It is fascinating and I love to see it. Ultimately though, why not build your own? I think that in part is what we're beginning to see-- highly customized and personalized software. I take most of these as inspiration these days and just build my own. Nothing you can't hammer out with a few good Claude sessions.
But why build your own from scratch when you can do it on top one extensible UI platform that already has all of your organizational context, not to mention coding agents already built-in that you can just tell it to build & deploy your personalized workflow softwaree from scratch? ;-)
Because it's fun! I don't "really" use LLM for anything serious yet, but after discovering playing with a custom harness I licked blood. Like going back to edit any message to start a new conversation branch off that, or being able to go back and remove things that turned out to to be dead ends from the history. Not really "useful" [0] for now, I always wondered about that and I was dumbfounded how trivial and tiny a first MVP was. Of course, that's also because it is trivial, it's just chat, it affords none of the usefulness you'd want for any "work", but it's still very interesting.
And for some simple recurring tasks or local housekeeping stuff, having something where I 100% know what files it reads or writes will surely be useful. E.g. if I wanted to watch out for certain topics on HN, I'd rather make something myself that grabs the feed and turns that into a list of titles and topic ID, which is probably 1% of what the front page HTML would be, and then only have the LLM process that output -- rather than telling an LLM to do all that every time. Even if tokens may not be that precious, and the difference in "cognitive performance" not worth speaking of, that would feel way neater to me.
That's an option right? When I say roll your own-- it could be extending another platform. There's definitely different levels. Customization is in reach regardless of where you want to start and stop your stack.
These new experiences are built atop layers and layers of bedrock libraries that we, generally, don't reinvent. Or we reinvent one or two.
What's notable to me is that ux doesn't so far generally have this behavior. That as per this post people just build a new app, a new experience.
I want to believe over time we'll have better composable & malleable ux experiences atop broader platforms for us. That over time the "go it alone" path has other worth ways to innovate that use a more substantial shared base. It's dangerous to go it alone, and doing so equipped with just our wooden sword and some courage and perhaps an LLM wisp is an amazing adventure, but I think the survival rate & impact would be much better if we had more general ux systems that supported better innovation atop them, and if less people did the pure "why not build your own" path.
I understand you take. And I definitely agree that we could use a lot better ux tools than we have right now. Maybe one of these prototypes will lead to that though. I'll look at it more as creative engineering. Not everything is meant to be distributed or consumed. But that doesn't mean that it's not worthwhile to try and prototype different flows and options. I have seen hundreds of harnesses, I've only used a couple-- but they've inspired me to build something that fits me better and works for me. If somebody comes out with something that helps unify the UX experience and provides better building blocks for me to work with, I'll definitely switch to that and rebuild my experience on top of that. But for now this is what we have. Admittedly this reflects my level of experience and my expertise too. I don't have a lot of experience working with building UX frameworks, so I adapt what I can.
People are using the UX systems though, just that it's the agent that is gluing them together now. Under the hood, it's still React plus component libraries like shadcn, lucide, tailwind, axios and so on.
The agents are so good at this, there's literally no point in not doing it that way.
These are all surface veneer but so so so much less than an app. The entire rest of the owl has to be redrawn, and all the guts inside the owl redesigned.
UX does have this, and it has for many years. You might remember Bootstrap way back in the day, and everything looking like bootstrap. These days it seems like everything is based on tailwind under the hood.
Right now shadcn/ui might be the leading candidate but there are plenty including old school bootstrap, material / MUI, a bunch of various flat-themed ui, etc. They often come in the form of a react component library or whatever but they're generally really solid. I also found one in svelte that I can't immediately pull up but it was nice looking too.
I virtualize mine with proxmox and pass the necessary NICs through to the VM. Snapshots make rollbacks a breeze. You can also run HA pairs if you want to do updates without downtime.
I don't work for a nation state, but I work for a large financial that requires adversarial testing and it's absolutely increased the ability to operate and attack these environments. The floor is definitely easier to reach at a minimum. Vulns and 0days are far easier to exploit, time to exploitation from the release of vulns is nearing zero, largely aided by LLM reverse engineering -- why wouldn't this apply to nation state level adversaries?
