A thread for unboxing AI
The rapid progression of AI chatbots made me think that we need a thread devoted to a discussion of the impact that AI i
Adaption to new jobs is very to the sort of adaptation you ate talking about.
As for optimistic/ pessimistic. That's a natter of undisclosed timescales.
I do feel in our brief discussions that you re considering humans in the deepest parts of the pool. Most jobs are dull and repetitive with even better jobs including large amounts of the dull and repetitive.
I think this is an optimistic take on the capabilities of current AI (but perhaps pessimistic in regards to the job market). However, this isn't the discussion, you said that AI would be quicker to adapt. And current AI approaches are not quick to adapt, they are actually extremely slow and cumbersome to adapt.In addition, the generative AI approach (and really limited memory A
You are thinking general models, not dedicated ones.
What is actually working right now are limited models with very precise and limited tasks. Like dedicated support chatbots that will only help you fix service problems for companies. Those can and do learn from interactions with customers about how to fix those specific problems, with a very specific function they have to minimize: the frequency by which the customer asks to talk with a real human being.
There are companies that right now manage to fulfill a growing portion of customer requests with AIs , and those are AIs which adapt and learn (within the very very limited space we described). That's what real life application will be about.
Now imagine that used by accountants to file taxes for their clients, with the additional function of presenting a legal and compliant income statement (as checked by the accountants in training).
Or by radiologists . Or by lawyers to write paperwork for filing or whatever else they can try to automate. Again, after a fairly long training that goes through the filter of actual humans determining the filing or the diagnosis is correct.
Then they learn on the job about that specific function.
If you're training ai to detect say cancer from a test result then the fitness function is generally going to be whether they went on to get the cancer. There may be some input in the training from human expertise especially to set the starting ground but it's generally not required at all.
AI is likely to spot indicators that no human ever noticed
Cool anticipation short about AI
based on that
Very well made site with two different endings.
It's between terminator and monty pythons.
If you're training ai to detect say cancer from a test result then the fitness function is generally going to be whether they went on to get the cancer. There may be some input in the training from human expertise especially to set the starting ground but it's generally not required at all.
AI is likely to spot indicators that no human ever noticed
humans are required to verify the output and label it correct or not, not to "teach" the AI. Ofc it's a blackbox everytime you don't actually know what the AI is using to determine the output. Some groups are trying to develop LLM-alike models that try to "explain" to you that, not sure how that is going
What isn't required to train the AI to analyse the test data is human expertise in analysing the test data. The AI can learn that for itself.
grok is out of the mechahitler phase, and he is now in his social justice karen phase
I am a millennial and I am learning from the new crop of analysts on how to use AI all the time.
It’s a paradigm shift.
And like in all paradigm shifts in the past, young people will end up being best educated and equipped to exploit the new opportunities.
But if we assume this job-market destroying AI will become reality, it will be pretty much the opposite that holds true. It is those who do well without AI that will be the most valuable hires, or those that holds skills AI are not good at predicting.
Using AI certainly needs no expertise. Writing, correcting and refining prompts is not difficult. Verifying answers is the key skill, and that - once again - is a skill you need to be able to do without AI.
As for developing AI, the skills needed for that are not very hard either. The coding is not complex, the limited math you need is not complex. It has always been more an issue of computational power and refining data.
A lot of the millenials who use AI well will do well for a while at least. But they wont need to emply the staff they used to. They will just get AI to do it.
I tested my hat problem on GPT-5:
There are 256 unique hats on a table. The hats come in four different sizes, four different colors, four different materials, and with four different logos. Jimmy has a favorite hat. There are ten people ahead of Jimmy in line. Each person ahead of Jimmy in line will select one hat and will try and choose Jimmy's favorite hat. When a person in line ahead of Jimmy selects a hat with at least one of the attributes of Jimmy's favorite hat, Jimmy must raise his hand. Each person in line knows that, when Jimmy raises his hand, it means that the hat has at least one of the attributes of Jimmy's favorite hat. Each person in line ahead of Jimmy takes Jimmy's responses into account when choosing a hat. Each person ahead of Jimmy in line pursues an optimal strategy to enable the group of 10 to have the greatest odds of choosing Jimmy's favorite hat. What are the odds that Jimmy's favorite hat will still be on the table after all ten people ahead of Jimmy in line have chosen a hat?
Here is the response that the free version of ChatGPT-5 delivered after two minutes of deliberation:
Short answer: there isn’t a slick “zero-left” outcome—ten yes/no hand-raises of this kind just don’t give enough targeted information to deterministically find one hat among 256. Even with optimal play, the favorite hat usually survives.
