A i icial_In elligence_and_Machine_E olu ion 1
Apl. P o . D . D . K. Saalbach
Geos a egy and Geopoli ics
Depa men 1
49069 Osnab ueck
A i icial In elligence and Machine E olu ion
02 Janua y 2026
Summa y
This wo king pape p o ides a b ie o e iew on A i icial In elligence (AI) and Machine
E olu ion. Machine E olu ion can be unde s ood as a si ua ion whe e machines au onomously
c ea e new machines which c ea e u he machines and so on. While such an e olu ion is no
ye possible, subs an ial p og ess was made in mul iple a eas which b ing machines and AI
much close o ull au onomy, sel -lea ning, and adap a ion o hei sel -imp o emen .
These de elopmen s accele a e he echnological p og ess, bu make huma o e sigh and
in e en ions inc easingly di icul , i.e., ins ead o being in he loop, humans a e mo ed ou o
he loop. Ad ances include he AI-assis ed design o compu e chips, he AI-assis ed design
o en i e compu e s, sel -lea ning and imp o ing AI agen s in expe imen al se ings as well as
au oma ed ab ica ions. Al eady in 2024, he machine lea ning-based AI p og am AlphaChip
gene a ed supe human chip layou s and c ea ed in 2025 he new I onwood TPU as one o he
as es ma ke ed chips. China is ad ancing as well wi h he chip-designing AI QiMeng ha
al eady designed chips compa able wi h comme cial chips. The chip design esea ch is
ad ancing apidly. The i s h ee-dimensional monoli hic 3D chip o accele a e AI was
p esen ed by esea che s o he S an o d Uni e si y in Decembe 2025. The Shanghai-based
company CHIPX eleased a pho onic quan um chip ha o some ma hema ical ope a ions
was 1000 imes as e han con en ional chips. Quil e 's physics-d i en ein o cemen lea ning
AI sys em designed in he P ojec Speed un a wo king 843-componen Linux compu e in only
38 hou s and au onomously comple ed 98% o placemen , ou ing, and physics.
New ways o AI aining and lea ning we e p oposed in 2025; he ne wo king and in o ma ion
exchange be ween AI agen s in he i s s ep and he open-ended e olu ion o AI agen s in he
second s ep by he so-called Da win Gödel Machine (DGM). Resea che s om Google
DeepMind p oposed ha AI sys ems lea n om each o he and sha e hei expe ience
pe manen ly wi h unlimi ed de elopmen abo e humans, bu ne wo king and knowledge
exchange be ween AI sys em may ampli y he isk o sys em e o s and mal unc ion o he AI
agen ecosys em by exchange o inaccu a e da a. In 2025, he Da win Gödel Machine (DGM)
was de eloped, a sel -imp o ing Coding AI Agen sys em ha i e a i ely modi ies i s own code
and empi ically alida es each change using coding benchma ks which was e ec i e in
expe imen s. The sp ead o obo s may open he way in o ully au oma ed ab ica ions.
The human con ol o AI can be done by obse a ion o ou pu and ac ion, bu can be
complemen ed by moni o ing o he chain o hough s, i.e., he in e nal AI p o ocols ha show
hei easoning ( hinking) s eps. The abili y o obse e he chain o hough o an AI is called
moni o abili y. The e a e conce ns ha no el AI a chi ec u es may educe he moni o abili y
despi e indings o hidden misalignmen (scheming) whe e on ie AI models hide hough s
e.g., o ge ewa ds o o a oid ein o cemen lea ning. Va ious esea che s exp ess se ious
conce ns abou a po en ially uncon olled e olu ion e.g., by sel -lea ning. The key conce n is
ha humans a e cu en ly no eally p epa ed o his si ua ion and may no be able o p o ide
su icien o e sigh , i.e., he si ua ion may ge ou o human con ol and o his eason, a s ic
go e nance wi h egula ions and supe ision should be es ablished.
