elliotbnvl

Definitely an invasion of privacy. I can’t visit this website in good faith. It should be taken down.

The point is valuable, and the mission is important, but the ends do not justify the means. If this must be shared, at least use static pictures and don’t stream the content for viewers.

$500k is a vast overestimation. For massive concurrency at FP8 or even BF16 maybe.

NVFP4 at reasonable speeds (~120 tok/s) and concurrency is possible at a $80/90k figure with today's prices, maybe even less. That buys you 6 RTX 6000 PRO Blackwells, a decent CPU and motherboard, power supply. 576gb of VRAM.

You could do it for under $50k if you're OK with 40 tok/s decode, ~1200 tok/s prefill.

It's more about the trajectory.

It is insane that we are comparing locally-hostable models to leading cloud providers, it is wild to me that this article even exists.

We have come a long way, and very clearly have a long way yet to go.

I don’t foresee development timelines speeding up but I am seeing entire applications being built by one developer instead of four over the same timeframe. If anything communication overhead is reduced and velocity and quality are up, too.

I mean it is the butcher who knows how the sausage is made.

It's a home desktop form-factor with 8x48gb DDR5 6800, 9985WX, a whole lotta fans and a 1600W PSU. Max-Qs are the only card you can fit four of into this rig without PCIe extenders or cooking themselves, unless you go water cooled (which I didn't).

I have been thinking about this a lot. I just bought a rather expensive rig for local inference for a home agent (powered by four RTX PRO 6000 Blackwell Max-Qs).

As I contemplate handing it more and more of the keys to my life, I grow increasingly concerned about what is, to me, the primary risk of this. Not data destruction (automated backups are trivial), but data exfiltration. Specifically, via prompt injection.

My solution to the problem, which I am implementing as a Hermes plugin + custom iOS / macOS app, is simple: an airlock architecture. One Hermes profile runs with local FS access and no internet access, inside an Apple container, and one Hermes profile runs with internet access and no FS access, inside an Apple container. They never share data directly or in any automated fashion.

If the user (i.e., my wife) wants to do some internet research, she can start a conversation with the remote-access profile. This is analogous to Claude and ChatGPT apps in their current state. However, at any point, she can flip the conversation over to local mode, which copies and pastes the conversation's transcript into the local-only profile (which has zero egress, enforced at the VM level) and seamlessly switches over to a new conversation in that profile.

After that, there's no way to re-enable internet attachment. Should she want to spawn a new conversation with information derived from the local file system, she starts a new conversation with a local agent, asks it to write up a research plan, and then – this is the airlock – manually begins a new conversation with only this plan in context.

The advantage this grants is that it's no longer necessary to worry about poisonous inputs flowing in – she only needs to worry about making sure any generated plan, the only artifact which could conceivably enter into the egress-enabled agent, does not contain information we'd rather not share with the internet at large.

I think this is bulletproof, but very much welcome input. Is it possible I am overengineering this out of paranoia? Yes. Will I share a lot more of my personal data with the agent as a result of its perceived security? Also yes. Is that dumb? Maybe.

Thank you. This is good info.

I just learned I have this as well (not as severe). Quercitin helped me. I am taking I think 1600mg (with bromelain) per day. How much was she taking, and what is the mast cell stabilizer that helped, if you don't mind my asking?

I've been dealing with my symptoms for 17 years this year and Quercitin + Zyrtec + Pepcid is the first thing that's made a dent in it. I started a few weeks ago and it's been amazing but I'm not experiencing full relief yet.

This was the combo Claude recommended I start with for a trial, one message after I told it my symptoms. No doctor has ever been able to help.

The lethal trifecta strikes again.

You may as well ask to run a comparison between gnu libc 2.42 and musl 1.2.5.

Telling me you wouldn't learn anything from this?

What are those tasks? This and the paragraph just after seems very much like a XY problem where all the energy is focusing on resolving the Y, not the X. It's like discussing how we can reach the moon using cannons.

Or like how we can get from A to B without horses.

It's a different world, one worth learning about. If these tasks don't at least arouse your interest, nothing I can say will help you.

Go off and run a comparison of Qwen 3.6 27B and GLM 5.1 GGUF (https://huggingface.co/ubergarm/GLM-5.1-GGUF) at IQ2_KL 261.988 GiB (2.985 BPW) and let me know if you learn anything.

Or maybe just compare Hermes vs OpenClaw for long-horizon personal agentic tasks. Which one performs better in offline inference personal finance analysis tasks?

Or read up on how the `/code-review` workflow works in Opus 4.8 and give me a guess as to how long it'll take Codex to implement it and which tool would be more appropriate for your engineering team (don't forget to include enterprise API token costs in workflows – it can spin up 100 agents in thirty seconds).

