I mean, it sounds like reviews and tests are already their standard practice, and explicitly part of their AI practice. So it should have worked, right?
I don't think most commenters have read the article. I can understand, it's rambly and a lot of it feels like they created a thesis first and then ham-fisted facts in later. But it's still worth the read for the last section which is a more nuanced take than the click-bait title suggests.
If you want server-side compilation, you could just run the xslt transform in ci/cd. It would still be a simpler solution than Jekyll in some regards, but I probably wouldn't do it for more than hobby projects
Either way, to be clear: the 28.4% -> 22.4% is human performance vs human performance (before and after "exposure to AI"). There are no numbers provided on accuracy with the use of AI.
Except the paper doesn't say that the doctors + AI performed better than doctors pre-AI. It is well documented that people will trust and depend on AI, even if it is not better. It is not clear in the paper, but possible that this is just lose-lose.
Using the formula for black hole density, a black hole of this mass would have an average density about the same as the near-vacuum atmosphere of Mars(!)
No, whether this tech could lead to a post-scarcity future is debatable. The naive part is thinking that it will lead to it, given the way we use it and the power structures that profit from it.
Is it confirmed that synthetic data was used for gpt-oss training? I didn't pick up on that in the press release or see it elsewhere. Did I miss it or is Sean speculating that it is the case?
The idea of US private efficiency is overblown. It doesn't take long working in a large US company to see massive delays, red tape, duplicate work, self-sabotage, or favoritism. Even when it is fast, that doesn't equate to efficient. It's not uncommon to see 8 figures put into throw-away work.
"AI" is just a misleading and unhelpful term, exactly because it causes people to assume that there are properties we associate with intelligence (abstract thought, planning, motivations, emotions) present in anything given the term. That is easier to correct when someone is referring to a logistic regression. I think that "AI" has clung to LLMs because they specifically give the illusion of having those properties.
Most of the cost of power is not the cost of the electrons. You have to account for the transmission and distribution infrastructure, the operation of the grid, the safety programs, customer service, etc. Retail rates bake that all in. When you run local solar, they are still providing all of that for you when you need it, plus the distribution of the power you sell back to them. What would be "fair" would be to net meter at the value of the power, minus the cost of the grid and backup generation. The cold truth is that this would still be a negative number, i.e. what would truly be fair is for the homeowner to pay a fee to the power company even when they are running off of their own solar 24/7.
My point is that their rant is not about that definition, even if they reference it. Again, they are not wrong in their concerns at all! It's just that having a PFAS'd internet is not a "sign of model collapse". A sign of model collapse would be new LLMs having degraded performance.
> You’ll spend 5-10 minutes knocking it back into your own style
You lost me here. I have often found it to be far more than a 10-minute style issue, but fundamental misunderstanding of the code purposes that I need to fix.
Ok, yes, but that's not what model collapse is. I was expecting an interesting finding on LLM training stalling due to AI pollution of the internet (the training data). Instead this a rant about the pollution of the internet. All good points and it will lead to model collapse, but this is a clickbait headline.
I don't want to hate, what you built is really cool and should save time in a data scientist's workflow, but... we did this. It won't "automate most of the ML lifecycle." Back in ~2018 "autoML" was all the rage. It failed because creating boilerplate and training models are not the hard parts of ML. The hard parts are evaluating data quality, seeking out new data, designing features, making appropriate choices to prevent leakage, designing evaluation appropriate to the business problem, and knowing how this will all interact with the model design choices.
Between the lines, you highlight a tangental issue: execs like Zuckerberg think easy/automatable stuff is 90%. People with skin in the game know it is much less (40% per your estimate).This isn't unique to LLMs. Overestimating the benefit of automation is a time-honored pastime.
> They just seem to be regular right wing by American standards.
Correct, both have an agenda of subverting the rule of law to violate human rights of immigrants, under a flag of white nationalism. There is a single word for this.
Oh cool, horrendous things like this have been done for years. I guess it's fine then, human rights violations aren't real if someone else did them too. /s