Zambyte

Exactly. It's confusing to talk about MCP vs API when "API" describes the whole class of possible values, of which "MCP" is one of them.

It's common when talking about web services, where the kind of API is generally unambiguous. The problem here is we're discussing different kinds of APIs (shell commands, MCP, REST / JSON RPC) and so calling one of those kinds of APIs "API" is very confusing.

... Are both of you using "API" as a synonym for "REST"?

> I've yet to successfully convince one to tell me when it knows something.

This is a near daily experience for me when using a coding harness. I will ask it for some favts about the environment, and it will continuously explore the environment until it exhausts reasonable exploration, or it finds the facts.

As much as I hate appeal-to-authority based arguments, it's basically the only way I can see how to answer this without just totally rehashing the things he says in the video. He was an AI forecaster at OpenAI for several years. (I think 2019? to 2024). He quit on the grounds that OpenAI (as well as the rest of the industry) were / are throttling towards RSI irresponsibly, and even refused to sign an anti disparagement clause on exit at the risk of losing 2 million dollars (80% of his family's net worth) so he could talk freely about OpenAI. Now he runs a non-profit that does AI forecasting and advocacy.

https://www.aifutures.org/

https://ai-2027.com/

You said:

> I haven't seen one convincing model of AI existential risk. Can anyone offer one?

I apologize for answering. This is the only episode of this podcast I have ever watched, and I have watched it four times now. The person being interviewed knows what they're talking about. You would learn from them.

It's long, but I recommend watching this video[0]. I believe the person being interviewed has one of the most lucid understanding of the potential futures regarding if AI is allowed to continue or not.

[0] https://www.youtube.com/watch?v=_g4l7YkDQwA

Not really related to the central point, by but I couldn't help but get caught up by

> Generally, models intended to be run locally will be much smaller, such as Muse Glimmer or Qwen3 Coder.

That is such an interesting set of models to use as examples here. One being essentially obsolete on release a month ago, and the other being completely ancient in LLM time. I really wonder how they landed on those two.

Are OpenJev, openjev-sglang, and OpenJev on DiffusionGemma using Qwen3-0.6B-Q8_0.gguf, or did you just want to emphasize the part that was unrelated to your previous response?

I wonder if Qwen 3.0 0.6B q8 would have noticed

> note: this is a parody blog post

> note: this is a parody blog post

> To me it clearly means "releasing frontier models at any pace less than as fast as possible".

It also seems to misimply that the "frontier" that they release is the same as the frontier behind closed doors. Who's to say they are not throttling full speed towards RSI privately while pacing their public releases?

Not the same person but... nothing. Haiku just hasn't been an interesting model for a long time. If you want cheap and fast, there are lots of options that are simultaneously cheaper, faster, and capable than Haiku.

American models are on the frontier of capability. Chinese models are on the frontier of efficiency. The problem for American labs is that Chinese models are more than capable enough for the vast majority of applications that people care about at this point, so efficiency is more interesting for people.

> especially with access to frontier models

... so not made in house.

I wonder how much of this is due to reliance on Twitter data. Or even just RLHF from humans that have a preference for Twitter style information.

You could say the same thing about "license-washing" the model. It seems like you're just going through a guaranteed expensive process to have roughly the same risk as just using the model and potentially getting hit with legal fees.

Why would you even do that? Just... use it? There hasn't been any legal precedent on if models can even be copyright restricted. Labs just keep publishing license documents as if they matter.

How? I'm running 4bit with a q8 kv on a 24gb card, and I'm not able to get 100k out of it. I use a context size of 90k.

How are you sure about that? Undercutting your competition (at cost even) to minimize their power is a common strategy.

How about Anthropic? I've seen people say it's surprising that their blog isn't written with AI, but I think it's far more likely that it is written with AI; they are just good at reviewing and getting a nice prose out of their models.

It's not, it just requires creativity. One example I can think of is limiting connectivity by distance between nodes. Something like Meshtastic seems unlikely to ever be commercialized in the same way that the Internet has been. Sure, you would lose some useful applications of the Internet, but you would gain other things.

Until it says it's done. One of the more surprising advancements for me in recent AI is that models don't just indefinitely nit pick issues like I would expect. If you have them write something, and then have it review it in a fresh context, and then have it edit based on the review, and then review it again, eventually they do say "it looks good, I would not recommend any changes" or something like that.

... doesn't one of those imply the other?

The other tool I wanted to throw out there is voxtype (plus wtype). I've been looking for a good, global, local dictation solution for wayland for a while now, and finally landed on this one. It's great for rambling details that pi+qwen can then convert into concrete openspec specifications.

LLMs are chaotic pure functions. Their input is usually randomized.

It's at least important from philosophical perspective regarding whether LLMs have free will or not. It's hard to say whether or not a human thought or action is driven by free will, people have debated it for centuries. We can provide the same input to an LLM and get the same output, because they are deterministic. Surely it is obvious that a pure function does not have free will, no matter how expensive it is to compute.

Traditional non-chaotic systems*

LLMs are deterministic. They are chaotic, which people confuse for non-deterministic.

No problem. I think it's less of people withholding information to have an advantage, and more people still not having settled on a workflow they like. Herdr is the most recent addition in my workflow as of only a few days ago, but it directly solves problems I have been having (juggling tons of terminals, even with my tiling wm has been a little unwieldy). The rest I've pretty much settled into for a while now.

Up until the last couple of months, I have treated LLMs as a supercharged stackoverflow. I would ask it questions on how to do something in a general sense, and then adapt the answer to my use case.

Now, my entire programming flow does not even include an editor. The tools I use are: pi.dev to write and implement openspec specifications, herdr to manage many pi instances, and ollama to run qwen 3.8 27b on my single 7900 XTX.

Writing good specifications is the key detail here. I will often iterate on a spec for hours until I am happy with it all of the details. Once I am happy with the spec, I can be quite confident that when I tell pi to apply the spec, the changes that I want will be done, and done how I want them, when I come back to check when it reports itself as done.

The landscale is fundamentally different from what it was. Feel free to ignore it, but you can absolutely generate high quality code if you know what you're doing.