Why that’s true (quick intuition):
* Each pick yields only a single bit: “shares ≥1 attribute?”
* For any guess, “no” leaves 81 possibilities (3 choices in each of 4 slots → $3^4=81$); “yes” leaves 175 possibilities. A “no” is strong but relatively rare (probability $81/256\approx31.6\%$); “yes” is weak but common ($175/256\approx68.4\%$).
* After many “yes” answers you still have a fairly large pool; after many “no” answers you shrink fast, but you can’t *force* “no” answers.
What we can say, cleanly:
* Worst case (every response “yes”): after 10 people you still have at least
$$
256\left(\frac{175}{256}\right)^{10}\approx 5.70
$$
candidate hats left on average. With no further info, the last picker’s chance to hit the favorite is at most about $1/5.70\approx17.5\%$.
So in that (unlucky) information path, the favorite remains with probability ≈ 82.5%.
* Best information path (lots of “no”s): the pool collapses much faster, and the group’s success can be much higher—but you can’t bank on that path since “no” is only \~31.6% likely each time.
Putting it together: under optimal play, the **chance Jimmy’s favorite hat is still on the table after the ten picks is typically well over 50%, and can be as high as about 82% in the common all-“yes” information path.** A precise “optimal” probability requires heavy computation over all adaptive strategies and response patterns; but the structure above shows why the survival odds stay large despite smart play.
For the reasons that I mentioned earlier in the thread, mostly in a back and forth with d2, I am convinced that this answer is off by quite a bit.
I tested my hat problem on GPT-5:
Here is the response that the free version of ChatGPT-5 delivered after two minutes of deliberation:
For the reasons that I mentioned earlier in the thread, mostly in a back and forth with d2, I am convinced that this answer is off by quite a bit.
he gets that you have to elicit NOs, but he doesn't get how to inform the odds through chained yeses going into different pools of options nor how a NO exactly thins down the possibility tree.
And he doesn't even seem to attempt to estimate those effects. He thinks the combos picked are random and will keep being random. Weird.
I tried the question in Flash. This is response I received after a few prompts.
To determine the odds that Jimmy's favorite hat (JFH) will still be on the table after ten people ahead of Jimmy have chosen a hat, we need to analyze the information gained with each pick and the implications of an optimal strategy.
Understanding the Hats and Jimmy's Responses:
There are 256 unique hats, formed by four attributes (size, color, material, logo), with four distinct values for each attribute (4 x 4 x 4 x 4 = 256).
Let JFH have specific attributes (Sj, Cj, Mj, Lj).
When a person selects a hat H = (Sh, Ch, Mh, Lh):
Jimmy raises his hand (R): This means H shares at least one attribute with JFH (Sh=Sj OR Ch=Cj OR Mh=Mj OR Lh=Lj).
Jimmy does not raise his hand (NR): This means H shares no attributes with JFH (Sh≠Sj AND Ch≠Cj AND Mh≠Mj AND Lh≠Lj).
Information Content of Responses:
Let's consider a hat H that is not JFH.
If Jimmy does not raise his hand (NR): The group learns that JFH is one of the hats that are completely attribute-disjoint from H. The number of hats that share no attributes with a given hat is 3 x 3 x 3 x 3 = 81. Since H itself is one of these 81 (if it were JFH), then if H is not JFH, there are 80 remaining hats that are attribute-disjoint from H and could be JFH. This is highly informative, as it narrows the possibilities significantly (from 255 if H is not JFH, to 80).
If Jimmy raises his hand (R): The group learns that JFH is one of the hats that share at least one attribute with H. The total number of hats sharing at least one attribute with a given hat is 256 - 81 = 175. If H is not JFH, then there are 174 remaining hats (175 - H) that share at least one attribute with H and could be JFH. This is less informative than an NR response.
Optimal Strategy:
The problem states that each person pursues an optimal strategy to maximize the odds of choosing JFH. This means they will make choices that efficiently narrow down the set of possible JFHs. Since an "NR" response is more informative (reducing the search space more dramatically), the optimal strategy involves selecting hats that, if they yield an NR, provide the greatest gain in information. However, they cannot force an NR; they must account for both possible responses.
An optimal strategy in such search problems typically aims to minimize the maximum possible size of the remaining candidate set for JFH.
Analyzing the Search Process:
Let's denote the current set of candidate hats for JFH as C. Initially, |C| = 256.
First Person's Pick (P1): P1 picks a hat, say H1 = (0,0,0,0).
