Why on earth wouldn't you need to look at the P value? You give good arguments for the soundness is the methodology, but you still need to look at the results and their statistical significance.
Their P value is 0.02, which is good but certainly not definitive. Also the effect is kind of small, 3.5% reduction in diagnoses.
> he was described as the most corrupt official in Chinese history... Heshen is remembered as one of the richest men in history... His total property was... reputed to be equivalent to the imperial revenue of the Qing government for 12 years.
I think the concept of "AGI-complete" was interesting but had been falsified. LLMs have jagged intelligence, meaning that they are good at some things while being counterintuitively bad at other things that seem much easier. Especially given that a lot of software engineering can be trained via RL, it's entirely plausible that they will get extremely good at that while lagging in other things.
Totally agree. I'm a scientist, and like most scientists I have some specialized skills that most of my colleages don't. AI has empowered them to learn and build things that they might have otherwise needed me for. But there have been quite a few cases where it led them very far down a wrong path. This has started happening way more often in the last few months.*
We've known since the beginning that AIs confidently say incorrect things. But now that they can speak confidently about very complex topics, and mostly say correct things, we are letting our guard down and lots of subtle falsehoods are slipping through.
*In one case, I was able to put things back on track because the AI suggested my colleague talk to me; somehow it figured out we were co-workers.
Last I heard, Cerebras chips were entire wafers and would be extremely expensive. How could OpenAI possibly have enough of these to serve a popular model at scale?
I used to think this was the explanation, but I was told by a particle physicist that this is actually not correct. Unfortunately I don't remember the correct argument (and I'm not sure I understood fully it in the first place)
The estimate that AI companies need to replace 27% of jobs to service their debt is interesting. But at least Anthropic and Meta seem to have their eyes on replacing software engineers.
There are ~1.6M software engineers on the US [0], earning a bit under 150k/year on average [1]. If AI companies captured all of that spend, that amounts to about 250B/year. The article assumed that they need around 300B/year to keep up with their debt.
At least based on Meta's recent behavior, forcing 30-50% of developers to switch to data labeling, it looks like that is actually their game plan.
> Atlas would need to learn new factory tasks in a day or two and reach 99.9% reliability before it could be truly useful on the floor
Progress in robotics has been impressive, but is there any evidence that we are approaching this point? How many days are needed to teach a robot a task at even 90% reliability? Given that most companies are still only showing of demos, that number looks to be way more than 2...
Yeah it's sort of alarming when you think about hooking up models to take action in the real world and telling them it's just a game. Several scifi stories have it as a plot twist that humans think they are playing a game but are killing actual people. I'm not sure if the same twist shows up for AIs but it seems like an increasingly real possibility.
That is incredible. 2.5 hours underwater, 1.5 hours of CPR. They were instructed not to start rewarming him until he could be given more comprehensive treatment at a hospital. They list 'death' as a differential diagnosis...
He didn't come out unscathed though. They describe his progress:
> At 6-month follow-up, he was giving short commands, standing without support, riding a tricycle, eating soft foods, and relearning simple tasks. Peripheral neuromuscular weakness continued to improve.
which is quite limited for an 8-year old, but remarkable considering the circumstances.
Naive question: will it not be possible for ad blockers to upgrade to ManifestV3? Is there something about it that makes ad blocking much harder. What does Manifest actually do?
All this proves is that economic interdependence doesn't completely prevent war. Who knows how much more violent the 20th century would have been without it.
I know Google's track record prior to the DOW contact is far from perfect, but is it really so hard to understand why it crossed a line for a lot of people? Why are we acting like ad tracking/targeting and developing autonomous lethal weapons are morally comparable?
Economic interdependence is also a big factor. If you depend on someone selling you things or buying them from you, you have a lot more to lose by invading them.
Is a success story line this still possible with coding assistants, or do they basically pull up the ladder that this guy climbed? I don't have enough insight into the job market right now to know.
They've been coming faster and faster for me. First I was blown away by GPT2, specifically the fake news article about talking unicorns. Just stringing together a few sentences while maintaining logical coherence was very impressive at the time.
Then it was models like Minerva that could actually solve math problems, and the discovery that LLMs were one-shot learners and could write code.
After that, the improvement felt pretty steady, with IMO gold feeling like a watershed moment.
And recently OpenAI's solution to the planar unit distance problem is starting to actually freak me out a bit.
Their P value is 0.02, which is good but certainly not definitive. Also the effect is kind of small, 3.5% reduction in diagnoses.