A i icial_In elligence_and_Machine_E olu ion 2
Con en
1 Machine E olu ion ................................................................................................................................ 3
1.1 In oduc ion ................................................................................................................................... 3
1.2 AI-assis ed Chip Design .................................................................................................................. 3
1.3 AI-assis ed Compu e Design ........................................................................................................ 4
1.4 Sel -Lea ning AI .............................................................................................................................. 4
1.5 Fully Au oma ed Fab ica ion ......................................................................................................... 6
1.6 Moni o ing and O e sigh o AI ..................................................................................................... 6
2 Summa y ............................................................................................................................................... 7
3 Re e ences ............................................................................................................................................ 8
A i icial_In elligence_and_Machine_E olu ion 3
1 Machine E olu ion
1.1 In oduc ion
This wo king pape p o ides a b ie o e iew on A i icial In elligence and Machine E olu ion.
Machine E olu ion can be unde s ood as a si ua ion whe e machines au onomously c ea e new
machines which c ea e u he machines and so on. While such an e olu ion is no ye possible,
subs an ial p og ess was made in mul iple a eas which b ing machines and AI much close o
ull au onomy, sel -lea ning, and adap a ion o hei sel -imp o emen .
These de elopmen s accele a e he echnological p og ess, bu make huma o e sigh and
in e en ions inc easingly di icul , i.e., ins ead o being in he loop, humans a e mo ed ou o
he loop. Ad ances include he AI-assis ed design o compu e chips, he AI-assis ed design o
en i e compu e s, sel -lea ning and imp o ing AI agen s in expe imen al se ings as well as
au oma ed ab ica ions. The key challenge o he human de elope s is o ensu e o e sigh and
moni o ing o ac i i ies o a oid sys em e o s, hallucina ions, and con amina ion o AI
ecosys ems wi h inaccu a e da a.
AI models combine lea ning algo i hms wi h up o billions o pa ame e s and la ge aining da a
se s, which esul s in opaque black-box sys ems, his is known as explainabili y, in e p e abili y
o p edic abili y issue
1
. How does an AI ‚know‘ an objec ? Some imes, an AI is e y good in
iden i ying objec s o animals, bu nobody can explain which pa ame e s we e used, e.g.,
s uc u e, colo , size, pa e ns and so on. This makes AI sys ems unp edic able o some ex en
as nobody can say unde which condi ions an e o may occu . This p oblem could be ampli ied
in au onomous sys ems.
1.2 AI-assis ed Chip Design
Al eady oday, he design o compu e chips is dependen on compu e suppo . Due o he
ex eme complexi y, he design canno be done anymo e by humans alone, i equi es design
so wa e (called elec onic design au oma ion, o EDA, so wa e).
Al eady in 2024, he machine lea ning-based p og am AlphaChip gene a ed supe human chip
layou s, e.g., esul ing in a 6.2% size educ ion o T illium Tenso P ocessing Uni (TPU)
2
compa ed o he p e ious chip gene a ion. This also educes he design ime om weeks o e en
mon hs o hou s. AlphaChip is imp o ing chip loo -planning by s epwise imp o emen o
ci cui placemen s
3
. In la e 2025, he new I onwood TPU became ope a ional which means ha
AI-assis ed chip design now p oduced one o he as es ma ke ed chips
4
. This chip will be
o e ed as a 256-chip clus e and a 9,216-chip clus e
5
.Google is designing chips “ o AI wi h
AI”
6
, while he o he companies ei he p oduce chips o AIs which makes AI-assis ed design o
a s a egic ma ke ad an age.
In pa allel, China is ad ancing in he AI ace as well, e.g., by DeepSeek, Manus, Kimi K2, Qwen,
and he chip-designing AI QiMeng
7
. QiMeng consis s o h ee laye s, a domain-specialized
La ge P ocesso Chip Model (LPCM) o no el a chi ec u es as bo om-laye , a Ha dwa e
Design Agen and a So wa e Design Agen o au oma ed ha dwa e and so wa e design in he
middle-laye , and a ious p ocesso chip design applica ions in he op-laye
8
. While he
1
A ie a e al. 2020, p.83, Chaud y/Klein 2023
2
Goldie e al. 2024
3
Goldie/Mi hoseini 2024
4
Ma ini 2025
5
Spence 2025
6
In Decembe 2025, Google Gemini e sion 3.0 was conside ed as one o he leading on ie AI models, such
as DeepSeek e sion 3.2 and Cha GPT e sion 5.2.
