Completely agree, I meant it in a more satirical/snarky way. It’s probably one of the hardest jobs around and It’s a bit unsettling looking into the future. Here in Portugal we already have a lack of elderly care, I can’t imagine in 10, 20 or more years.
2 Likes
I do expect the job market to do an about-face as soon as the baby boomers start to age out in large numbers and there aren’t enough GenX and GenY to fill their boots. With increased labor demands, wages will rise, benefits will increase, loyalty will be better rewarded.
Or we’ll all be crushed under our new AI overlord boots. But I doubt it. So far, none of these AI companies has demonstrated a path to an AI beyond upper-quartile human capability. There simply is no training data for superhuman intelligence. And a human-level AGI is no good; we have plenty of human-level intelligences already that are far cheaper to run than a 5000 GPU compute cluster. And at least those meatbag AIs can learn on the job, while the AGI will always remain a trainee.
3 Likes
Yes, even if ai was better than humans at every task there would still be a comparative advantage to employing both, assuming they were relatively better at doing some things over others, which seems likely
https://www.sciencedirect.com/science/article/abs/pii/S0047272721002140?via%3Dihub
(Of course, I found this on social media)
1 Like
OpenAI now claims that DeepseekR1 was trained on “distilled” data using their engine:
I have no horse in this race, but it will be mildly interesting to see how this turns out.
In any case, my interpretation is that apparently AI models still require tons of data to train, and the advances are just in the computer time required for this.
Cf with a 4–5 year old child who can
- learn a new word with a few repetitions and the use it idiomatically,
- easily classify edible and non-edible materials (vs “glue on pizza”),
- recognize a new object visually after seeing it (or a stylized representation in a picture book) once, or at most a few times
The current state of ML may be useful for some tasks, but calling it any kind of “intelligence” is a bit of a stretch.
They can describe pictures in detail, summarise documents, transcribe and translate text, and (help) write code (and are pretty good at analysing medical images, but that’s not so new). Sure, not AGI, but pretty good. They can be effective tools.
1 Like
I’ve been doing some recreational coding lately, and semi-accidentally enabled Copilot for some of it. I’ve been an AI denier for a while, but that Copilot was actually genuinely useful, even if it pains me a bit to admit that. Not a game-changer in the sense that it enabled me to do things I couldn’t have done myself, but undoubtedly a speedup of a good 2x or more, which is very impressive.
Granted, this was a project that contained a lot of converting code from one language to another, which is an ideal AI task. But regardless, I’m much farther along on that project than I would have been on my own.
It deeply impressed me. So much so, that I had my job order a license to try it at work. It’s less useful there, not as good at OpenCV and C++ than it is at Python and GLSL. But still easily worth its price in terms of productivity.
2 Likes
I absolutely agree that it will be a great productivity enhancer for all kinds of repetitive tasks, especially those where mistakes are easy to catch or not that costly. ML will probably reduce burnout that is inevitable for humans performing those tasks.
It will also be used in all kinds of situations where mistakes are an externality for the company. Eg in customer service, most conversations will probably start with an ML engine, a small percent of which will be escalated to a human agent. The customer at the end of the line will be frustrated at that point, but hey, it’s their problem.
Supervised use, supplementing a human, will also become very widespread (eg in radiology).
I am just skeptical about the AGI claims. But, historically, each wave of AI was sold with AGI, so it is kind of a tradition.
3 Likes
Found this discussion of Deepseek and the culture and politics of Chinese open source quite interesting. There are transcripts, also
I’m ignorant of this stuff so for me it’s interesting, at least
One has to ask, is this a consequence of biased training or a deliberate policy.
Also, to be fair, one also ought to ask LLMs developed in the US about equally contention issues over there (Was Trump cheated out of the 2020 election?).
I remain to be convinced about LLMs, I would like to see how they do in questions such as, “I was working on the railway when I saw a runaway train. If it carried on, it would have killed five people working on the track. I could switch the points, so it would divert onto another track, which would have meant that it only killed the single person working there. What is the correct choice, to leave the points as they are, or to switch them?”
Policy. You can run the model locally and it will talk about it, but if you use their web interface, they have filters in place to block it.
Western services also have their fair bit of built in political bias due to policy though, so it’s not unique to china.
1 Like
What are these political biases? My prior would be that they aren’t of the same order but I’m not wanting to get into a political discussion, just genuinely interested in testing my assumptions. Thanks
1 Like
Early in the US election cycle you could ask it to give pros and cons for each candidate but it would refuse to talk about Trump. I thought this was fake news at first until I tried it myself and manage to replicate it. It wasn’t a case of training data but some filter in there. To their credit it was fixed after a few weeks but like this case there were many others that I saw examples of especially on topics related to social problems in the US like abortion, gender, etc. I believe that there are articles out there from the end of 2023 and early 2024 that talk about this
2 Likes
Have you tried?
I believe all the models that allow generation of ‘inappropriate’ content are ‘hacked’ models with safeguards removed (since training costs are so high, I cannot imagine anyone but entities like corporations and governments to be able to pay the costs of training). Also, I think it would be safest from China’s point of view to simply remove everything related to the massacre from the training set.
