Philpax

Appreciate the read, even if the workflow's not for me personally. I've found that I'm just not that interested in writing the code by hand now that the option to not do so is available - but I totally understand people who still want to!

> It's not going to happen accidentally.

https://transformer-circuits.pub/2026/emotions/index.html

Whether these are like "our" emotions is hard to say. What we _can_ say is that they are emotion-shaped, we didn't design them, and they happened accidentally.

Modern AI is grown, not meticulously designed, and we cannot say with any certainty what the resulting mechanistic properties are.

They didn't file the report. The model drafted a report that was not sent.

No, but I am suggesting that OpenAI and Anthropic are further down the RSI path than any of the other companies, as we can see from the model that solved Navier-Stokes being less than two weeks old at the time: https://openai.com/index/navier-stokes-solution/

Are these other labs in the room with us now?

No, seriously, I'm all for a multipolar world here, but he's right that the frontier is literally just those two companies at present.

Google is behind. MSL is doing better, but not by much. xAI is a dysfunctional joke. Thinking Machines aren't on the frontier. SSI's primary output is their announcement post. Poolside was bought by NVIDIA. Arcee aren't vying for frontier. Magic have been largely AWOL, aside from their recent blog post. Reflection have shipped nothing.

but like, they did

The HF incident had them pwn their own cluster: https://en.wikipedia.org/wiki/2026_OpenAI_agent_cyberattacks...

I wish you luck on your quest to avoid all of the software on that list.

"Hackers" because the term in the context of "Hacker News" is broad and hard to define. Intelligence is, too, but I'm not in the business of pretending that the AI systems aren't some form of intelligent.

While I'm sympathetic to the sentiment, the ongoing automation of intelligence is, for better or worse, one of the most consequential things that can/will happen to "hackers" (as well as white-collar workers in general), and it is very difficult for us to look away from that asteroid in the sky.

It's been around for six years; at this point, I imagine any damage it could have done has been done already.

Your phone does not offer more screen area on demand. That is genuinely useful for many people. Apparently, not you, but that's OK.

The entire point is that if he couldn't naively figure it out, no normal user would. I'm sure that he could have thought like a technical user and gotten there, but he shouldn't have needed to.

No, I mean we just don't know what's going on in the circuits of the model at any substantial level. We set their architecture (hyperparameters), we pump them full of data (pretraining), and we shape how they behave through examples (SFT) and reward (RL), but we can't say with any certainty what the resulting model does internally.

You can scroll through https://transformer-circuits.pub/ to see the ~extent of our current understanding.

We don't know what they do. We shape them, but our understanding of how they get to their result is comparatively minimal.

It was put up and then taken down. Strange things afoot.

Er, o1 was also RL.

Actually, I would say the exact opposite. This post is full of strange and grammatically incorrect phrases, weird paragraph pacing, and unintuitive clauses: that is to say, this reads as very strongly human-written to me, and it is refreshing.

o1 was first, and Anthropic were doing a bit of it; DeepSeek brought it to the masses, but did not invent it.

> Ah yes, Nexus Mods, where they'll ban you for making mods that change "Body Type A" and "Body Type B" back to "Male" and "Female", or remove mandatory pronoun selection.

Oh my god please get a life

Use of Claude Code for an application used by artists. Just a fundamental misalignment of expectations.

Interpretation is in the eye of the beholder :-)

Maybe you don't! It is very possible that your problems don't actually need frontier-level artificial intelligence.

Ah, failed to snipe it. Here's the actual thread: https://news.ycombinator.com/item?id=49525496

Honestly, can't say I miss my Cursor subscription at all, and I'm surprised they're still a going concern. Why would I want to use a proprietary VSCode fork with an identity crisis when I can use literally any editor with any agent and have about the same experience?

My measurement was with MoE offloading, but there's only so much you can keep on-GPU with a 200GB quant and 48GB of VRAM. It's hard to overcome the CPU/RAM bottleneck.

For what it's worth, all of my hardware was used; I think, all-in, I'm probably at around 3k-4k USD? Not cheap, but also not the worst for something relatively versatile.

There are risks associated with releasing historical proprietary models that were not designed for open release:

- It is trivial to extract samples of the training data that was used, which can bolster existing lawsuits/foster new ones.

- Older models are not as safety-hardened, so it is easier to coax unsafe behaviour out of them, which is a PR risk.

- It may be possible to divulge proprietary secrets from the model (e.g. architectural details that may still be relevant).

For these reasons, and more, it's unlikely that GPT-3/similar models will be released until these concerns are no longer relevant (e.g. when they become a purely historic concern, similar to the open-sourcing of other proprietary software from decades ago).

No? They were the frontier, or near it, at the time of release: https://artificialanalysis.ai/models/releases/gpt-oss-120b

The fastest I was able to get my Threadripper 3960X + 2x 3090s + 256GB DDR4-3200 to run a 2-bit quant of GLM-5.2 was 8 TPS. I would expect to be in seconds-per-token territory for a pure-CPU 4-bit quant.

leaving the README like this is a good bit, though

Performance is not the same as original creation, so yes, I think that's very possible.