping features.
  • General takeaway:
  • Expect more “context-aware” and chainable assistant experiences that integrate search, pricing, and checkout into AI-assisted workflows.

Windsurf IDE and Cascades by Codium#

  • Windsurf IDE (Cascades) pitch: a code-indexing-first IDE that helps you answer questions about your codebase quickly.
  • Codium background: known for indexing and fast code-base queries; aims to accelerate understanding large codebases.
  • Actionable takeaway:
  • If you work with large codebases or monorepos, try Windsurf/Cascades-style workflows to test how quickly you can answer code-base questions and navigate dependencies.

XAI and multi-agent workflows in Verel AI SDK#

  • XAI integration in the Verel AI SDK (Verel likely referring to the platform in the talk).
  • Mistal (mistl) approach: one pattern is to tag multiple agents and compose them into a workflow.
  • Human-in-the-loop example: one agent drafts a YouTube title/description, passes output to a blog-post agent, and prompts for your feedback.
  • The big idea:
  • Multi-agent flows without requiring heavy coding could become a standard pattern for end-to-end content generation, data pipelines, or deployment pipelines.
  • Actionable takeaway:
  • Experiment with a small multi-agent chain (e.g., video outline → title/description → blog draft) to see how well the hand-off works and where you want human review points.

Practical takeaways for developers and builders#

  • Build fast: leverage the new copilots and multi-page tooling to prototype and ship faster.
  • Search-first workflows: adopt fast code search and repo discovery to reduce friction in starting new projects.
  • Memory-aware workflows: for long-running tasks, experiment with memory-enabled models to maintain context and reduce prompt overhead.
  • Multi-agent workflows: start with simple agent stacks and a human-in-the-loop for quality control; scale up gradually as you validate outputs.

If you want, I can tailor these notes to emphasize a particular topic (e.g., code tooling, memory in AI, or multi-agent workflows) or adjust the level of detail for a deeper dive.

Transcript

hey everybody welcome to the show today we're going to be covering a lot of topics quick so you will be up to date on everything that happened in AI that's worth knowing as a developer as a builder as someone that uses these tools so the first one I want to call out this is a for people that are using SQL and this has been out for a little while but it just keeps getting better and better so if you're writing SQL and you want to have a co-pilot for that this is specifically trained on postgress and it's fantastic it's in superbase you don't even have to be using the super base backend as a service and if you are using that then you get to select the schemas and it'll give you more specific stuff but I found that very helpful second one is vzo came out with their really it's it's a competitor to bolt. new where it allows you to have multiple Pages created so I two shotted this which is a big dashboard and it has some cool stuff in it to keep you updated but it'll lets you actually create all these different files so what we have 15 maybe different files in here and it's writing different SVG stuff which svgs typically are very Troublesome for llms because there's a lot of numbers involved but this is a big deal it's going to be really helpful I think it's way better because they're actually tied to deploys so you could go and you could five shot a website and then click deploy and it's just tied in right there it also has this URL nav bar which looks really good you have the development preview so you can see where this is going where because they own the you know they kind of position themselves as the frontend cloud it's getting closer and closer to that well you'll be able to build build build build build by just prompting test it on here and then either share it or deploy it you can also do this select thing so you can say hey I want to make this inline change you can open this in new tab you can full screen it you can refresh it you can see where this is going it's getting a lot better you can also just do this this has existed but this is going to be a big deal in the long term and another piece is they actually bought a company called grep so grep is really fast code search it's not as good in terms of the breadth of GitHub search so if I did a search for nextjs it says 55,000 results here but if I did it on GitHub it's like 450,000 and just some of the filtering isn't quite as good but what they're doing is they're taking this and they're going to have it as a standalone product let's see if I can find the right Tab and it'll have a standalone tool an API and integrated search engine to vzer Inver cell so really fast search across the whole stack that's great news and the way that I see this going is you could probably use it to find a repo that you want to fork or you want to add to so you're not starting from scratch so if I said I wanted to work on this iron Cal thing which if I remember correctly it is a spreadsheeting software written in Rust if I wanted to take that and Fork it and work on it then I could do that I could search it really quickly it could bring it into v0 I could probably see the whole repo there and then start editing it that's interesting the bolt. new which I kind of talked about a lot of people have talked about this I don't really like it that much because it's kind of Standalone and I feel like if you're actually a developer it's not that helpful but it's great for people that are getting excited about coding it's actually open source at least part of it and it's worth checking out it's getting a lot of stars it's been out for a little while but it just shows you if you were to make agent stuff and you wanted to see like well how does this code thing work on the right how do they bring this in how do they allow inline editing well it turns out you can go and find that component and you can see oh they use code mirror what's code mirror and it lets you just learn so that's a really helpful one another the piece is actually on Gemini so Google Gemini launched memories we're used to seeing memories in open AI but the difference is the context window is much larger so that is probably something to keep an eye on and as well as Microsoft they did a big announcement saying that they're going to have I think they called it like infinite memory they basically said that they're close to having infinite memory and of course I can't find it but Microsoft's making those claims other things that are happening are mistro launched lay chat used it it's what all the other ones do just a little bit worse has citations and stuff it is funny that they are having a mission to develop highly competitive models while ideally making money in the process I think that's just is a funny way of of positioning it and then perplexity added one uh click checkout they also added a bunch of shopping stuff so I don't know what that means for the infamous company called bolt which was that guy's kind of a Shyer something that I haven't used so I'm not comfortable talking about it very much but this wind surf IDE is supposed to be really good because they have this thing called Cascades and it's built by a company called codium codium is known I think they started with just rag for your code base so answering questions about your code base and being really good at that so they're really good at uh basically indexing everything so I'm excited to try this out there's a lot of really good feedback online about it but because I haven't used it I don't really want to talk about it I've talked about zad in the past and I think that that's really promising because just the performance of it and it doesn't get in the way xai now is in the verel AI SDK we talked a little bit about this mistal thing it has canvas one pattern that I think will be interesting and other people will do is you can tag agents so if you had multiple agents you can bring them in here and have them probably do different tasks so you can see a world where you're stringing those together and they're doing the different tasks that you want and it looks something like this pattern where instead of it saying the name of the component at the top probably be the name of the agent probably the steps that it took and then you can have a human in the Loop where if you had an agent that's job is to write descriptions for this YouTube video and titles then it would have the output and then that would hand that to a blog post agent so it could see what the title and description was and then draft something for that and then ask you for the feedback but I think that's that's likely where stuff like that will go when it comes to multi- agentic flows don't require coding that anyone can jump in and that's going to be a game changer so as always if you enjoy the video make sure you like it and subscribe to make sure you're in the loop on all this stuff that's happening every week in the quickest way possible and I'll see you in the next video