To a guy who knows nothing about nuclear physics, an atomic bomb is very abstract until it blows up and becomes very concrete. Should we wait until it explodes to consider the risk?
No, layer looping increases effective depth, but it still has to go through decode. So it's more like they increased number of layers from 100 to 200 without increasing number of parameters.
"Latent reasoning" is rather trivial - you can just replace unembed-embed step with a MLP. But labs don't do that largely because they want to read the output of unembed.
The additional layers provide additional computation without going through one or more dec/enc cycles in between. Whether or not that impacts interpretability of the final token stream depends entirely on the maximum depth permitted (and how efficient the model in question is).
You need to train data for a BERT-based classifier, and then there's a risk that it will pick up specific biases from the data instead of what you want.
As far as I understand, the idea of Jev is zero-shot or few-shot classifier: it learns a lot of stuff at pre-training, but unlike a classic LLM it doesn't need to learn how to chat, so it can be much smarter at a particular size
Education system was set up in XIX century. It's not clear how much of it is necessary in XXI century.
Back in the day religious books were copied by scribes educated in a monastic tradition. Now printers can print them in a completely godless manner but the result isn't any worse.
It's low because there's no competition. MDs are like medieval guild: once you're in, you're set for life. Restrictive regulations are lobbied by MD associations, which limit competition.
Ha ha, but the joke is kind of wrong. While anyone with an MD degree is entitled to the honorific "doctor" that doesn't mean they can legally practice medicine. Becoming a licensed physician generally requires completing a residency program, and the worst students sometimes fail to get matched to the limited number of residency slots.
Yes it's a joke, but to quote some numbers in the US: something like 93% of the US graduates matched to a residency program [1], although the numbers were a lot lower for international students applying for US residency. So if you're talking about US medical students, you have to be in the bottom 10% to flunk out.
The attrition for lawyers is higher, neither has anything on PhDs where the number of "practicing" doctors drops off constantly during and after the degree, and only really levels off when O(10%) of them land faculty jobs.
Probably better solution is to upskill nurses + AI to do handle all the simpler tasks like prescribing standard treatments, etc. There's already a concept of mid-level practitioner which can be expanded.
This is so hilariously false and completely inverted that it shocks me that people not in medicine can believe it.
Out of respect for you as a fellow intelligent HN commenter — you are deeply misinformed, and I would urge you to reconsider your perspectives on this.
Every study shows that utilization of APPs and nurse practitioners Leads to decreased quality of care and a significant increase in utilization of other healthcare resources, like the emergency department and imaging, that better-trained physicians don't need.
Less-trained providers misdiagnose cancers, refer patients to the wrong specialists, overprescribe antibiotics, and generally cost the system significantly more in overall health load than if we had better-paid general practitioners.
There is an argument that not enough physicians go into general practice, which is true, but it's because subspecialties are in such high demand that they're generally better paid. The unfortunate fix is that we need to find a way to better compensate primary care, even though Medicare physician reimbursement rates continually decline and our health insurance system is not well structured to support this kind of primary care model.
I've left a few links below if you'd like to read them:
There are several papers demonstrating AI is at least as good at diagnostics as fully qualified doctors. So why would NP + AI be worse? AI should compensate for the lack of knowledge.
And I'm not saying NP + ChatGPT - it should be properly calibrated system which would defer to a 'proper doctor' in more complex cases.
> AI is at least as good at diagnostics as fully qualified doctors
If the prompt is an expert-written board question! Not so with inferior prompts [0]. Critically, you need deep medical knowledge to interact correctly with the agent.
What you're asking is basically: "If we take someone out of a three month dev bootcamp, and have them prompt Claude, why can't they be as good as a four year CS grad?"
I doubt that you would feel similarly about expertise in your own field.
There have been a huge progress in AI reasoning in the past 2 years. GPT-4o mentioned in the article would struggle with high-school math problems, OTOH GPT-6 can solve problems beyond capability of professional mathematicians.
I'd wager GPT-6 would not depend on high-quality prompts, although it might still be good to get a trained person to enter information and do a sanity check.
Yes, point taken. I am well aware of the intelligence explosion and I am very AI pilled.
But the core problem still exists based on the clarity of the prompt and the prompter.
Seeing my wife's depth of intuition and expertise as a physician, compared to my comparatively caveman-like prompts when asking for health advice, there's a massive delta in accuracy.
Have you ever had a significant health issue and tried to get help from a model? Even something like GPT 6 Pro? The firehose and confusion you will generate trying to figure out answers are far less helpful than having directed interaction with a physician.
Following up on this because I've been thinking about why this is. I believe a big part of it is just that the intuitive parts of medicine are learned essentially by internship and therefore are not well encoded into the models. Unlike software, where you can have deterministic output that the models can train on, medical outputs are very fluid and dynamic. What works well in a paper, even though we claim to do evidence-based medicine, may be very far from what a "good" physician does in practice.
The issue is that the AI is at best what's in papers and medical records, which often forgoes the core thing that might lead a physician to uncover something or take a different approach with the patient.
> my ideal sandboxing is "prevent writing to anything outside this dir but still allow reading to most things so that I don't have to manually copy things into a container/VM"
That's what Codex does out of the box, and it's not good against malware - i.e. a rogue npm packet (or even just codex after prompt injection) can read your ssh key and send it to the attacker.
As I said, opaque rules built into the harness rub me the the wrong way. They could change in an update without anything making it clear. Plus, I don't use Codex outside of work (I don't have any active paid subscriptions LLM offerings).
> it's not good against malware - i.e. a rogue npm packet (or even just codex after prompt injection) can read your ssh key and send it to the attacker
Ignoring the repeated references to software I don't personally use, I never said I was trying to hedge against malware. The use case for me is when I'm running agents directly based off of prompts that I give them and asking them to modify some files. If I wanted a solution for running code I didn't trust, I wouldn't rely on what I wrote, because that's not the intended use case at all.
Making model to consists of many small modules is inefficient on GPU, especially as routing adds data dependencies, etc, and especially with pytorch (compared to a custom kernel).
The difference might be smaller on a CPU which has limited parallelism.
But it's basically equivalent to a very deep model which might be problematic for training.
Luddites were complaining about having their lives upended and being thrown into generational poverty. That is what wide-scale "disruption" looks at the receiving end.
It seems obvious then that the answer is supporting the much smaller displaced people, not stopping progress for the entire population (and future generations).
I'd say this social contract which got "broken" never existed in the first place.
Even before AI we heard lots of complaints like "I made a popular open source library which is now used by corps with trillion-dollar market cap and I don't get anything out of it; halp". There was always some kind of a conflict, now the nature of the conflict just changed
reply