EVAL

EVAL
Yi˘
gi Kola ∗
Sea le, Washing on, U.S.A.
[email p o ec ed]
Abs ac
This a icle in oduces EVAL, a musical wo k ha explo es wo con empo a y
possibili ies wi hin a in age en i onmen — namely he Commodo e 64 and i s
MOS 6581/8580 SID sound chip. One o hese possibili ies is AI in e ocompu ing:
a simpli ied decode , adap ed o he Commodo e 64 BASIC en i onmen , gene a es
sho melodic lines de i ed om a p e- ained a ia ional au oencode . The o he
in ol es eal- ime sound p ocessing and imp o isa ion using a SID emula o . The
pe o mance, con olled li e ia wo Nin endo Swi ch JoyCons, begins as a play ul
in e ac ion be ween human and machine and akes an unexpec ed u n.
1 In oduc ion
1.1 Mo i a ion
The p ojec explo es he Commodo e 64 (C64) and i s MOS 6581/8580 SID (Sound In e ace De ice)
sound chip wi hin a con empo a y compu a ional, concep ual, and a is ic amewo k. I ollows
se e al objec i es: in eg a ing gene a i e AI in audio/music in o a in age compu ing en i onmen ,
using he SID chip in he con ex o eal- ime sound p ocessing, and b inging hese wo aspec s
oge he wi hin he amewo k o a musical composi ion/pe o mance.
While a ious ools exis o musicians o expe imen wi h he SID’s unique soundscape, a p o-
g amma ic, eal- ime, and modula app oach has been lacking. This gap is add essed wi h
SIDOsc
,
a cus om Supe Collide Uni Gene a o (UGen). The gene a i e AI componen o he p ojec , as
discussed in he ollowing sec ions, eme ges om bo h echnical and a is ic limi a ions. Echoing he
e hos o he C64 demoscene, he p ojec explo es how he unc ionali y o a a ia ional au oencode
can be app oxima ed wi hin he cons ain s o he Commodo e 64.
2 Backg ound
2.1 Re ocompu ing, SID, and AI
Re ocompu ing is a h i ing subcul u e ha a ac s p og amme s in e es ed in he challenge o
c ea i e p og amming wi hin limi ed esou ces. In oduced in 1982 and discon inued in 1994, he
Commodo e 64 has p o ided a e sa ile en i onmen o he p og amme s. The sys em’s MOS
6581/8580 SID sound chip o e s a cha ac e is ic soundscape wi h i s ou wa e o ms, h ee one
gene a o s, and h ee en elope gene a o s, among o he componen s.
The appea ance o c oss pla o m emula o s such as VICE enabled musicians and p og amme s o
explo e he C64 wi hou o iginal ha dwa e. VICE no ably ea u es eSID, a obus SID emula o
de eloped by Dag Lem and a ailable as open-sou ce so wa e (Lem, 2023).
The e is an exis ing body o li e a u e on a i icial in elligence wi hin he con ex o 1980s pe sonal
compu ing, wi h a pa icula ocus on he C64. Examples include ule-based sys ems (B ain and
B ain, 1984) as well as expe imen s wi h a i icial neu al ne wo ks (Walke , 1987). Recen ly,
∗Independen A is /Resea che
P oceedings o he 6 h Con e ence on AI Music C ea i i y (AIMC 2025),
B ussels, Belgium, Sep embe 10 h-12 h
a i icial in elligence pa adigms ha e e-eme ged in e ocompu ing, ypically in ol ing compu a ions
pe o med ex e nally due o ha dwa e cons ain s (Bild, 2024).
2.2 Rela ed wo ks
Nick Bild’s Commodo e 64 AI Image Gene a o signi ican ly in luenced his p ojec (Bild, 2024).
Bild uses a p obabilis ic p incipal componen analysis algo i hm o ob ain a se o weigh s, mean
alues, and a co a iance alue o scale he posi ions o sp i es on he sc een. The PPCA is ained
wi h 94 hand-made sp i es.
Alexande Sashno ’s AY-8912 emula ion lib a y and i s AY UGen o Supe Collide p o ided
addi ional inspi a ion. The lib a y emula es he Gene al Ins umen AY-3-8910 soundchip used in
in age sys ems like ZX-Spec um 128 and A a i ST.
3 Abou he wo k
3.1 Sys em desc ip ion
3.1.1 Usage o AI
The wo k ea u es "machine" and "human" in e ac ions wi h he SID chip. The machine in e ac ion
u ilizes a a ia ional au oencode (VAE) ained on 94 a ays o 64 bina y alues, each ep esen ing a
monophonic melody o ma ed as an 8×8 g id, whe e he x-axis co esponds o ime and he y-axis o
pi ch. A e aining he VAE, decode weigh s consis ing o an 8x64 ma ix and a 64-elemen bias
a e s o ed. All alues a e ounded o he nea es en h.
Figu e 1: A 8x8 ep esen a ion o a melody aken om he aining se .
