Category Archives: AI

Google Street View scripts and A.I. emojis

Leave a reply

When capturing video for Google Street View, Google recommends using 1 frame per second video for walking, and 5 frames per second for biking and lower speeds. A full 30 or even 60 fps video is unnecessarily huge and will take much longer to upload and process … and most of the frames will be discarded by Google anyway.

I had one of the A.I.s (probably CoPilot) automate using the ffmpeg open source command line tool so I could batch convert files in a directory. A very rough work-in-progress version is on my GitHub now:

allenhuffman/GoogleStreetViewScripts: Scripts for converting videos before uploading to Google Street View

I have noticed the A.I.s are starting to put emojis in things — including code and scripts they generate!

I don’t even know how to type emojis in uMacs or VI ;-) but apparently they are supported these days.

Have you noticed the increase in emojis in A.I. responses lately?

I’d end this post with an emoji, but I do not know how to type one in WordPress . . .

Could the CoCo run a tiny Large Language Model (AI) thing?

2 Replies

An interesting post appeared on the Color Computer Facebook group the other day… Youngstown Ken posted:

AI on the CoCo? This is a simple language experiment that I think can be done on the coco if anyone would like to try the code. I asked Grok to write a basic program to read a 10MB corpus of sentences from a dictionary and then build a simple Markov graph db to predict the next word in a sequence.

– Youngstown Ken, via Facebook CoCo group

The attached link went to a Grok-generated conversation about the subject. You can read the short interaction here, and see the generated Python and BASIC-ish code:

https://x.com/i/grok/share/sXRGgHcIROsl7PPNXnF8D34MC?fbclid=IwY2xjawI15sRleHRuA2FlbQIxMQABHZVbpT6JU-bIDIQJoVd4KxBf4gJFb2oSJmYX9t9Y3sIShpw0nK6nmP8laA_aem_Q4MMZzgAmLsauqsrU5ID6g

So naturally, someone needs to try this and see if it would even work. AI is terrible at BASIC, though I did have some success talking it through some simple 32-column text screen game code when ChatGPT first made a big public splash a few years ago.

Let’s look at the second code variation it produced:

10 REM Two-Level Markov Chain Predictor for TRS-80 CoCo
20 REM Vocabulary: 1000 words, Trigram model fits in 512KB

30 DIM V$(999): REM Vocabulary array (1000 words)
40 DIM F(999): REM Frequency array for vocab building
50 DIM M(499,3): REM Model: 500 trigrams max (w1,w2,w3,prob)
60 N=0: REM Number of trigrams loaded

70 REM Pass 1: Build vocabulary from corpus.txt
80 OPEN "I",#1,"CORPUS.TXT"
90 FOR I=0 TO 999: F(I)=0: NEXT I: REM Initialize frequency
100 IF EOF(1) THEN 140
110 LINE INPUT #1,L$: REM Read line
120 GOSUB 500: REM Process line for word frequencies
130 GOTO 100
140 CLOSE #1

150 REM Sort top 1000 words by frequency (simplified)
160 FOR I=0 TO 999: V$(I)="<UNK>": NEXT I: REM Default to <UNK>
170 W$="":C=0: REM Temp word and count
180 OPEN "I",#1,"CORPUS.TXT" : REM Reopen for word assignment
190 IF EOF(1) THEN 230
200 LINE INPUT #1,L$
210 GOSUB 600: REM Assign top 1000 words
220 GOTO 190
230 CLOSE #1

240 REM Pass 2: Build trigram model
250 OPEN "I",#1,"CORPUS.TXT"
260 IF EOF(1) THEN 300
270 LINE INPUT #1,L$
280 GOSUB 700: REM Build trigrams
290 GOTO 260
300 CLOSE #1

310 REM Input sequence and predict
320 PRINT "Enter two-word sequence (e.g., 'the quick'):"
330 INPUT S$
340 GOSUB 800: REM Extract last two words and predict
350 PRINT "Next word:";W3$
360 GOTO 320: REM Loop for more predictions

500 REM Subroutine: Count word frequencies
510 W$="":P=1
520 IF P>LEN(L$) THEN RETURN
530 C$=MID$(L$,P,1)
540 IF C$=" " OR P=LEN(L$) THEN GOSUB 550: GOTO 570
550 IF W$<>"" THEN FOR I=0 TO 999: IF V$(I)=W$ THEN F(I)=F(I)+1: GOTO 560: NEXT I
560 W$="":P=P+1:GOTO 520
570 W$=W$+C$:P=P+1:GOTO 520

600 REM Subroutine: Assign top 1000 words (simplified)
610 W$="":P=1
620 IF P>LEN(L$) THEN RETURN
630 C$=MID$(L$,P,1)
640 IF C$=" " OR P=LEN(L$) THEN GOSUB 650: GOTO 670
650 IF W$<>"" THEN FOR I=0 TO 999: IF V$(I)="" AND F(I)>C THEN V$(I)=W$:C=F(I): GOTO 660: NEXT I
660 W$="":P=P+1:GOTO 620
670 W$=W$+C$:P=P+1:GOTO 620

