Subject: Re: Can anyone get Edge Copilot to work without logging in after July 2026?
On Sun, 8/9/2026 1:41 PM, Andy Burns wrote:
Maria Sophia wrote:
Moving forward, given Gemini limits free chats as does ChatGPT, I'm going
I presume access to the AI chatbots is also blocked (without login) from the Firefox sidebar?
<https://support.mozilla.org/en-US/kb/ai-chatbot>
I have all AI (apart from website language translation disabled).
Well, you have to be able to take a joke.
I've tried out three models so far. Not really all
that adventurous at running local models (on inadequate hardware).
16GB of RAM would handle the 7GB or so Gemma 4 E4B needs (a bit weak model).
64GB of RAM would handle the 35GB or so Gemma 4 31B needed.
128GB of RAM would handle the 80GB the OpenAi 120B one needed.
Two are loaded in a Windows copy of LMStudio.
One is loaded in a Linux copy of LMStudio.
The tale starts with the following query.
Write a program in the C language that computes the mathematical constant PI to one million digits
I started out, not really caring about the code :-)
The mission here, was to see if an LLM-AI could "actually finish a mission". Can it work its little pencil, until the pencil is just a stub.
The longest question runtime, was 50 minutes. Another of the runs
took 23 minutes.
1) With E4B, all it did was write a file full of "comment text"
/* Comment text about how I would write your program */
Utterly... useless. Like tits on a chest freezer.
2) So I tried 31B. First I had it attempt the feat in the C language.
It burbled along for a bit, talking to itself. It always
announces "what the BEST algorithm is". Then it just happens
to mention one of the older or inferior algorithms and it writes
code for that! Then in the comments, it happens to mention
"If I implemented the Binary Splitting Algorithm, that would
be 200 lines of code." That's its excuse for why the homework
assignment is late. A virtual dog ate the homework.
3) Now, knowing from previous work I've done, that the C++ version
of Bignum code is half the length of the C version (and runs at half
the speed, so this is not a free lunch), I "cut the AI a break" by
reducing that BSA-toughness issue, to a potential 100 lines of code.
Write a program in the C++ language that computes
the mathematical constant PI to one million digits
This time it writes Chudnovsky, without BSA. And this time, it
volunteers the comment about BSA, that "it would take five hundred
lines of C++ to implement BSA and I could make an off-by-one error
while writing it". So now I know, it is pulling its complexity
estimates from the AI nether region (the hole where the slop comes out) :-)
OK, so now I go online, and ask for a Chudnovsky with BSA. It's
almost the first hit in the Google search. It is C++ .
(
https://gist.github.com/komasaru/68f209118edbac0700da )
/*
* Compute PQT (by Binary Splitting Algorithm)
*/
PQT Chudnovsky::compPQT(int n1, int n2)
{
int m;
PQT res;
if (n1 + 1 == n2) {
res.P = (2 * n2 - 1);
res.P *= (6 * n2 - 1);
res.P *= (6 * n2 - 5);
res.Q = C3_24 * n2 * n2 * n2;
res.T = (A + B * n2) * res.P;
if ((n2 & 1) == 1) res.T = - res.T;
} else {
m = (n1 + n2) / 2;
PQT res1 = compPQT(n1, m);
PQT res2 = compPQT(m, n2);
res.P = res1.P * res2.P;
res.Q = res1.Q * res2.Q;
res.T = res1.T * res2.Q + res1.P * res2.T;
}
return res;
}
Notice that this is 23 lines of code :-) Nether regions, my ass.
Now, is that code "scary" or what ? Does your little AI heart
tremble at the thought of coughing up a hairball like that ?
The code is recursive and calls itself, but that does not
normally cause shyness in AI.
So let's run that puppy. I ran it for two million digits,
because the above code is missing the "+ 64 bits" thing
you're supposed to add for the guard digits, which helps ensure
the last digits in your millions of digits, are correct.
$ ./chudnovskyBSA
**** PI Computation ( 2000000 digits )
TIME (COMPUTE): 0.664249 seconds.
TIME (WRITE) : 0.172128 seconds.
That code runs *ten times* faster than SuperPI 1.5 XS .
You could re-write the code in C and shave more time off it.
I don't think humans are obsolete, quite yet. Tip of the
hat to komasaru for the code.
Paul
--- PyGate Linux v1.5.19
* Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)