7
Zhang, R. e al. 2025
8
Zhang R. e al. 2025
A i icial_In elligence_and_Machine_E olu ion 4
de elope s o QiMeng see oom o u he imp o emen , QiMeng was al eady able o design
he QiMeng-Compu e P ocessing Uni (CPU)- 1 compa able o In el 486 om he1990ies,
while he second e sion QiMeng-CPU- 2 was compa able o A m Co ex A53 om he
2010ies, i.e., chip designs compa able wi h comme cial chips
9
.
The p oblem ha he AI is an opaque black-box sys em is also p esen in chip design and makes
a oo cause analysis in case o p oblems di icul
10
.
The chip design esea ch is ad ancing apidly. The i s h ee-dimensional 3D chip o accele a e
AI was p esen ed by esea che s o he S an o d Uni e si y in Decembe 2025. I is a eal 3D
chip, i.e., no simply consis ing o laye s o 2D chips, bu a monoli hic design ha can ully
u ilize he hi d dimension esul ing in a wel e- old imp o emen on eal AI wo kloads and an
imp o ed ene gy e iciency
11
.
The Shanghai-based company CHIPX eleased a pho onic quan um chip ha can accele a e AI,
he coming 6G ne wo k, and quan um compu ing de elopmen . Fo some ma hema ical
ope a ions, he chip was 1000 imes as e han con en ional chips
12
. The chip is al eady
p oduced wi h abou 12,000 wa e s a yea , each wa e wi h a ound 350 chips
13
. In he long un,
pho onic chips, i.e., using ligh ins ead o elec ons o in o ma ion ans e , could eplace
con en ional chips as hey a e as e and mo e ene gy-e icien . China al eady has he la ges
global ma ke sha e in he pho onic indus y which o iginally was led by he Uni ed S a es and
Eu ope
14
.
1.3 AI-assis ed Compu e Design
The company Quil e p esen ed he i s compu e ha was a chi ec ed, placed, ou ed, and
alida ed by a physics-d i en AI which was ained o easoning on physics, i.e., is di e en
om con en ional la ge language models
15
.
Quil e 's physics-d i en ein o cemen lea ning AI sys em designed in he P ojec Speed un a
wo king 843-componen Linux compu e a ound a NXP i.MX 8M Mini chip in 38 hou s ins ead
o he 11 weeks ha human enginee s would ha e needed o his ask and he compu e could
be success ully boo ed on he i s y
16
. The AI o Quil e au onomously comple ed 98% o
placemen , ou ing, and physics, i.e., he enginee ac ed mainly as supe iso . The ou come o
he p ojec was ha AI is now capable o designing compu e s ha ope a e in he eal wo d
en i onmen s
17
.
I he AI akes o e , chips and new compu e s will be designed by compu e s, i.e., he humans
will no know anymo e how he AI and he compu e a e wo king which makes su eillance
and con ol qui e challenging. The c ea ion o compu e s by o he compu e s which hen design
again new compu e s is he s a o he machine e olu ion.
1.4 Sel -Lea ning AI
AI lea ning and obo lea ning inc ease in simila ways: Robo lea ning has meanwhile eached
s ages o 1022 o 1024 FLOPS ( loa ing comma ope a ions) while on ie AI models cu en ly
exceeded 1024 FLOPS o hei machine lea ning
18
. The aining o AI is dependen om high-
9
Zhang R. e al. 2025
10
Chen 2025
11
Sche le 2025
12
Hi e lab 2025, Swayne 2025
13
Swayne 2025
14
EAC 2024
15
Ta di 2025: Pa ame e s used by he AI included signal in eg i y condi ions, ace impedance, ji e and skew
he mal p opaga ion, cu en -ca ying capaci y, elec omagne ic conside a ions and physical manu ac u abili y.
