General > General Technical Chat
OpenAI ChatGPT for electronics design
xrunner:
--- Quote from: BrianHG on February 27, 2023, 12:38:52 pm ---I guess anything codded, or any other text at all beginning in 2023 onward is now suspect to being partly or fully done by ChatGPT. Now, you may never again be able to hire employees where you can be sure their skills are genuine, or that they have any native authentic skills at all other than being able to 'tickle' ChatGPT in the right way to achieve results.
Ten years from now, such natural skills will begin to die off and those with said skills even lateron will be of a special breed.
--- End quote ---
Yea, I guess it's the same analogy when the calculator came out. I remember when kids (the more well-off kids) brought them to school. I created a real dilemma. "Why do I have to learn how to solve math problems on paper when I can use this calculator?". Now it isn't a problem. They will say the same thing about writing "Why do I have to write an essay or report when this tool on the internet can do it for me?"
A lot of changes are coming in society, because this is just the beginning. Contrary to other opinions machines, IMHO will have intelligence eventually. Or they will be able to simulate it so well that you can't tell. If you can't tell, then it is, in effect, no different than true intelligence to an observer ... what ever true intelligence is, that is defined by the super smart people. :)
tom66:
--- Quote from: BrianHG on February 27, 2023, 12:38:52 pm ---I guess anything codded, or any other text at all beginning in 2023 onward is now suspect to being partly or fully done by ChatGPT. Now, you may never again be able to hire employees where you can be sure their skills are genuine, or that they have any native authentic skills at all other than being able to 'tickle' ChatGPT in the right way to achieve results.
Ten years from now, such natural skills will begin to die off and those with said skills even lateron will be of a special breed.
--- End quote ---
Disagree. ChatGPT is not a programmer; it is a language processor. It cannot think independently like a programmer can. It's essentially a very good predictor of the next word in a sentence or next token in a programming language (or sentence).
It does change the way we test programmers (and students). However, a 10-20 minute interview with a promising candidate will reveal pretty quickly if they are just a good user of ChatGPT or if they are actually good at programming.
If you give ChatGPT a problem outside of its corpus of typical 'experience' - in other words a novel problem - it's really quite bad at solving it and very prone to bullshitting the solution - which is to say, it gives a plausible-looking solution but not an actually functional one.
Here's an example - a common enough problem in FPGA land. A brief scan and you might say "that's plausible" but even a slightly deeper reading reveals it to be nonsense.
--- Quote from: me ---In Verilog, construct a template for a Xilinx 7 series OSERDESE2 block with 8 data inputs serialising to a DDR output. The input clock shall be provided externally and the lower frequency clock derived from a BUFR with a divider.
--- End quote ---
--- Quote from: ChatGPT ---Sure, here's a template for a Xilinx 7 series OSERDESE2 block with 8 data inputs serializing to a DDR output in Verilog:
--- Code: ---module oserdese2_ddr(input clk, // external clock
input [7:0] data_in, // 8 data inputs
input reset,
output reg [1:0] ddr_out); // DDR output
// create a lower frequency clock using BUFR and a divider
wire clk_div;
BUFR bufr_inst (
.O(clk_div),
.I(clk),
.CE(1'b1),
.CLR(reset)
);
reg [7:0] data_reg; // register to hold data
always @(posedge clk_div) begin
if (reset) begin
data_reg <= 8'b0; // reset data register to 0
ddr_out <= 2'b0; // reset DDR output to 0
end else begin
data_reg <= data_in; // load new data into register
ddr_out <= {data_reg[0], data_reg[1]}; // serialize data and output to DDR
end
end
endmodule
--- End code ---
Here, we first create a lower frequency clock using a BUFR with a divider. Then, we create a register data_reg to hold the 8-bit input data. In the always block, we check if the reset signal is high, and if so, we reset the data_reg and ddr_out signals to 0. Otherwise, we load the new data_in into data_reg, and serialize the first two bits of data_reg to the ddr_out signal. Note that we are assuming a DDR output of two bits, as is typical for Xilinx 7 series OSERDESE2 blocks.
