3 ms·

Both new Sonnet and gpt-4o still fail at a simple: "How many w's are in strawberry?" gpt-4o: There are 2 "w's" in "strawberry." Claude 3.5 Sonnet (new): Let

Both new Sonnet and gpt-4o still fail at a simple: "How many w's are in strawberry?" gpt-4o: There are 2 "w's" in "strawberry." Claude 3.5 Sonnet (new): Let me count the w's in "strawberry": 0 w's. (same question with 'r' succeeds) What is artificial about current gen of "artificial intelligence" is the way training (predict next token) and benchmarking (overfitting) is done. Perhaps a fresh approach is needed to achieve a true next step.

fassssst 2y ago
They are trained on tokens not characters.
redox99 2y ago
There's always that one tokenization error comment
ssijak 2y ago
Can we stop with these useless strawberry examples?
wild_egg 2y ago
It's bad at directly working on classical computer problems like math and data processing. But you can do it indirectly by having it write a program that produces the correct result. Interestingly, I didn't even have to have it run the program, although usually you would write a tool which counts the number of w's in "strawberry" and return the result Which produced: Here's a simple Python function that counts the number of 'w's in the word "strawberry" and returns the result: ```python def count_w_in_strawberry(): word = "strawberry" count = word.count('w') return count # Call the function and print the result result = count_w_in_strawberry() print(f"The number of 'w's in 'strawberry' is: {result}") ``` This tool does the following: 1. We define a function called `count_w_in_strawberry()`. 2. Inside the function, we assign the word "strawberry" to a variable called `word`. 3. We use the `count()` method on the `word` string to count the occurrences of 'w'. 4. The function returns the count. 5. Outside the function, we call `count_w_in_strawberry()` and store the result in the `result` variable. 6. Finally, we print the result. When you run this code, it will output: ``` The number of 'w's in 'strawberry' is: 1 ``` This tool correctly identifies that there is one 'w' in the word "strawberry".
- jandrese 2y ago
  I always thought the halting problem was an academic exercise, but here we see a potential practical use case. Actually this seems pretty dangerous letting the LLM write and automatically execute code. How good is the sandbox? Can I trick the LLM into writing a reverse shell and opening it up for me?
  - okasaki 2y ago
    I'm not sure it's still the case, but I've had ChatGPT run shell commands. But I don't know what you could do since it's ephemeral, doesn't have internet access or root. Plus I'm sure they have security scanning.
int_19h 2y ago
There are models around that will answer this question perfectly. But why would anyone care? If you want to count w's in a word, that's a Python one-liner, you don't need an LLM for that (although the LLM can write it for you).

fassssst 2y ago

They are trained on tokens not characters.

redox99 2y ago

There's always that one tokenization error comment

ssijak 2y ago

Can we stop with these useless strawberry examples?

wild_egg 2y ago

It's bad at directly working on classical computer problems like math and data processing. But you can do it indirectly by having it write a program that produces the correct result. Interestingly, I didn't even have to have it run the program, although usually you would write a tool which counts the number of w's in "strawberry" and return the result Which produced: Here's a simple Python function that counts the number of 'w's in the word "strawberry" and returns the result: ```python def count_w_in_strawberry(): word = "strawberry" count = word.count('w') return count # Call the function and print the result result = count_w_in_strawberry() print(f"The number of 'w's in 'strawberry' is: {result}") ``` This tool does the following: 1. We define a function called `count_w_in_strawberry()`. 2. Inside the function, we assign the word "strawberry" to a variable called `word`. 3. We use the `count()` method on the `word` string to count the occurrences of 'w'. 4. The function returns the count. 5. Outside the function, we call `count_w_in_strawberry()` and store the result in the `result` variable. 6. Finally, we print the result. When you run this code, it will output: ``` The number of 'w's in 'strawberry' is: 1 ``` This tool correctly identifies that there is one 'w' in the word "strawberry".

jandrese 2y ago
I always thought the halting problem was an academic exercise, but here we see a potential practical use case. Actually this seems pretty dangerous letting the LLM write and automatically execute code. How good is the sandbox? Can I trick the LLM into writing a reverse shell and opening it up for me?
- okasaki 2y ago
  I'm not sure it's still the case, but I've had ChatGPT run shell commands. But I don't know what you could do since it's ephemeral, doesn't have internet access or root. Plus I'm sure they have security scanning.

int_19h 2y ago

There are models around that will answer this question perfectly. But why would anyone care? If you want to count w's in a word, that's a Python one-liner, you don't need an LLM for that (although the LLM can write it for you).