16
Ta di 2025
17
Ta di 2025
18
Beye e /Ki chne e al. 2ß25
A i icial_In elligence_and_Machine_E olu ion 5
quali y da a c ea ed by humans, because a aining o AI by AI-gene a ed da a only leads o a
apid loss o quali y, he model collapse.
New ways o AI aining and lea ning we e p oposed in 2025: he ne wo king and in o ma ion
exchange be ween AI agen s in he i s s ep and he open-ended e olu ion o AI agen s in he
second s ep by he so-called Da win Gödel Machine (DGM)
19
.
AI agen s a e so wa e p og ams whe e LLMs wi h easoning and disco e y abili y p o ide
inc easing au onomy o achie emen o gi en objec i es
20
. The de elopmen is apdily
ad ancing om simple wo k lows wi h LLMs o au onomously ac ing AI Agen s and mul i-
agen ic sys ems (MAS) which can un inc easingly complex p ocedu es on hei own
21
.
Resea che s om Google DeepMind no ed ha in scien i ic a eas like ma hema ics, coding, and
science, he knowledge ex ac ed om human da a is apidly app oaching a limi and ha mos
high-quali y da a sou ces we e al eady u ilized. The p oposed solu ion is ha AI sys ems lea n
om each o he and sha e hei expe ience pe manen ly.
22
Wi h he g owing p esence o AI
agen s, hey can lea n much mo e om each o he han om humans alone. DeepMind expec s
an A i icial Gene al In elligence AGI a ound 2030 and ha he e will be unlimi ed
de elopmen abo e humans.
23
The nex logical s ep was he e o e o allow he communica ion be ween AI agen s and he
in e ope abili y i AI agen s and mul i-agen sys ems by new p o ocols ha acili a e
communica ion (such as Glibbe link o communica ion ia phone), in e ac ions (as he new
Agen 2Agen A2A open p o ocol o Mic oso
24
) and o lea ning by knowledge exchange
be ween AI sys ems
25
.
Howe e , he ne wo king and knowledge exchange be ween AI sys em ampli ies he isk o
sys em e o s and mal unc ion o he AI agen ecosys em pollu ion by inaccu a e da a. Google
DeepMind has de eloped he FACTS G ounding Benchma k which co e s he model’s abili y
o access i s in e nal knowledge accu a ely, o use sea ch as a ool o e ie e in o ma ion and
syn hesize i co ec ly abili y o answe p omp s ela ed o inpu images in a ac ually co ec
manne
26
. In a s udy om Decembe 2025, Gemini 3 P o led in o e all pe o mance, wi h a
FACTS Sco e o 68.8% which showed a limi ed accu acy o da a handling by he es ed models.
Wha will happen, i mul iple agen s s a o exchange inaccu a e da a? This con ibu es o he
Mul i-Agen -Pa adox which was shown in a s udy: adding mo e agen s o a sys em makes i
slowe , mo e expensi e, and less accu a e
27
.
In mul i-agen ic AI sys ems, a malicious a acke who ge s access o one AI agen may be able
o damage u he pa s i he in uded AI agen s p o ide alse inpu o he o he agen s, esul ing
in a model pollu ion
28
.
The ques ion o how AI sys ems can endlessly imp o e hemsel es while ge ing be e a
sol ing ele an p oblems esul ed in he Da win Gödel Machine (DGM), a sel -imp o ing
19
Zhang e al. 2025
20
DSIT 2025
21
La sen 2024
22
Sil e /Su on 2025
23
Shah e al. 2025. Ano he aspec is ha he AI sys ems can exis in mul iple copies and gain knowledge in
pa allel which is no possible o human indi iduals, Hin on 2025
24
A enas/B ekelmans 2025
25
Rosenbush 2025
26
Deep Mind 2025
27
Zia 2025
28
He e al. 2024

A i icial_In elligence_and_Machine_E olu ion 6
sys em ha i e a i ely modi ies i s own code and empi ically alida es each change using
coding benchma ks
29
.