Doing anything special to handle the syncing of .claude? Kind of a pain point for me jumping between Windows/Linux and having to deal with path differences.
> Any concerns as far as privacy or data protection?
We moved to OpenCode Go ($10/mo), so we could switch between DeepSeek v4, GLM 5.1, and Qwen 3.7 models run by providers in EU, US, & Singapore that OpenCode FAQ claims don't use retained data for training.
What about data and privacy?
The [OpenCode Go] plan is designed primarily for international users, with models hosted in the US, EU, and Singapore for stable global access. Our providers follow a zero-retention policy and do not use your data for model training.
> Qwen 3.7 models run by providers in EU, US, & Singapore that OpenCode FAQ claims don't use retained data for training.
Note: Alibaba Cloud is the only company that currently offers Qwen 3.7 models (they haven't released any open weight versions yet), and according to OpenRouter, they retain prompts for an unknown period. So they might not explicitly use your data for training, but they do store it indefinitely on their servers and can potentially[0] use it for all sorts of other purposes.
[0]: Disclaimer: I haven't read their privacy policy. Just pointing out that it's not so simple.
Using OpenRouter, going to migrate to DeepSeek's official API soon. I'm not using it for anything commercial or for private data so I have no privacy qualms.
Makes sense. Privacy is my only real hang up with DeepSeek. Both of the big SOTA providers have become extremely filtered. Things that I could do one version ago are now getting refusals. Anthropic is almost unusable. ChatGPT is slightly better. Even with a "cyber exception" in place and a vetted account. They are going to force me to take my business elsewhere.
GitHub Copilot refuses to do any security testing or proof-of-concepts for exploits. While I understand why, we pay for Enterprise and I’m working on our proprietary code base. It’s incredibly annoying.
I’ve actually had luck taking the analysis from GHCP and pasting it into our M365 Copilot and getting a useful poc to stick into my bug reports.
You can always run deepseek yourself, v4-pro and flash are open weights. It's a little tricky to get the hang of self deploying open weight models but you do fully own your deployment substrate and privacy narrative at that point.
I get mine from bulksupplements. Gummies and many gym focused powders are extremely overpriced. Buying in bulk you should be able to get it for ~$.15 a day at 5g.
Long term studies haven't shown any harm. I've taken 30g a day for well over a year now for both the physical performance and the cognitive benefits. I had brain fog following a bad COVID infection which this has helped with. My family also has a history of Alzheimer's and dementia (both sides), I figure it can't hurt.
The only side effects I've encountered have been mild GI discomfort, and that only rarely (mostly when restarting after a vacation etc-- I drop the routine when I travel). Roughly similar to having a morning coffee at its worst.
I've found I get much thirstier when on this high of a dose. If you're not already a big water drinker I'd definitely invest in a nice insulated bottle to drag around.
The standard studied dose _is_ 5g, but the common advice you hear in the gym community is that you will see cognitive benefits from 10-20g or more, depending on your body mass. 30g isn’t uncommon. And, many believe that 5g may be too little for optimal muscle gain, again depending on your body mass.
Adverse effects, as mentioned, are mostly GI-related, some people experience bloating especially with monohydrate. Another effect is an increase in water retention, which, in turn, means you need to drink more water. It’s generally regarded as safe for the kidneys unless you already have reduced kidney function.
So, overall, not really a big deal to take well over the studied 5g dose. Not to mention the recommended loading dose is 5g 4x daily.
It is definitely more than the standard dose. I was already taking 5g for weightlifting purposes. After some of the studies suggesting mental benefits I titrated my dose up to 30g (the high end of the studies). Just added an additional 5g per week until I hit 30. I add it in powder form to my daily protein shake-- usually in the morning. While it could be a placebo, I personally feel there's a noticeable difference. I buy in bulk so it ends up being less than $.90 a day to dose at this level. It's an extremely cheap supplement.
I'd say that this is a bit relevant to the entire field of cyber security and a good chunk of development roles. If you're not concerned about how password hashing (which is a key component of understanding cracking) works as developer-- I'm not sure what to say. While not all of the in-depth research is probably needed. It's definitely relevant to many technical fields. I work in offensive security and we use tools like this daily in our industry. And no we are not cyber criminals.
