lucisferre

I wonder if it is more of a fixation problem. In that by asking it not to do something you increase the likelihood of it doing that thing because the words for the thing are there.

Sort of like how sometimes drivers fixate on a thing they are trying not to hit and as a result run right into it.

Evidence or not, this is probably the most satisfying explanation I've heard for this behaviour.

My own suspicion is that LLMs are not becoming more general-purpose over time, as these companies had hoped, and ongoing development in that direction is stalling. They will likely need to move in the direction of more specialization and train LLMs for specific use-cases. However, this would also be admitting that they are not on a direct path to AGI.

AI needs typing? Why? That seems counterintuitive to me.

I agree on ORMs, but that would apply with or without AI for me personally. ORMs and ActiveRecord are an obvious and well documented anti-pattern at this point in time.

I think it is fair to argue that prompts are not a safety layer at all and can't be relied upon for much.

"Make no mistakes"

I wonder if some of this explains why people have been finding with Opus 5 that running it with lower settings than "High" is producing better or at least just as good results.

> Not so fast. A 2025 paper(opens a new tab) from Northeastern University and the University of California, Berkeley on frontier open-source LRMs showed that between 30% and 60% of their “thinking steps” had “minimal causal impact” on the answers the models produced to benchmark math questions. Chop half of them out, and a model’s performance barely suffers. “We want to be careful when we review these chain-of-thought prompts because they may not be linked to the final output,” said Weiyan Shi(opens a new tab), one of the study’s authors

This comment is so brutally ironic. Wow.

You are technically correct. The best kind of correct.

This is the critical problem with all of this.

Daylight savings time changes, can't be globally banned fast enough really.

Darling, we'll always have W_q, W_k, W_v, and W_o.

Given the IP situation, is shopping it around even on the table?

They are mostly "faster" than the majority of humans. They are rarely better than experienced and talented humans at the majority of tasks they are able to do. They are better on both scales on a small thin slice of work tasks.

I dunno I was listening to someone explain this on a podcast and to put up something on the scale of a typical land based data center would make it, by orders of magnitude, the largest spaced based object ever made and most of the size would be the radiators.

I have the same car and want to do this, but not for the reasons the author noted but because the GPS unit in the car is broken when paired with Carplay and has the wrong compass heading causing navigation to be completely useless.

I have reported this to Toyota multiple times with videos detailing the problem and they have denied the problem and ultimately when faced with the evidence simply refused to fix it.

I've been a big fan of Toyota's Production System and their management culture, but this experience has really diminished the brand for me. I realize these problems exist with all cars today. The pattern seems to be to foist low-quality hardware and software on their customers and take no responsibility for the results. Software bugs aren't what they consider a "typical car problem" so they simply don't fix them.

Man I thought this was about the new Stargate show which they are filming in the UK. I was trying to square why OpenAI would be in charge of that.