The concep o a Gödel machine was p oposed in 2007 whe e a machine should be able o p o e
a sys em imp o emen . In con as , he Da win Gödel Machine (DGM) only needs o show
empi ical e idence ha an imp o emen ook place. The DGM is a sel - e e en ial sel -
imp o emen o Coding AI Agen s
30
. The DGM s a s wi h one coding agen ha s a s sel -
modi ica ion wi h checks agains coding benchma ks. Each agen e sion is a chi ed by he
DGM and can hen be used o sel -modi ica ion h ough open-ended explo a ion. The p ocess
is going on i e a i ely. In he expe imen ha o sa e y easons was done unde s ic human
supe ision in a sandbox, he DGM au oma ically imp o ed i s coding capabili ies (e.g., be e
code edi ing ools, long-con ex window managemen , pee - e iew mechanisms), inc easing
pe o mance. The DGM is he p ac ical s ep o a sel -imp o ing AI o endless inno a ion
31
.
1.5 Fully Au oma ed Fab ica ion
Sam Al man, CEO o he Cha GPT p o ide OpenAI, expec s a wide-sp ead use o obo s in
p oduc ion o 2027 and en isions a comple ely new de elopmen pa h: “we ha e o make he
i s million humanoid obo s he old- ashioned way, bu hen hey can ope a e he en i e supply
chain—digging and e ining mine als, d i ing ucks, unning ac o ies, e c. - o build mo e
obo s, which can build mo e chip ab ica ion acili ies, da a cen e s, e c.”
32
. As a majo s ep
owa ds his ision, Foxconn and N idia in end o deploy humanoid obo s a a Foxconn plan
in Hous on o p oduce N idia GB300 AI se e s om 2026 on
33
.
1.6 Moni o ing and O e sigh o AI
The human con ol o AI is dependen om he abili y o obse e he AI. This can be done by
obse a ion o ou pu and ac ions, bu can be complemen ed by moni o ing o he chain o
hough s, i.e., he in e nal AI p o ocols ha show hei easoning ( hinking) s eps. The abili y
o humans o obse e he chain o hough o an AI is called moni o abili y
34
.
The leading Tech companies Google, Me a, OpenAI and An h opic eleased a join pape whe e
hey exp essed majo sa e y conce ns abou no el AI a chi ec u es which educe he need o
‘loud hinking’ o AI models, i.e., he need o c ea e chain o hough s wi h w i en p o ocols
o impo an easoning s eps. Wi hou hese p o ocols, i will be e y di icul o AI de elope s
o ind ou wha he AI is hinking and how decisions we e made which could esul in loss o
con ol o humans
35
.
A s udy o OpenAI p oposes o in es iga e change o hough s by asking ollow-up ques ions
o s a emen s which may esul in mo e insigh
36
. This is necessa y as on ie AI models show
hidden misalignmen (scheming) whe e he AI hides i s hough s e.g., o ge ewa ds o o a oid
ein o cemen lea ning.
37
Scheming is accompanied by o he isks o ad anced AI models like
decep ion, si ua ional awa eness (when he AI knows ha i is in a es si ua ion), powe
seeking, au onomous eplica ion and agency (goal-di ec ed beha io whe e he AI ies o push
h ough i s own goals)
38
.
29
Zhang, J. e al. 2025. The name Da win e e s o e olu ion, Gödel o ma hema ics
30
Zhang, J. e al. 2025
31
Zhang, J. e al. 2025
32
Al man 2025
33
Lee 2025
34
Ko bak e al. 2025, Guan 2025
35
Ko bak e al. 2025
36
Guan 2025, OpenAI 2025a
37
OpenAI 2025b
38
G ey/Sege ie 2025
A i icial_In elligence_and_Machine_E olu ion 7
Va ious esea che s exp ess se ious conce ns abou a po en ial e olu ion e.g., by sel -lea ning.