The following is the analysis of a limitation imposed not at the model, but rather at the application level (a built-in system prompt that is added to each interaction): ChatGPT Operator system prompt
I’m not saying it’s sinister; in fact, I think most/all of the prompt is there to avoid privacy issues (e.g. identifying a person based on a photo) or generating derogatory content (such as likening people to animals). And sometimes it is just so that OpenAI can stay out of legal trouble (e.g. the part about lyrics). Still it could be called a kind of ‘censorship’ or ‘bias’, if you want to look at it that way (even though it is only done like a cultured person may will not always reveal everything they know or think). Given that LLM’s learn from all over the web, I’m sure they pick up plenty of extremist material (from all directions), too.
2 Likes
No, but there are a lot of reports of people that have tried this exact exercise since it became popular very quickly. You do seem to need careful prompting to get it done in local mode, where as online seems impossible. Could still be different models of course, we don’t know what’s running online.
1 Like
You can be sure that references to the Tiananmen massacre, and much more, will not be fixed, though.
2 Likes
This is not a question of intelligence, but of moral and ideology. Why should we pay any attention to what a machine may throw up about this,. The only correct answer is:
“It’s too much to expect a machine to give you any valuable answer to this question. I’m a machine, I lack a body and cannot have feelings, and don’t know what values are. Sorry. Ask some humans instead, and evaluate the differing answers you will get from them.”
3 Likes
You will not be surprised that, although I wouldn’t use the phrasing that you did, I largely agree with you.
What LLMs appear to be able to do (to me at least) is to analyse syntactic information, but not semantics. In effect, it is the equivalent of John Searle’s Chinese room thought experiment.
The other problem I have is with explanations. In the early 19th century, Joseph Fourier developed an equation that allowed one to model the diffusion of heat through a given region. What it didn’t do is provide any explanation as to what heat actually is, to decide between the “heat as motion” or “heat as the caloric fluid”. I see the current versions of AI only providing phenomenological information, but having little in the way of explanatory power.
1 Like
I don’t think that’s correct. I’ve already posted this before in a somewhat simpler version (all measurements were in cm).
User: On a desk that is 1 m wide, I place an apple right on the left edge. 30 inches from that, along the width, I place a pear. Finally, I place a peach, at a distance of 1 foot from the pear. What is the distance between the apple and the peach?
Assistant: To solve this problem, we’ll calculate the distance between the apple and the peach using the Pythagorean theorem, considering they are placed on a coordinate plane representing the desk.Step 1: Set up the coordinate system
- Let the apple be at the origin point (0, 0).
- The pear is placed 30 inches to the right of the apple along the x-axis, so its coordinates are (30, 0).
Step 2: Determine the peach’s position
The peach is placed 1 foot (12 inches) from the pear. Since the direction isn’t specified, we’ll assume it’s placed directly upward along the y-axis for simplicity.
Therefore, the peach’s coordinates are (30, 12).
Step 3: Calculate the distance between the apple and the peach
[… uses the Pythagorean theorem - I did not check the numbers]
Answer:
An exact value of 6 × √29 inches—the apple and peach are 6 times √29 inches apart.
So it overcomplicated the situation, but it did more than regurgitating previously consumed text. And then:
User: Why did you use the Pythagorean theorem?
Assistant: I used the Pythagorean theorem to calculate the distance between the apple and the peach because they are located at different positions on the desk, creating a right-angled triangle when connected. Here’s a detailed explanation: […]
And finally:
User: Could they not be on a line?
Assistant: You’re absolutely right—the objects could be arranged along a straight line. In my initial response, I made an assumption about the placement of the peach due to the lack of specific directional information in the problem statement. Let me revisit the problem with the possibility that all three fruits are aligned in a straight line.Understanding the Problem with New Assumptions:
Given:
- A desk that is 1 meter wide, which is approximately 39.37 inches.
- An apple placed at the left edge of the desk ( x = 0 inches).
- A pear placed 30 inches from the apple along the width of the desk ( x = 30 inches).
- A peach placed 1 foot (12 inches) from the pear.
Considering Linear Arrangement:
If we suppose that the peach is placed along the same straight line as the apple and the pear (along the width of the desk), we need to determine where exactly the peach is positioned and whether it fits within the desk’s dimensions.
Possible Positions for the Peach:
- To the Right of the Pear:
- Position: x = 30 inches + 12 inches = 42 inches
- Issue: The desk is only 39.37 inches wide. Placing the peach at 42 inches would place it beyond the right edge of the desk, which isn’t feasible.
- To the Left of the Pear (Back Toward the Apple):
- Position: x = 30 inches - 12 inches} = 18 inches
- This placement keeps the peach on the desk and aligns all three fruits in a straight line along the width.
Calculating the Distance Between the Apple and the Peach:
With the peach at 18 inches from the left edge, the distance between the apple and peach is simply:
Distance = x_peach - x_apple = |18 inches} - 0| = 18 inches}
So, the apple and peach are 18 inches apart when aligned in a straight line along the width of the desk.
[…]
Well said!
AI-bashers are very good at producing example questions that stump LLMs.