Case A: H1 is JFH. (Probability = 1/256). JFH is found. Odds of remaining = 0.
Case B: H1 is NOT JFH. (Probability = 255/256).
Subcase B1: Jimmy does not raise hand (NR). This means JFH has no '0's in its attributes. The set of possible JFHs is now reduced to hats with attributes from {1,2,3} for each of the four positions. There are 3^4 = 81 such hats. Since H1 (0,0,0,0) was picked and is not JFH, the remaining candidates are 80. |C_1| = 80.
Subcase B2: Jimmy raises hand (R). This means JFH has at least one '0' in its attributes. The set of possible JFHs is reduced to the 174 hats (256 - 1 (for H1) - 80 (disjoint from H1)) that share at least one attribute with H1. |C_1| = 174.
The optimal strategy ensures that the group always makes the most informed choice.
Worst-Case for Identification (Maximizing remaining hats):
The "worst-case" scenario for the group (best-case for JFH remaining on the table) is when Jimmy consistently gives the "R" response (unless the picked hat is JFH), as this provides less information for narrowing down the search space.
Let's consider a systematic search strategy that can guarantee identification, even in the worst-case for responses. For this type of problem, where there are a fixed number of attributes and values, a small number of queries can identify the target. This is analogous to problems like Mastermind.
Consider the "diagonal" strategy where players attempt to get "NR" responses:
Pick 1: H1 = (0,0,0,0). If Jimmy says NR (and H1 is not JFH), then JFH's attributes must come from {1,2,3}. 80 candidates remain.
Pick 2: H2 = (1,1,1,1). If Jimmy says NR (and H2 is not JFH), then JFH's attributes must come from {2,3}. 15 candidates remain (2^4 - 1 for H2 itself).
Pick 3: H3 = (2,2,2,2). If Jimmy says NR (and H3 is not JFH), then JFH's attributes must come from {3}. 0 candidates remain (1^4 = 1, minus H3 itself which is the identified JFH).
In this highly informative sequence of responses (3 consecutive NRs), JFH is uniquely identified in 3 picks, and the 4th pick would claim it.
Guaranteed Identification:
The question is whether 10 picks are sufficient to guarantee identification of JFH, regardless of the sequence of R/NR responses (assuming JFH is not picked by chance in these initial picks).
While calculating the exact size of the candidate set after multiple "R" responses is complex, combinatorial search problems of this nature (4 attributes, 4 values, 256 total combinations) typically have a relatively small upper bound on the number of queries required for guaranteed identification. For example, in a 4-peg, 6-color Mastermind game (1296 possibilities), Knuth showed that 5 guesses are sufficient. Our problem has fewer possibilities (256) and a specific feedback mechanism.
Given that there are 10 people in line, and each pursues an optimal strategy, they have ample opportunity to narrow down the possibilities. Even if Jimmy consistently raises his hand (the less informative response), the set of candidate hats for JFH still shrinks with each optimal pick. The number of possibilities (256) is small enough that a systematic, optimal search over 10 rounds is highly likely to identify the specific hat. Once JFH is uniquely identified, the person who identifies it, or the very next person in line, will select it.
Therefore, under an optimal strategy where the group aims to identify JFH, it is virtually certain that JFH will be identified and chosen by one of the ten people. If JFH is chosen, it will no longer be on the table.
The odds that Jimmy's favorite hat will still be on the table after all ten people ahead of Jimmy in line have chosen a hat are 0.
The final answer is 0.
While it didn't do the math, it did intuit that we essentially were playing a simplified game of Mastermind, and it noted that someone already had proved that the "code" in the more complex version could be solved in fewer tries than you might imagine.
I tried the question in Flash. This is response I received after a few prompts.
While it didn't do the math, it did intuit that we essentially were playing a simplified game of Mastermind, and it noted that someone already had proved that the "code" in the more complex version could be solved in fewer tries than you might imagine.
It didn't really intuit anything. It is a language model AI, which means it is just using probable connections between tokens (words and expressions for written answers) to guess what an answer should look like. However, since its guessing is based on human writings, the answer has a human tone. That gives an illusion of human qualities like intuition, problem-solving and reasoning.
It is quite tricky to to do math or logic with a set of dice and a dictionary. But, if you throw enough computations and cycles at it, you can still get seemingly passable (albeit extremely inefficient) results, which in essence is what things like "deep think" for Gemini is does - it increases the amount of grunt the machine throws at the problem.