Pos - aining, a simpli ied decode implemen ed in C64 BASIC gene a es melodic "p o o ypes" om
s o ed weigh s and biases. We ind his me hod o be an e ec i e comp omise, si ua ed be ween he
ambi ious ask o aining a a ia ional au oencode di ec ly wi hin a BASIC en i onmen and he
simple al e na i e o embedding a s a ic, p e-compu ed p o o ype as a BASIC a ay.
The decoding u ilizes a ixed la en ec o
z
which con ains 8 alues.The p o o ype is calcula ed ia
x=σ(W·z+b)
, whe e
W
and
b
a e he weigh ma ix and bias a ay, and
σ
is he sigmoid ac i a ion
unc ion which maps each o he 64 elemen s in o
[0,1]
. Fo C64 BASIC, we use an accommoda ing
piecewise-linea app oxima ion o he sigmoid unc ion:
2
s(x) = 


0,i x≤ −2
x+2
4,i −2<x<2
1,i x≥2
Figu e 2: Piecewise-linea app oxima ion o he sigmoid unc ion and i s implemen a ion in C64
BASIC
The p o o ype is hen linea ly mapped in o an a bi a y equency ange by applying a piecewise
ans o ma ion: o each o he 8 columns we ex ac he ow index
n ∈ {0, ..., 7}
o he "1 bi ",
no malize i o n
7, scale and o se o he in ended SID egis e span ia
F=jn
7×80 + 160k,
and inally w i e (o POKE) he 16- alue Fin o SID’s equency egis e s.
3.1.2 Real- ime usage o SID
The "human in e ac ion" aspec employs
SIDOsc
, a Supe Collide UGen de eloped by he au ho .
Essen ially a w appe a ound Dag Lem’s eSID emula o , he UGen is capable o p o iding eal- ime
oscilla o beha io ia he emula ion o SID’s ou oscilla o s – iangle wa e, saw oo h wa e, pulse,
and noise. Fo he li e p ocessing, he UGen is en iched h ough nonlinea sample eedback using
Fb1
, a UGen dedica ed o single sample eedback and eed o wa d ope a ions. This en i onmen
p o ides a ich imb al soundscape, which is hen explo ed by a numbe o con ol pa ame e s linked
o wo Nin endo Swi ch JoyCon con olle s ia OSC messaging o e Blue oo h ansmission ia
joycon-osc lib a y.
Two JoyCons con ol wo independen s eams o e en s. The le JoyCon adjus s eedback ia s ick
mo emen s, whe e e ical and ho izon al mo ions o he s ick a e mapped o he pi ch and ampli ude
alues. Simila o he condi ions o a no-inpu mixe eedback, he ampli ude a gumen does no
a ec he o e all ampli ude, bu d as ically changes he imb al p o ile o he sound. The o e all
ampli ude his pa icula s eam is manipula ed ia he accele ome e s o he le con olle . The
igh JoyCon manages addi i e syn hesis ac oss ou SID oscilla o s, modula ing hei independen
ampli udes h ough accele ome ic da a.
Figu e 3: Two JoyCons and hei unc ions.
3.2 Na a i e
EVAL begins as a play ul in e ac ion: he AI p esen s a monophonic melody, u ging he pe o me ’s
imp o isa ional esponse. The pe o me has comple e eedom in hei decisions. Howe e , each
3
a emp is me wi h ailu e. Ini ially, his is communica ed in a suppo i e one, bu he p omp s
g adually escala e in se e i y. Ul ima ely, he scena io is e ealed o be a igged a is ic e alua ion
wi hin a ic ional au ho i a ian go e nmen , o ches a ed by a commi ee o e seeing cul u al a ai s.
The isual componen o he wo k, w i en in 6502 assembly using KickAssemble u ilizes e ec s
such as as e ba s ha a e cha ac e is ic o he Commodo e 64 aes he ic (see Figu e 4).
Figu e 4: A sc eensho om he isual componen o he wo k.
4 Fu u e wo k
Cu en ly, no eal- ime e alua i e eedback loop exis s be ween he AI and pe o me , which is
consis en wi h he piece’s na a i e whe e an e alua ion wi h a p ede e mina ed conclusion is
being conduc ed. Fu u e i e a ions aim o inco po a e genuine in e ac ion and e alua i e p ocesses,
explo ing deepe human-machine dynamics. Addi ionally, implemen ing he ixed-media aspec s on
ac ual Commodo e 64 ha dwa e emains a compelling di ec ion.
Re e ences
Bild, N. (2024). Commodo e 64 AI image gene a o . A ailable a
h ps://gi hub.com/
nickbild/c64_gen_ai. Las accessed 13 Ap il 2025.
B ain, K. and B ain, S. (1984). A i icial In elligence on he Commodo e 64. Sunshine Books.
Lem, D. (2023). eSID: MOS 6581 / 8580 SID so wa e emula ion. A ailable a
h ps://gi hub.
com/daglem/ eSID. Las accessed 13 Ap il 2025.
Walke , J. (1987). Neu on ne wo k associa i e memo y p og am. A ailable a
h ps://www.
ou milab.ch/documen s/commodo e/B ainSim/. Las accessed 13 Ap il 2025.
4