700 REM Subroutine: Build trigrams
710 W1$="":W2$="":W3$="":P=1:W=0
720 IF P>LEN(L$) THEN RETURN
730 C$=MID$(L$,P,1)
740 IF C$=" " OR P=LEN(L$) THEN GOSUB 750: GOTO 770
750 IF W$<>"" THEN W=W+1:IF W=1 THEN W1$=W$ ELSE IF W=2 THEN W2$=W$ ELSE IF W=3 THEN W3$=W$:GOSUB 900
760 W$="":P=P+1:GOTO 720
770 W$=W$+C$:P=P+1:GOTO 720

900 REM Subroutine: Add trigram to model
910 IF N>=500 THEN RETURN: REM Limit to 500 trigrams
920 FOR I=0 TO 999: IF V$(I)=W1$ THEN W1=I:GOTO 930: NEXT I:W1=999
930 FOR I=0 TO 999: IF V$(I)=W2$ THEN W2=I:GOTO 940: NEXT I:W2=999
940 FOR I=0 TO 999: IF V$(I)=W3$ THEN W3=I:GOTO 950: NEXT I:W3=999
950 FOR I=0 TO N-1: IF M(I,0)=W1 AND M(I,1)=W2 THEN M(I,3)=M(I,3)+1:RETURN
960 M(N,0)=W1:M(N,1)=W2:M(N,2)=W3:M(N,3)=1:N=N+1
970 RETURN

800 REM Subroutine: Predict next word
810 W1$="":W2$="":P=1
820 IF P>LEN(S$) THEN GOTO 850
830 C$=MID$(S$,P,1)
840 IF C$=" " THEN W1$=W2$:W2$="":P=P+1:GOTO 820 ELSE W2$=W2$+C$:P=P+1:GOTO 820
850 IF W1$="" THEN W1$=W2$:W2$="": REM Handle single word
860 FOR I=0 TO 999: IF V$(I)=W1$ THEN W1=I:GOTO 870: NEXT I:W1=999
870 FOR I=0 TO 999: IF V$(I)=W2$ THEN W2=I:GOTO 880: NEXT I:W2=999
880 W3$="<UNK>":P=0
890 FOR I=0 TO N-1: IF M(I,0)=W1 AND M(I,1)=W2 AND M(I,3)>P THEN W3$=V$(M(I,2)):P=M(I,3)
900 NEXT I
910 RETURN

This is surprising to me, since it does appear to be honoring the 2-character limit of Color BASIC variables, which is something my early interactions with ChatGPT would not do.

The original prompt told it to do something that could run on a CoCo within 512K. Even Super Extended Color BASIC on a 512K CoCo 3 still only gives you 24K of program space on startup. And, as I recently learned, if you try to use a CoCo 3’s 40 or 80 column screen and clear more than about 16K of string space, you get a crash.

I think I may give this a go. I do not understand how any of this works, so I do not expect good results, BUT I do want to at least try to see if I can make the program functional. If it expects to use 512K of RAM for string storage, this program is doomed to fail. But, 24K would be room for 1000 24-byte words, at least.

To be continued … maybe.

In the meantime, check out that Grok link and see what you can come up with. And let me know in the comments.

AI Music: SUNO versus UDIO

1 Reply

About a week ago, YouTube decided to show me this video:

While I was aware that AI-generated music was a thing, I had no idea it had gotten this good.

I signed up and spent the next few days playing with SUNO. It produced some pretty amazing and surprising results.

Then, a few days ago, I was shown a video speculating on some new AI music generator that was about to be announced. The claim was it would be 2X better than SUNO.

Yesterday, that service was launched — UDIO (like studio with the ST I guess?). It was down most of the day (overloaded) but late last night I got to play with it, mostly letting it produce random lyrics and tunes

Here is an UDIO generated song about being in love with a VCR:

https://www.udio.com/songs/1t4BXqC7FFw9zGZXic5UCf

Udio looks quite promising, with some features missing from SUNO (like the ability to extend a song and add to the START; SUNO allows you to extend from a time code, only).

Initial testing did not let me make any “complete song” like I have been doing with SUNO, but I expect this must be possible with enough attempts – their gallery is full of them.

Vocal quality of Udio is stunning. SUNO sounds great most of the time, but there is an auto-tune/harmonizer digital sound on vocals sometime. I have not heard that yet on any of the dozens of things I have had Udio create.

The music sounds good, but SUNO seems to generate more “wow, I like this song” tracks than Udio. The three “songs” I had SUNO make are just kind of drifting melodies, with no clear chorus or hook or anything. I have certainly heard songs like that, but right now I’d give SUNO the advantage of making “actual songs” versus tunes.

Check them both out and see what you think.

SUNO AI creates CoCo songs…

Leave a reply

The state of A.I. music generator is far beyond where I expected it to be. In fact, I did not even think this would ever be possible. But here it is.

I wrote up some lyrics and worked with SUNO until it got something interesting. Here is the 80s rock version:

I wrote the words to this, but SUNO is able to create lyrics as well. If you are not a writer, you can just say “country song about a boy and his dog named Fred, down by the lake on a summer evening” and away it goes.

Speaking of country, I used the same lyrics and did a shorter version as a country song…

The styles of music this thing can make is pretty open – rock, country, doo wop, barbershop quartets, opera, death metal, blues, jazz, etc. Just for fun, here’s an 80s new wave kind of synthesizer pop song:

Music production had already gotten very simple thanks to loops you just drag in to build songs, so the real impressive part of SUNO is that it can create the singing.

Check out https://www.suno.ai and have your mind blown, too. If there is any interest, I will do a video showing what I have learned about using this tool.

Until then…