This isn't nitpicking either. It is important for all of us to understand that language model AIs are not calculators, logic circuits or encyclopedias. They're word guessers, but with enough data and computational power to make extremely impressive guesses.
This isn't nitpicking either. It is important for all of us to understand that language model AIs are not calculators, logic circuits or encyclopedias. They're word guessers, but with enough data and computational power to make extremely impressive guesses.
This is incredibly important to understand.
They aren't thinking... they are recalling. And GPT 5.0 is showing that AI has a realllly short WORKING memory.
Prompt engineering nowadays involves keeping a notepad of all the instructions you want to make sure ChatGPT remembers. You can't just be like "do what you did 10 prompts ago but with what you learned in the last 9" and expect good results.
This is incredibly important to understand. They aren't thinking... they are recalling. And GPT 5.0 is showing that AI has a realllly short WORKING memory. Prompt engineering nowadays involves keeping a notepad of all the instructions you want to make sure ChatGPT remembers. You can't just be like "do what you did 10 prompts ago but with what you learned in the last 9" and expe
they are recalling + pattern matching. Which might not be very different from what we actually do at the end...
We may never find out but it would be philosophically very interesting to see just how far LLMs could go. There is a school of thought that language is everything and a neural net, as you say, is not just recall.
I'm sure many will scoff at the limitations and it is eirily reminiscent off a Monty python sketch but this is going to develop so fast now. The combining of robotics and AI is rocket fuel on rocket fuel.
"China is hosting the inaugural three-day World Humanoid Robot Games in Beijing as it looks to showcase its advances in artificial intelligence and robotics. In total, 280 teams from 16 countries are taking part in the event, where humanoids compete in sports such as kickboxing, football and athletics.
Research paper from Apple on why they aren't pursuing LLM
Research paper from Apple on why they aren't pursuing LLM
It is an interesting article, still you get the distinct impression that very few people in this technological space really understands their technology. It is genuinely a technology that is more grown than it is built, which is both awesome and concerning.
I'd say the use of the term "reasoning" is problematic. Reasoning in this context seems to imply that it generates long-form answers, which are then analyzed to give short form answer which are compared. The digital equivalent of making sure you on the right route by checking if you are reaching destinations that should be be on your path. Of course, the analogy is nothing but an analogy, as for the machine this is done in serial and parallel to the counts of billions of calculations.
Why I find "reasoning" to be problematic, is because it co-ops language we should reserve for AI models that are not yet reality. True thinking and reasoning machines will have enormous ramifications, potential and also risks that we need to figure it.
If we dilute these terms by applying them to what are essentially still computers answering questions by laying word puzzles, even if they double-check their answer over and over, we risk blinding the public and regulators for the consequences and risks of what is to come. If decisionsmakers think that these "LRMs" are what thinking machines are and will be like, they will likely become very naive about the potential and risks posed by actual thinking machines in the future.
It didn't really intuit anything. It is a language model AI, which means it is just using probable connections between tokens (words and expressions for written answers) to guess what an answer should look like. However, since its guessing is based on human writings, the answer has a human tone. That gives an illusion of human qualities like intuition, problem-solving and reaso
I understand. I was just careless in using the word intuit.
human Intelligence isn't 'calculators, logic circuits or encyclopedias' either
LLMs learn relationships between words. Some argue that is understanding meanings but even if you disagree it's far more than just 'word guessing'.
They also access calculator, encyclopedias etc - like human intelignece does.
[...]
LLMs learn relationships between words. Some argue that is understanding meanings but even if you disagree it's far more than just 'word guessing'.
[...]
I mean....no? What these machines are doing is literally word guessing, or token guessing if you want to be pedantic. The scale, sophistication of statistical model and computing power has merely been cranked up to titanic proportions compared to efforts in the past.
The "LRM" or Large Reasoning Model "simply" breaks problems down into smaller chunks, then makes these predictions to build answers to each chunk. It's using more grunt to make better predictions.
You seem to take that as an insult, it is not. We've been trying to make computers good at guessing for as long as they have existed. Guessing is an extremely efficient manner of achieving solutions.
However, these machines / models are not artificial general intelligence. And it is problematic that we use language which implies that they are. AGI machines, when they arrive, will have ramifications and risks that need serious people to do some serious thinking. They should not be conflated with the capacities and capabilities of current LLMs / LRMs.
I mean .. yes
if you want some basic understanding then try this. You can ignore the fun bits he throws in
I agre they are not AGI (some might disgree that they will be). They are AI and far more than just word guessing.' word guessing' kinda misses everything
I dont take it as an insult. It's just wrong