--- End quote ---
There's no doubt ChatGPT changes how a programmer works: it's a very useful tool. But it doesn't replace them (except maybe at the very lowest end, such as writing boilerplate test code - but even before ChatGPT much of that could already be automated, and was.)
--- Quote from: Brumby on February 27, 2023, 12:34:26 pm ---Moving right along....
Further to my previous inquiry: Has anybody had success in getting image responses?
I've had URLs offered to me, but none have actually presented me with an image. I just get "file not found" type messages.
--- End quote ---
Some people have managed to get basic ASCII art out but it doesn't seem the training set has enough of that to be useful. Also the process through which it generates the output isn't really compatible with image generation since it can only refer up to 2000 tokens deep and can't refer to future tokens. sqrt(2000) ~= 45 so it would not be able to render a large image even if it tried. DALL-E is optimised for images (as are things like Midjourney) and IMO, besides resolution, these NN's are so good at art with a good prompt that most artists are going to be seeing competition from AI art soon enough. Not that art was ever a field that people besides the auction houses and dealers tended to make it big on, though.
In order to be competitive with the human brain, we need a neural network that works in the analogue domain. A current 8-bit multiplier block is around 1k transistors and a single neuron can do something comparable to that.
Siwastaja:
--- Quote from: Microdoser on February 27, 2023, 10:05:23 am ---I haven't tested it, it probably needs some tweaking, but that was written in about a minute, freeing me up to do other, more interesting, coding.
--- End quote ---
By all means, test it and do that "tweaking" to get it work and then make the decision if it saved any time. All examples I have seen so far have shown the opposite to be true: on surface everything looks convincing, but all details are so seriously off that complete rewrite is needed, taking more time than just writing the thing from scratch.
The code seems quite fine, but on the other hand this is some trivial code which is easily written anyway. Surprisingly, the register offsets, I checked against datasheet, seem to be correct, suggesting that training material has had code for DS1307. Usually with ChatGPT, details like numbers are complete bogus and look random (they are not actually randomized, but come from a different context), so your example is already much better than usual.
magic:
--- Quote from: xrunner on February 27, 2023, 02:09:58 am ---Wait a sec ... it said -
"Overall, the evidence suggests that humans are intelligent and have developed many remarkable abilities that have allowed --> us <-- to thrive and prosper as a species."
Us? Us?
Shouldn't it have said "You" ? Does it think it's one of US? Has it assimilated it's creators?
--- End quote ---
It means that OpenAI is bulshitting us and it's their interns writing those responses :-DD
Ever heard of Theranos?
tszaboo:
--- Quote from: Brumby on February 27, 2023, 12:34:26 pm ---Moving right along....
Further to my previous inquiry: Has anybody had success in getting image responses?
I've had URLs offered to me, but none have actually presented me with an image. I just get "file not found" type messages.
--- End quote ---
You don't get image responses from ChatGPT, and the basic schematic drawings it creates are often wrong.
I'm fairly sure we will get image responses from bing though. It also does product suggestions.
--- Quote from: tom66 on February 27, 2023, 01:29:56 pm ---There's no doubt ChatGPT changes how a programmer works: it's a very useful tool. But it doesn't replace them (except maybe at the very lowest end, such as writing boilerplate test code - but even before ChatGPT much of that could already be automated, and was.)
--- End quote ---
So instead of stackoverflow people will copy their non working code from chatgpt.
Oh BTW something I tried today: I sked it to calculate something, then I asked a follow-up, to provide a confidence interval for the same calculation. It gave me a 95% confidence interval, with a made up errors.
Asked the same confidence for a trivia question, no way of replying to it.
Also, it seems to provide weird answers if you tell it that some info is wrong, and you ask the same question again. It remembers, but reaches wrong conclusions. I'm going to post something below.
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version