macleginn

I am an emigrant now. When I was still living in my East European country of origin, my wife and I would regularly hold parties. We liked going to other peoples’ places more than hosting but most of our friends would not throw parties themselves, so we had to fill the gap.

"Moral panic" were the exact thoughts that came to my mind when I read this yesterday: https://www.theguardian.com/technology/2026/sep/11/mps-urge-...

The text is mostly about rustc_codegen_utc; maybe this should be reflected in the title?

Also the use of italics there is rather jarring.

This is how deniable threats work. The promise of "not being nice" in conjunction with "ruining the career" is as clear a threat as there can be in writing.

I think this should be in the title of the post. 'OpenAI allegedly threatening to ruin a prominent researcher's career', or smth like that.

Recurrent depth and chain-of-thought are two completely different concepts. In the former approach, the output of the layer gets is rerouted as input to the same layer, potentially several times. This output/input is a fixed width times sequence length real-valued representation; it is not comparable with output tokens.

Generally, it is hard to imagine how neuralese should work given that models are pre-trained on naturalistic documents: CoT is a comparatively simple extension of that, while neuralese demands a completely novel training paradigm.

This is why almost 100% of hard and social scientific discourse reads almost exactly like this these days.

‘But the idea that you’d need explicit self-referentiality before you could get convincing and world-changing conversational intelligence?’ — has anyone ever formulated this actual idea or anything logically equivalent? This looks like a mighty straw man.

I am duly impressed by the powerl of the nameless internal AI, but not a single human contributor's name listed anywhere? Did someone at least make this model a coffee?

What about writing code by hand?

Array programming where we compare first and look for the first failure later will not help much here if runs are short because by itself it doesn't give you early termination and you may spend a lot of time on wasted comparisons.

An unintended consequence may be the separation between master's and PhD programmes, like what is normally done in Europe.

I find it fascinating that two of the most (if not the two most) original, inventive, and outright weird British writers of the 20th c., Ballard and Mervyn Peake, spent their early years in China.

"It will help us solve some of the biggest problems society faces from accelerating drug discovery to developing new clean energy sources to creating novel advanced materials" — but these are not the real problems plaguing modern developed societies, are they? What developed societies really need to figure out right now is how to distribute the already available resources without making people miserable, and so far AI hasn't been helpful.

For context, I have access to MS Copilot through my workplace. To see what it looks like, I have tried to login through https://copilot.microsoft.com/ , where I was informed that my account, although recognised, is not yet supported. However, I can get more or less the same chat window, with access to all the data, through https://m365.cloud.microsoft/ A redirect could have been useful.

The overall resemblance to Ticket to Ride's Indian map is rather striking.

Given the sheer volume of the output, it is conceivably difficult to maintain the quality of the plots/characters, even if the animation is stable or even improving. As a byproduct of that, more and more ambitious/batshit source texts are being animated, which sometimes works well (Dandadan) and sometimes doesn't (Uzumaki).

They are poor at generalising from a small number of examples; this is why the real generalisation power is achieved in pre-training.

Unremarkable base model will remain an unremarkable fine-tuned model that memorised a couple thousand of input-output pairings.

They normally don't do self-set-up real-time monitoring though.

Ok, cool, depends on application, probably.

Yes, they do have money to burn, and this will bring some improvements for sure, but active learning has never really worked out, has it? And even 10% of the educated population doing this for, like, 50 years is not that much data, while normally each accuracy percentage is more and more data-expensive.

Not the worst way to make money, but if internet-scale data were not enough to reduce errors to a somewhat tolerable margin, how much data do they hope to collect in this manner?

> Noone is really caring about hallucinations on point facts these days though, it is much more about complex reasoning tasks.

The boundary is pretty thin there though. E.g., Gemini recently told me that a certain papers claims that two frameworks are mathematically equivalent, while the paper shows the opposite, and yesterday Google's AI overview told me that no World Cup matches were scheduled for that day despite their being several of them. The model probably used complex reasoning to arrive at both (incorrect) answers, but superficially they look like basic errors of fact.

The main problem here is that hallucination suppression doesn’t generalise. We can penalise models for incorrect answers on a wide range of questions, but this doesn’t lead to the emergence of a coherent worldview, which, coupled with logical abilities, is the only true remedy against hallucinations. With current architectures, hallucinations will likely persist on open-domain tasks forever.

It's been 63 euro for some months now.