This part of the industry is struggling more and more (or more likely silently taking more risks to hide the struggle). It has no path to sustainability and it starts to show.

More in depth look at the launch white paper and the issues covered in the PR. Not much survives scrutiny... there's nothing special with this model.

Indeed, and it's going to get even crazier at some point. I guess somewhat soon but who knows...

One of the dark sides of our industry, and this is is accelerating at a worrying pace. Maybe it's time to look at and fix the whole hardware life cycle?

Are we surprised it's mostly a PR stunt? Not at all. Of course, I agree a lot with the conclusion: we can't trust any claim from those companies. They try to present themselves as labs but mostly try to disguise marketing as research...

Stop looking at the shiny toy, remember the ethics behind them...

It's first a great marketing stunt. The model is likely not the secret sauce though.

Long but very precise piece about why you can likely ignore LLM for development purpose. Starting from older Fred Brooks work is spot on. Indeed whatever will remain of LLM based tools in the years to come, it's much smarter to focus on fundamental skills than chase the new tools. At least, I'm trying to do my share in getting myself and others better at the craft.

Definitely interesting approach. I think neurosymbolic approaches are what we ultimately need so I'm probably biased. At least it means using LLMs for what they're good at (language skills) and only that. Then rely on proper code symbolic models which do the reasoning heavy lifting. I'd expect it can give nice output with smaller models.

It's getting clearer that the industrial LLM complex will have a hard time meeting its targets.

Excellent piece, it show quite well the problem of skipping the "grunt work". Without it you can't really learn your trade (be it astrophysics or anything else). It also shows how the incentives on scientific careers are wrong. It's not new, but when LLM agents become available, things are definitely changing for the worst.

So much this... There are so many organisational problems that churning code faster is likely not what you need. When did we start to obsess with the number of lines of code?

It feels like it's supercharging an old bias... We tend to confuse confidence for competence.

Real innovations come from constraints. The frugal AI movement is clearly where we will see interesting things emerging. Interestingly, those approaches are closer to what AI is about as a research field than the industrial complex which got unleashed with all its extractive power.

The price hike on RAM due to the LLM as a service bubble is really killing interesting fields. Can't we have nice things? Will the arm race end soon?

Unsurprisingly, they need to find new data to feed the monster...

Not peer reviewed as far as I can tell. That said if confirmed by other studies this feels like an important paper. The language flattening might be real and this will have lasting cultural impacts.

I personally think this is where it'll head after the bubble pops. We should be able to recover enough material to have something viable to run locally. The question will be "where the updated models come from?", it might be the public sector helping there and hopefully those will be truly FOSS and ethical (like Apertus).

Or why this latest trend in genAI hype is a fool's errand.

Excellent piece, indeed legal is not the same as legitimate. More often than not the law is lagging behind and things might be wrongly "fixed" at a later date. In that interval that's when our communities need to build its own tools to protect the commons. We're clearly reaching such an inflection point. Interestingly, I think there's is a difference of reaction between the people with a Free Software culture and the ones with an Open Source culture.