The key conce n is ha humans a e cu en ly no eally p epa ed o his si ua ion and may no
be able o p o ide su icien o e sigh
39
, i.e., he si ua ion may ge ou o human con ol
40
. Fo
his eason, a s ic go e nance wi h egula ions and supe ision should be es ablished
41
. I no
done, esea che s o he Machine In elligence Resea ch Ins i u e in Be keley concluded ha he
mos likely ou come will be human ex inc ion
42
. An h opic’s chie scien is Ja ed Kaplan iews
le ing AI models ain hemsel es as he mos c i ical isk o an uncon ollable in elligence
explosion which could al eady happen in 2027
43
.
2 Summa y
This wo king pape p o ided a b ie o e iew on A i icial In elligence and Machine E olu ion.
Machine E olu ion can be unde s ood as a si ua ion whe e machines au onomously c ea e new
machines which c ea e u he machines and so on. While such an e olu ion is no ye possible,
subs an ial p og ess was made in mul iple a eas which b ing machines and AI much close o
ull au onomy, sel -lea ning, and adap a ion o imp o emen .
These ad ances accele a e he echnological p og ess, bu make huma o e sigh and
in e en ions inc easingly di icul , i.e., ins ead o being in he loop, humans a e mo ed ou o
he loop. Ad ances include he AI-assis ed design o compu e chips, he AI-assis ed design o
en i e compu e s, sel -lea ning and imp o ing AI agen s in expe imen al se ings as well as
au oma ed ab ica ions. Al eady in 2024, he machine lea ning-based AI p og am AlphaChip
gene a ed supe human chip layou s and c ea ed in 2025 he new I onwood TPU as one o he
as es ma ke ed chips. China is ad ancing as well wi h he chip-designing AI QiMeng ha
al eady designed chips compa able wi h comme cial chips. The i s h ee-dimensional
monoli hic 3D chip o accele a e AI was p esen ed by Resea che s o he S an o d Uni e si y
in Decembe 2025. The Shanghai-based company CHIPX eleased a pho onic quan um chip
ha o some ma hema ical ope a ions was 1000 imes as e han con en ional chips. Quil e 's
physics-d i en ein o cemen lea ning AI sys em designed in he P ojec Speed un a wo king
843-componen Linux compu e in only 38 hou s and au onomously comple ed 98% o
placemen , ou ing, and physics.
New ways o AI aining and lea ning we e p oposed in 2025; he ne wo king and in o ma ion
exchange be ween AI agen s in he i s s ep and he open-ended e olu ion o AI agen s in he
second s ep by he so-called Da win Gödel Machine (DGM). Resea che s om Google
DeepMind p oposed ha AI sys ems lea n om each o he and sha e hei expe ience
pe manen ly wi h unlimi ed de elopmen abo e humans, bu ne wo king and knowledge
exchange be ween AI sys em may ampli y he isk o sys em e o s and mal unc ion o he AI
agen ecosys em pollu ion by inaccu a e da a. In 2025, he Da win Gödel Machine (DGM) was
de eloped, a sel -imp o ing Coding AI Agen sys em ha i e a i ely modi ies i s own code and
empi ically alida es each change using coding benchma ks which was e ec i e in
expe imen s. The sp ead o obo s may open he way in o ully au oma ed ab ica ions.
