fizx

I thought dsql was distributed transactions. :thinking:

I miss SF in the early 2010s.

- gorgeous 3-seat wine bar inside my laundromat

- hidden sushi restaurant in friends garage

- hole-in-the wall coffee shop with only cushion seats on the floor and $1.50 breakfast sandos, frequented almost exclusively by writing clubs

- corner store with half-stocked shelves, still using a cash register that printed receipts with a mechanical typewriter

Cursor + Gemini is rather good.

My hypothesis is that the greenfield projects make it easier to learn AI. I find it pretty easy to get value out of Cursor on 500k LOC legacy code bases, but I've also spent a few hundred hours on green field projects.

> Meta CEO Mark Zuckerberg is personally creating a new "superintelligence" team dedicated to building the world's most advanced AI platform, and splashing out nine-figure packages to hire top talent, the New York Times and Bloomberg reported Tuesday.

At a tech company of >1k engineers, this gets audited regularly by a dedicated data retention team.

Cursor+Gemini MAX is pretty good these days. It seems like Claude Code and C+GM leapfrog each other every month or two.

Cursor has a lot of potential leverage owning the developer and the training data streams and commoditizing the underlying model.

Writing has made people worse at memorization. This argument has been around since Plato.

What is this 1T index technique they seem so hyped about?

Once that auto-regressive model goes deep enough (or uses "reasoning"), it actually has modeled what possibilities exist by the time it's said "There are 2 possibilities.."

We're long past that point of model complexity.

If Trump believes that, it would reflect a complete lack of self-confidence in his negotiating skills.

Maybe you've discovered that learning pays compound interest.

AIs today basically fail because they've been trained to be aggressive editors of code. This makes the first steps feel amazing, gives you the most out of your first tokens, and helps win the evals focused on coding simple-to-moderate tasks.

Once they hit some threshold of project size, they overcommit, bite off too much or don't recognize that they're missing some context. Agents help this by allowing them to see their mistakes and try again, but eventually you hit some death loop.

I think someone around now-ish will realize that there should be two separate RLHF tunes--one for the initial prototype, and another for the hard engineering that follows. I doubt its that hard to make an methodical, engineering-minded tune, but the emphasis has been on the flashy demos and the quick wins. Cursor and folks should be collecting this data as we speak, and I expect curmudgeony agents to start appearing within a year.

Combine this with better feedback loops (e.g. mcp-accessible debuggers), the agent doing its own stackoverflow/github searches, and continued efficiency work driving token costs down by an order-of-magnitude every year or so, and agents will get very very good very fast.

In this atmosphere, humans will shortly exist to get context for the agent that it can't fetch itself, either for security reasons, or because no one's built the integration yet. And that will be short-lived, because integrations will always be built.

So I guess there's a window for the "copilot" reality, but it feels very very brief. I don't think agents will need humans for very long.

Hypothetically, yes. Practically, after having built ~10 small apps in a unified framework with Cursor and Claude Code, it doesn't seem true with today's AI.

AI is killing e.g. React Server Components and Svelte, but for different reasons.

Vibe coding doesn't care about the implementation details, so Svelte is dead.

A unified codebase might be better for humans. But a FE/BE database is better for AI, because you have a clear security boundary, separation of concerns, and well-known patterns for both individually.

The community seems rather divided as to whether these are intrinsic, or we solve these with today's tech, and more training, heuristics and workarounds.

I have it at pro.

Are you able to add 3-5 customers per month at the 25/mo level? If so there's a chance. If that many customers are signing up with just you and minimal marketing, you could probably 10x that with channel partners, then start working towards an enterprise version or additional marketing.

Another signal would be that you are able to successfully add a co-founder.

If it takes more than 6mo of post-launch to those milestones, I'd give up on the idea. Market's probably saturated or you're not solving a pain point.

Things only get more stressful. I'd work on backup plans or some job searching during those six months. Its nice to do the job search when the side project gets too much and vice versa. Also, it can be a nice side project if you re-enter the workforce.

Source: bootstrapped a SaaS company to several million in revenue.

I'd put a $1k long bet that a 3B param model, cleverly orchestrated, will achieve AGI* in the next ten years. These are the sorts of ideas that would help get us there.

Any takers?

*AGI defined as smarter than a FAANG staff engineer on similar tasks.

Nuclear materials have the advantages of being rare, dangerous to handle, and hard to copy over the internet.

What's the difference between a city park and e.g. the hundreds of miles of trails and national forest behind my house? Somehow that manages to be crime-free despite less usage. Would love a theory of how density works in your model.

That's absolutely fantastic, because if you have 1 good idea that's additive to the SOTA, you can test it for a dollar, not millions

same same

A background thread long polls the CDN, updating a local hashmap on change.

The scale is easy in practice, cause you outsource to a CDN. But everything takes time and has opportunity cost.

If I was a bootstrapped startup, I'd do a json file and then when I've outgrown, I'd hand write something that long-polls a CDN for updates, with a tiny rails or react app behind the CDN.

But these approaches are insane for companies above a certain size, where individuals are being hired and fired regularly, security matters, and feature flags are in the critical path of revenue.

Last time I looked at LaunchDarkly Enterprise licensing, it started at $50k/year, and included SAML.

Now that sounds like a lot, but if you're well past the startup stage, you need a tiny team to manage your homegrown platform. Maybe you have other things for them to do as well, but you probably need 3 people devoting at least 25% of their time to this, in order to maintain. So that's at least $175k/year in the USA, and if your company is growing, then probably the opportunity cost is higher.

You're not getting new high-quality textual data for pre-training from your chat service. But you are potentially getting a lot of RL feedback on ambiguous problems.

The problem isn't that criminals can use your service, its that the service provider really doesn't want to be liable for that happening, which generally only happens when you host illegal content.

Burning Man's been dying slowly since about 2018, so pre-Covid. When I say dying, I don't mean "getting mainstream & EDMified" like all the old Burners have complained about since forever. I mean, the mainstream has stopped taking interest, and increasing portions of the tickets are getting taken by internationals because its passé in the USA, but still on some people's travel bucket lists. This year was the first year in a while with extra tickets available, but the writing has been on the wall for a while.

I'm sure raving will come back in 20 years like most fashion.

ADHD?