eob

MLE | training data

Just writing hand wavily :)

As for practical use cases, one is to find an approximate optimization to a function

- You want to find the min/max of some probability distribution P(x)

- P(x) is too complicated to find a closed-form min, but you can draw samples from it.

- So instead, you carefully construct some OTHER probability distribution Q(x|θ) that you claim is structurally similar "enough" to P(x), parameterized by θ.

- Now you find the theta which minimizes the KL divergence KL(P(x) || Q(x|θ)), which is equivalent to delivering you the parameters of θ to Q(x|θ) that make it [approximately] "most" similar to P(x) without ever having minimized P(x)

It was a trick that came up a lot when AI consisted of giant Bayesian plate models for each specific task that you had to hand-optimize.

There used to be a hack to get around password protected screen savers in which you’d do the following (IIRC):

- Drag the password window around the screen furiously

- Press Ctrl-Alt-Del as you did it

I think it would just dump you straight to the desktop.

I remember doing it in Circuit City as a kid to access computers.

ChatGPT is amazing but ultimately unwieldy for directed, long running relationships more nuanced than general themed chit chat.

We’re in the nascent stages but I think there will probably always be a community of folks who want to add more nuance to the communication, whether it’s reveries that enact a mood or goal, tie-ins to other services, etc.

Eg imagine wanting to have your ChatGPT DnD master also keep some kind of score. It may be ultimately easiest to put a wrapper around a themed GPT window that imposes a predictable way to do that rather than require everyone to figure out how to prompt it correctly.

When it happens, I think it’ll happen on our phones first.

I think we’re in platform reliance mode for quite a while. I think of all of these SaaS companies as different icons on an abstract AWS console that doesn’t exist yet.

We wouldn’t bat an eye at using S3, EC2, and RDS as a host your own setup. The only difference here is that startups are moving faster than incumbents.

FWIW that’s one reason why Steamship (disclaimer: I’m the founder) aggregates all AI services under a single API key and interface. It’s to deal with the insane glue-code hassle of running this stuff on your own.

Is there any place where folks building AI dev tools at the [no code, low code, prosumer-code] intersection hang out?

I was just watching some of the Windmill videos (they're cool!) and am fascinated by the interplay between visual code blocks and code.

Selfishly, I wish there was a water cooler where those of us wrestling with these dev ux interfaces could trade tricks & hazards.

+1 to this from the Steamship perspective.

We're a hosting platform for AI Agents and in practice that means keeping track of a stream of usage events across agents, microservices (equivalent of Fly's VMs), and third-party services (OpenAI, Eleven Labs, D-ID, Pinecone, Replicated, etc) that roll up into an itemized consumption bill.

It might be this is just a super-niche billing situation, but we also haven't yet found an open-source solution that fits. Fully agree that a query layer atop a time-series database is the style of product that we're craving.

Metronome comes very close, but since it's not open source we're hesitant to make the jump before it's a perfect fit.

Congrats on the launch! It looks very cool!

100% agree with this.

The Agent & Tools metaphor makes a ton of sense for its simplicity, but has yet to scale beyond very simple agents.

I’m long on it as a programming model, but a lot of work is needed.

As a complete non-gamer, the ability to walk around the earth with a VR headset would justify buying one to me.

There are so many cities I've always wanted to walk down the streets of!

This is really neat. Do I need to give access to my data or just the schema?

I lost a decade of family photos using a Western Digital RAID-in-a-box setup (WD MyCloud or something)

Both hard disks fried at the same time.

More and more I'm seeing the playbook of:

1) Develop ops-related software 2) Package it in a single Docker image 3) Let folks spin it up on-prem

I think the fist person to write a playbook for this kind of SaaS is going to get a lot of readers. From the outside, it seems like such a well-designed product strategy: all the benefits of enterprise, without the heavy integration burden.

Anyone have requests for an agent framework that adds some of the capabilities Yohei on Twitter has been tinkering with?

Longer term planning, memory, etc?

Steamship | Frontend Developer & AI App DevRel | Remote (N. America)

Steamship is building Vercel for AI -- a platform for hosting the next decade of AI Apps.

Whether you're building ChatGPT bots, LangChain question-answering APIs, profile picture generators, or storybook writers, Steamship can host and scale it for you.

We are a seed stage startup, with amazing backing, only 4 folks currently.

You can email me ted(at)steamship.com

Has a court decided that yet? That’s huge if true.

Someone once told me a curation was impossible to copyright for similar reasons.

I know we're all suffering exhaustion from all things GPT, all the time, but the new possibilities that keep popping up continue to amaze me.

I would far rather ask AWS: "Which ALB threw the error for the request to /foo/bar" than figure out how to put the dots together and discover myself.

I figure it’s a stand-in for embodiment.

Instant hosting for vector search LangChains alone has me checking my wishlist of weekend projects :)

Really appreciate you giving it a go in the face of those errors.

In retrospect, that initial `ship try` CLI & those initial packages had too many moving parts for a first launch.

Part of what we're trying to do with this Prompt API template is strip things down to bare bones so that we can nail the basic "clone -> modify -> deploy" flow before layering back on bells & whistles.

If you're ever up for another go, give us a holler and we'll jump in if you hit snags like that again!

[Steamshipper here] This particular Prompt API template is a simplified starting place to what we're working on. So you're right to ask that question -- it's not so different than a lambda in this form.

More broadly: there are a lot of moving parts to a production AI backend: vectors, LLM chaining, long-running async tasks, persistent state, data conversion, etc.

We're working toward a framework that lets you snap together those pieces at hackathon speed... and then translate it into a multi-tenant, auto-managed infra stack out of the box.

Thanks for the feedback!

Working on it!

Right now our examples are pretty simple. But you can chain prompts manually and/or pull in services like DALL-E or voice gen.

We're working on the rails to do that in a way that'll scale well cleanly across API users of your app.

That "Ask the Bible" service that went viral the other day on Twitter had me thinking the same thing.

Especially for the set of problems with "Step 1: Hosting embeddings for some large corpus", it feels like there's a really useful role for offering static query/AKNN search atop popular datasets.

Re: earning -- 100%. We'd love to talk to you if you've got thoughts / requests (crew@steamship.com).

Re: DSLs -- what's the famous saying? "All services that survive eventually grow large enough to re-implement email"? :) We hear you though; if there's a way to support a standard QL atop the domain we're serving, we'll give it a good look.

Steamshipper here --

Totally fair! In our case we're just trying hard to maximize product iteration cycles before we build out billing, etc.

If we just provided a way to pass through your own OpenAI key, does that work better for you?

Steamship is Heroku for LLM apps.

We're working on the mechanics of deploying and scaling all the amazing prompt-based backends everyone is building --- chaining, prompt-search-prompt, image & audio generation.

This link is a simplified hosting harness we built building & launching a Python + Prompt API endpoint.

Hey jcq3

Anything in particular thing stopping you from shipping it online? Or just that it's not necessary for your use?

(Context: I've been working on `Heroku for LLM apps` and trying to understand where the value/frictions are)

Hey folks!

I'm in the AI space like a lot of you. And while I've found the rate of change exciting, I've also found it disorienting.

So I created this Substack to help me process it.

Rather than try to tell you EVERYTHING that's going on, I pick a tight theme, find 5 cool projects related to it, and then unpack what's going on through that lens.

My goal is for it to be an editorial, of sorts, for fellow builders & investors. And a celebration of awesome projects no matter how large or small.

We (steamship.com) use Swift on the server to power our Heroku-style AI apps platform.

It’s been mostly great. Only downside is the server-side swift community is smaller than others, so you don’t get the deep software library selection something like Python would have.