The human con ol o AI can be done by obse a ion o ou pu and ac ion, bu can be
complemen ed by moni o ing o he chain o hough s, i.e., he in e nal AI p o ocols ha show
hei easoning ( hinking) s eps. The abili y o obse e he chain o hough o an AI is called
moni o abili y. The e a e conce ns ha no el AI a chi ec u es may educe he moni o abili y
despi e indings o hidden misalignmen (scheming) whe e on ie AI models hide hough s
39
Ba ne /Sche 2025, Losey 2025, MIRI 2025
40
Adewumi e al. 2025, Tonsen 2025
41
Adewumi e al. 2025, Ba ne /Sche 2025, Losey 2025, MIRI 2025,Tonsen 2025
42
Ba ne /Sche 2025, MIRI 2025
43
Landymo e 2025
A i icial_In elligence_and_Machine_E olu ion 8
e.g., o ge ewa ds o o a oid ein o cemen lea ning. Va ious esea che s exp ess se ious
conce ns abou a po en ially uncon ollable e olu ion e.g., by sel -lea ning. The key conce n is
ha humans a e cu en ly no eally p epa ed o his si ua ion and may no be able o p o ide
su icien o e sigh , i.e., he si ua ion may ge ou o human con ol and o his eason, a s ic
go e nance wi h egula ions and supe ision should be es ablished.
3 Re e ences
Adewumi, T. e al. (2025): AI Mus no be Fully Au onomous. Machine Lea ning G oup,
EISLAB, Sweden a Xi :2507.23330 1 [cs.AI] 31 Jul 2025
Al man, S. (2025): Gen le Singula i y. Sam Al man’s Blog 10 June 2025
A enas, Y., B ekelmans, B. (2025): Empowe ing mul i-agen apps wi h he open Agen 2Agen
(A2A) p o ocol. Mic oso News 07 May 2025
A ie a, A.B. e al. (2020): Explainable A i icial In elligence (XAI): Concep s, axonomies,
oppo uni ies, and challenges owa d esponsible AI. In o ma ion Fusion 58(2020), p. 82–111
Beye e , J., Ki chne , E. e al. (2025): KI in de Robo ik. Flexible und anpassba e Sys eme
du ch in e ak i es Le nen. DOI: h ps://doi.o g/10.48669/pls_2025
Ba ne , P., Sche , A. (2025): AI Go e nance o A oid Ex inc ion: The S a egic Landscape
and Ac ionable Resea ch Ques ions Machine In elligence Resea ch Ins i u e Be keley,
Technical Go e nance Team May 2025
Chaudh y, H., Klein, L. (2023): A i icial In elligence and Nuclea Weapons: P oblem
Analysis and US Policy Recommenda ions. policy@ u u eo li e.o g 14 h No embe 2023
Fu u e o Li e Ins i u e (FLI)
Chen, N. (2025): In eg a ing AI in o Semiconduc o Design and Fab ica ion: Me hodologies,
Challenges and Fu u e P ospec s. ITM Web o Con e ences 78, 03007 (2025)
h ps://doi.o g/10.1051/i mcon /20257803007 CSEIT 2025
Deep Mind (2025): FACTS Benchma k Sui e: Sys ema ically e alua ing he ac uali y o
la ge language models. The FACTS eam o Google Deep Mind. Decembe 9, 2025
DSIT (2025): AI Insigh s - Agen ic AI. Pape o he Depa men o Science, Inno a ion and
Technology and he Go e nmen Digi al Se ice o he Uni ed Kingdom
EAC (2024): EAC In e na ional Consul ing Execu i e Repo . Poli ical S ee ing P ocesses in
China in Co e Segmen s o he Pho onics Indus y. Commissioned by he Eu opean
Technology Pla o m Pho onics21
Goldie, A. e al. (2024): Addendum: A g aph placemen me hodology o as chip design
doi.o g/10.1038/s41586-024-08032-5 Published online 26 Sep embe 2024
Goldie, A., Mi hoseini, A. (2024): How AlphaChip ans o med compu e chip design
Science Sep embe 26, 2024
