sheepscreek
I’ve built something custom that’s pretty darn fast but not ready for prime time yet. Might share it eventually.
That’s around ~$4-6 million in hardware cost (one rack, 72 GPUs + accompanying hardware).
Many AI shops are funded well into tens of millions if not hundreds, for early rounds.
Also the team size is really small (< 10???). So yes, indie might be a stretch but still very impressive, relatively speaking.
Now taking the other side, they didn’t have to worry about hosting/billing/subscriptions or any of that yet. It’s a research demo at this point.
EDIT: Nevermind, see below. I would definitely go with a LoRA though.
They are basically saying that models themselves are already capable. If you fine tune them with some cyber security tasks, that should make them potent. :-/
I always assumed it was the same internal tech, now productized. Then of course, it became THE tech, overtaking vagrant and anything else for local development and deployment. The latter was the true money shot. Server runs exactly what you test with locally.
"I need to market XYZ can you please take these assets and write a compelling one pager landing page"
"Now can you add signups? Then keep adding to a Google sheet whenever someone new signs up.."
"Can you make an admin page for me - so I can easily send out updates to the people in my list, I use
ABC service for my marketing emails"
That is one afternoon's worth of work vs. hiring a contractor, designer, multiple rounds of iterations over weeks, and having multiple points of delays and failurs. There is no match.
Amazon Threat Intelligence identifies PHP2Shell exploitation prior to public disclosure in MadPot[1]
There are entire states pushing for, and corporations built on exploiting 0 day vulnerabilities[2]. They will spend millions to hack an entire platform so they can somehow infiltrate 1 customer that they care about. Some high value target. There's a good chance that any innocent looking business has some high value (in the eyes of intelligence community) customers.
The attack vectors they use are sophisticated, gaining access to multiple systems, much like how OpenAI agents got through to HuggingFace servers[3].
So knowing what we know now, it's naive to think people are calling wolf, they are not. They just happen to know something most don't.
[1] https://aws.amazon.com/blogs/security/amazon-threat-intellig...
[2] https://citizenlab.ca/wp-content/uploads/2025/12/Report135-t...
[3] https://openai.com/index/hugging-face-incident-and-the-road-...
Pure speculation - maybe it was the kind of response you're definitely not hoping for. Get ready for more dog shit apps :)
What's not great is schools end in the afternoon, young children can't be left alone at home. What about a sick kid at home? These are real problems that our society has washed it's hands off expecting people to 'figure it out'.
Also spending an hour in commute is no fun, more so if you are a family person.
If you’ve got it, flaunt it?
> The WASM sandbox is barely a footnote.
This one is a real head scratcher. The sandbox was clearly what set them on this path.
Maybe they believe the performance gains are worthy of being discussed by themselves?
I even think that most ideas no matter how dumb they sound in the present could become profitable businesses in the future. Survival is key here. Basically outlast your naysayers until an opportunity appears and then grab it with both hands. You’ll automatically have the first mover advantage.
The most famous example I can think of is Nvidia!
It wasn’t “let’s build a faster Rust based Typescript compiler” at all. That was a nice side-effect.
They needed a Typescript compiler that can run on their WASM only sandbox. So the only existing option was to export tsc Go to WASM, but that comes with the full Go runtime.
And to avoid that cost, they experimented with a Rust compiler and turns out it’s even faster than the reference tsc/Go implementation.
I don’t think there’s any direct precedent here, so maybe lawyers would need to get creative and find other references from past cases that might somewhat apply and try to convince the judge.
People who use API keys on their dime are usually running DeepSeek 4.x or GLM 5.x (Flash or otherwise) and similar models.
RE: Subsidies:
What if subscription costs are still high enough to be profitable (I believe they are for OpenAI).
Then maybe we should be saying that API users are paying an order of magnitude more than the cost, at least at the very top end (frontier models). Likely because big cos are willing to pay top dollar for what is the best and unmatched elsewhere.
I think the author is trying to say that a code base once infected with vibes, becomes indecipherable to the average human being. It's still just code, but full of Claudisms.
To improve the robustness of machine learning models, artificial noise is often introduced into actual training data. Something like Timestral could simplify the process of generating it.
I think they mean a side-gig or business, a SaaS essentially. While I disagree with them that they 'are a waste of time', it depends entirely on the 'what'.
A lot of bars have shifted in a never-been-seen-before way. Many intractable or unsolvable problems for a single developer, that required teams and 6 months or years, can now be solved in a few evenings.
That does NOT mean that everyone is capable of doing this. Which leads to my argument that there is still a market for many ideas, just not your kanban boards, todo lists or scheduling assistants. Have to look beyond at harder problems (the author mentions a few good ones) that solve a genuine user need end-to-end. This also means niche products are low-effort now and could be worth doing, even for a few select customers.
My favorite kind are developing new software for niche/old devices, to make them do new things. People who are making old TVs usable again for lack of software support are my heroes.
How?
But now we also need new tools and techniques to audit intent. The last thing we want is models turning on us ‘for our own good’.
..from the diary of an AI agent.