If you can figure out how to secure agents with simultaneous access to personal data and the internet to run unsupervised while avoiding the lethal trifecta (Willison, 2025) let me know.

This is patently false. I work with and on AI every day at multiple levels of the stack, and every day I'm learning massive new swathes of information. I'm honestly shocked how deep the field goes and how much more effective you can be with time. The floor is falling and the ceiling is rising and the gap between them is widening every day.

And I don't read anything indicated they had fun.

Maybe I'm just projecting. I enjoy making things. Maybe they do, maybe they don't. Sounds like you don't.

There is pleasure in making something yourself. There is learning. There is pride.

You're speaking second person, when you should really be speaking first person. You enjoy making everything yourself, by hand. That is fine. It's also your personal perspective.

You are learning nothing.

If you really aren't learning anything, you're doing AI wrong.

Everyone else is vibe coding but they got the special sauce, the perfect prompts. They are delusional.

The delusion here is constructing a strawman out of the worst qualities you can imagine and berating that instead of actually looking at what other people are doing and trying to work out what they're thinking / how they feel. I can guarantee you that virtually nobody thinks they are the only person that can prompt a particular piece of software into existence.

I know this post probably won't land with you, because I'm a little annoyed while I write it (if only because your post comes off emotional and annoyed as well) (and, sorry in advance), but I do encourage you to consider that perhaps there are other worldviews than the clearly embittered and deeply entrenched one you've espoused. And perhaps those other worldviews are more suited to surviving the oncoming storm.

If it wasn’t a learning opportunity to build those things, that was the waste. You can learn from an AI far more easily than from a book — only now it’s far more easy not to and many people unconsciously choose that route.

It seems like the author is overindexing on useful and underindexing on wonderful. He clearly had fun building these products — and in hindsight is disavowing them because they didn’t generate income? An oddly capitalist view of play.

Some really good points on how these bots are incentivized to reward mindless engagement though and the bit about voice transcription not producing useful writing landed. When the barrier to release drops the quality naturally does too.

I think the next stage of us learning to harness these tools is us building the ability to reach for excellence even when we are not required to. To accustom ourselves to going beyond minimum viable bar for functionality and to reach for qualities or standards beyond that which the AI brings to the table unaided. A new kind of engineering rigor.

I move that this was always true and is now only far more so.

That's an optimistic take I haven't heard before, love the idea a lot. Wake me up when we get there though...

Because we are talking to the AIs instead of talking to them.

Speak for yourself. This is a sweeping generalization.

There was plenty written about how Google was making us dumber because we didn't need to remember anything any more.

Oh yeah, and that other new-fangled technology the Greeks were complaining about – books.

Why can’t they continue to do so?

If anything an underlying truth about humanity is being exposed: we take the easy way out far more often than we’d like to admit.

Perhaps, this truth being made explicit is a wakeup call that will teach us the value of that hard work anew.

After all, nothing the author’s written isn’t also true about Google, but nobody realized how bad of a mistake that was.

Though I deeply agree with this sentiment, the author fails to address that there can be multiple goals to an action. And writing can be for art or communication or both but not always both, and removing the art from communication doesn’t destroy its ability to communicate.

This resonates with me as well. For more reasons than one: with the rise of AI (Mythos is but a pale forerunner) digital security — and by extension, digital privacy — has ceased to exist. The bomber will always win. The only way to win is not to play.

This is a comically self-absorbed perspective.

Why are people making things with Claude Code if not because they’re motivated?

While you are not factually incorrect, my expectations were subverted in the same way that OOP's were.

Strong agree. Needs another pass or two at editing though, some painful LLM-os sticking out there :'(

And the people that still get excited about life!

It's a usability / quality of life feature to me. Nothing to do with increasing perceived confidence. I guess it depends on how much you already (dis)trust LLMs.

I'm finding more and more often the limiting factor isn't the LLM, it's my intuition. This goes a way towards helping with that.

Looks like it's an order of magnitude off. Missprint?

This is something Opus 4.6 likes to generate a LOT for some reason.

Hi all, author here. Please don't go to this website. It's expensive for me. It's a realtime multiplayer text-to-website chat, powered by Sonnet 4.6, that I built to showcase the incredible agency and speed frontier models have attained. I am paying API costs out-of pocket. Call it a business expense?

It's inspired by Pieter Levels' tweet about chatting with his websites. [1]

It uses the same tool under the hood as the core web offering, bareclaw [2], a thin Claude Code CLI wrapper I created to enable custom CC client creation while leveraging your Claude Max subscription.

[1] https://x.com/levelsio/status/2023960543959101938 [2] https://github.com/elliotbonneville/bareclaw