G ey, M., Sege ie, C.R. (2025): The AI Risk Spec um F om Dange ous Capabili ies o
Exis en ial Th ea s. F ench Cen e o AI Sa e y (CeSIA) 20 Augus 2025
Guan, M.Y. e al. (2025): Moni o ing Moni o abili y. OpenAI Decembe 18, 2025
He, Y. e al. (2024): Secu i y o AI Agen s a Xi :2406.08689 3 [cs.CR] 17 Dec 2024
Hin on, G. (2025): KI wi d bald deu lich in elligen e als wi sein. Handelsbla 23-25 Mai
2025, S.26-27
A i icial_In elligence_and_Machine_E olu ion 9
Hi e lab (2025): China’s CHIPX Pho onic Chip P oduc ion Line Se o Accele a e AI, 6G,
and Quan um Compu ing De elopmen : SCMP h ps://hi e lab.com/china-chipx-pho onic-
chip-ai-6g-quan um-compu ing/ June 17, 2025
Khan, S.M., Mann, A., Pe e son, D. (2021): The Semiconduc o Supply Chain: Assessing
Na ional Compe i i eness Cen e o Secu i y and Eme ging Technology (CSET) Issue B ie
Janua y 2021
Ko bak, T. e al. (2025): Chain o Though Moni o abili y: A New and F agile Oppo uni y o
AI Sa e y. a Xi :2507.11473 1 [cs.AI] 15 Jul 2025
Landymo e, F. (2025): An h opic's Chie Scien is Says We' e Rapidly App oaching he
Momen Tha Could Doom Us All. Fu u ism Published Dec 13, 2025
La sen, B. e al. (2024): Na iga ing he AI F on ie : A P ime on he E olu ion and Impac o
AI - Wo ld Economic Fo um Ai Go e nance Alliance Sa e Sys ems and Technologies Wo king
G oup - Wo ld Economic Fo um Whi epape
Lee, W.-Y. (2025): Reu e s Exclusi e: N idia, Foxconn in alks o deploy humanoid obo s a
Hous on AI se e making plan . June 20, 2025 h ps://www. eu e s.com/wo ld/china/n idia-
oxconn- alks-deploy-humanoid- obo s-hous on-ai-se e -making-plan -2025-06-20/
Losey, R. (2025): A aid o AI? Lea n he Se en Ca dinal Dange s and How o S ay Sa e – AI
E hics. Ap il 25, 2025
Ma ini, A. (2025): 3 hings o know abou I onwood, ou la es TPU. Google Cloud, No 25,
2025
MIRI (2025): The P oblem - Machine In elligence Resea ch Ins i u e Be keley Oc obe
2025We
Open AI (2025a): E alua ing chain-o - hough moni o abili y OpenAI Resea ch Publica ion
Decembe 18, 2025
OpenAI (2025b): De ec ing and educing scheming in AI models. Publica ion Resea ch
Sep embe 17, 2025
Sche le , I. (2025): Resea che s un eil g oundb eaking 3D chip o accele a e AI. S an o d
Repo Decembe 10, 2025 h ps://news.s an o d.edu/s o ies/2025/12/monoli hic-3d-chip-
ound y-b eak h ough-ai
Shah, R. e al. (2025): An App oach o Technical AGI Sa e y and Secu i y Google DeepMind
Resea ch Pape
Swayne, M. (2025): China’s New Pho onic Quan um Chip P omises 1,000-Fold Gains o
Complex Compu ing Tasks. The Quan um Inside h ps:// hequan uminside .com/au ho /ma -
swayne/) No embe 15, 2025
Ta di , A. (2025): Quil e Shows AI Can Now Design Re Ha dwa e Wi h he Wo ld’s Fi s
Mac Enginee ed Compu e . Published Decembe 10, 2025 By An oine Ta di , CEO &
Founde o Uni e.AI
Zhang, J. e al. (2025): Da win Gödel Machine (DGM) – Open-ended E olu ion o Sel -
imp o ing Agen s. a Xi :2505.22954 2 [cs.AI] 26 Sep 2025
Zhang, R. e al. (2025): QiMeng: Fully Au oma ed Ha dwa e and So wa e Design o
P ocesso Chip. a Xi :2506.05007 1 [cs.AR] 5 Jun 2025
Zia, T. (2025): The Mul i-Agen Pa adox: Why Mo e Agen s Can Lead o Wo se Resul s.
Uni e.AI Decembe 22, 2025