diff --git a/day17.py b/day17.py
new file mode 100644
index 0000000..402aff5
--- /dev/null
+++ b/day17.py
@@ -0,0 +1,334 @@
+#!/usr/bin/env python3
+
+import re
+from random import randrange
+from array import array
+from copy import copy
+
+class VM():
+    def __init__(self, reg: array = array('q'), mem: list = [], pc: int = 0):
+        self.reg = copy(reg)
+        self.mem = copy(mem)
+        self.pc = copy(pc)
+        self.outbuf = []
+
+    def run(self, abort: int = 0):
+        while self.pc < len(self.mem):
+            if abort>0 and len(self.mem) > abort: return
+
+            op = self.mem[self.pc]
+            self.pc += 1
+
+            if op == 0b000: # adv
+                self.reg[0] = self.reg[0] // (2**self._combo(self.mem[self.pc]))
+                self.pc += 1
+            if op == 0b001: # bxl
+                self.reg[1] = self.reg[1] ^ self.mem[self.pc]
+                self.pc += 1
+            if op == 0b010: # bst
+                self.reg[1] = self._combo(self.mem[self.pc]) & 0b111
+                self.pc += 1
+            if op == 0b011: # jnz
+                self.pc = self.mem[self.pc] if self.reg[0] != 0 else self.pc+1
+            if op == 0b100: # bxc
+                self.reg[1] = self.reg[1] ^ self.reg[2]
+                self.pc += 1
+            if op == 0b101: # out
+                self.outbuf.append(self._combo(self.mem[self.pc]) & 0b111)
+                self.pc += 1
+            if op == 0b110: # bdv
+                self.reg[1] = self.reg[0] // (2**self._combo(self.mem[self.pc]))
+                self.pc += 1
+            if op == 0b111: # cdv
+                self.reg[2] = self.reg[0] // (2**self._combo(self.mem[self.pc]))
+                self.pc += 1
+
+    def _combo(self, c) -> int:
+        if c == 0: return 0
+        if c == 1: return 1
+        if c == 2: return 2
+        if c == 3: return 3
+        if c == 4: return self.reg[0]
+        if c == 5: return self.reg[1]
+        if c == 6: return self.reg[2]
+        raise RuntimeError()
+
+    def flush(self, binfmt = False):
+        if not binfmt:
+            print(','.join(map(str, self.outbuf)))
+        else:
+            for o in self.outbuf:
+                print(f"{o:03b},", end="")
+            print()
+
+    def cmp(self, prog):
+        return self.outbuf == prog
+
+
+reg: array = array('q')
+mem: list = []
+pc: int = 0
+
+with open('day17') as f:
+    reg_a = int(re.match('Register A: ([0-9]+)', f.readline()).groups()[0])
+    reg_b = int(re.match('Register B: ([0-9]+)', f.readline()).groups()[0])
+    reg_c = int(re.match('Register C: ([0-9]+)', f.readline()).groups()[0])
+
+    reg = array('q', [reg_a, reg_b, reg_c])
+
+    f.readline()
+
+    mem_str = re.match('Program: (.*)', f.readline()).groups()[0]
+
+    for c in mem_str:
+        if c == ',': continue
+        mem.append(int(c))
+
+# day 1
+
+vm = VM(reg, mem, pc)
+vm.run()
+vm.flush()
+
+# day 2
+
+# enumerates all 3 bit number, then increases size by one
+# enumerates them again
+
+# -> we can find first boundaries by getting numbers where
+# output buffer length is too small or too large
+
+# but first we need to find where to even start
+# we know the reg_a from the puzzle input is
+# too small, but start from here and increas rapidly:
+
+reg_a = 17323786
+while True:
+    reg = array('q', [reg_a, reg_b, reg_c])
+    vm = VM(reg, mem, pc)
+    vm.run(abort=len(mem))
+    if vm.cmp(mem): break
+
+    if len(mem) == len(vm.outbuf): break
+
+    reg_a *= 10
+
+
+print(reg_a)
+vm.flush()
+
+# we get: 173237860000000
+# producing a program: 4,4,4,4,1,6,0,0,0,2,1,0,0,1,3,0
+# that's exactly the length we want
+
+# let's find our first set of boundaries so we can start a search
+known_bytes = vm.outbuf[-2:]
+init_reg_a = reg_a
+upper_bound = -1
+lower_bound = -1
+
+reg_a = init_reg_a
+while True:
+    reg = array('q', [reg_a, reg_b, reg_c])
+    vm = VM(reg, mem, pc)
+    vm.run(abort=len(mem))
+    if vm.cmp(mem): break
+
+    if len(vm.outbuf) > len(mem):
+        upper_bound = reg_a
+        break
+
+    reg_a *= 10
+
+
+print(reg_a)
+vm.flush()
+
+reg_a = init_reg_a
+while True:
+    reg = array('q', [reg_a, reg_b, reg_c])
+    vm = VM(reg, mem, pc)
+    vm.run(abort=len(mem))
+    if vm.cmp(mem): break
+
+    if len(vm.outbuf) < len(mem):
+        lower_bound = reg_a
+        break
+
+    reg_a //= 10
+
+
+print(reg_a)
+vm.flush()
+
+
+# that was fast again and did not produce very enlightening bounds,
+# nevertheless we have bounds:
+#
+# ub: 1732378600000000
+# out: 4,4,0,7,0,7,6,0,3,5,1,5,4,2,4,4,2
+#
+# mid: 173237860000000
+# out: 4,4,4,4,1,6,0,0,0,2,1,0,0,1,3,0
+#
+# lb: 17323786000000
+# out: 4,4,5,5,5,0,1,4,6,4,4,4,7,3,7
+
+
+# helper function
+def cmp_bytes(a,b):
+    if len(a) != len(b): return []
+
+    same = []
+    for i in range(1, len(a)):
+        if a[-i] == b[-i]:
+            same.append(a[-i])
+            continue
+        break
+
+    same.reverse()
+    return same
+
+# params for our algorithm to tune
+generation = 1
+population = 1
+mu = 173237860000000 # current best guess
+mu_hit = 2 # correct positions of mu
+mu_len = 16
+sd = 171505481400000 # standard deviation, (up-low)*0.1 - I pulled that out of my nose
+lo = 17323786000000 # lower bound
+lo_hit = 0
+lo_len = 15
+up = 1732378600000000 # upper bound
+up_hit = 0
+up_len = 17
+
+
+from day17_truncnorm import Truncnorm
+truncnorm = Truncnorm(mu, sd, lo, up)
+
+import random
+
+prodigy = mu
+prodigy_hit = mu_hit
+prodigy_len = mu_len
+
+mus = [(mu, mu_hit, mu_len)]
+
+evolve = True
+while evolve:
+    # each generation we birth 10000 kin. The new generation's genes center
+    # around the best produced offsprings so far. However, we allow for
+    # mutations with decreasing likelihood the more deformed the kin would be.
+    #
+    # The mutations are bounded. Mutations outside these bounds lead to
+    # a miscarriage.
+    #
+    # With each generation we re-asses and adapt our parameters.
+    #
+    # Occasionally a catastrophic event happens erasing the whole
+    # popluation. This hopefully prevents us from being trapped in a
+    # local optima.
+    #
+    # To preserve their legacy during these catastrophic events the mu
+    # collect and freeze prodigies - the most promising genes they
+    # know.
+    #
+    # All this allows us to search a search space probabilistically which
+    # would be way too large to search exhaustively.
+
+    catastrophy = random.randint(0,100)
+    if catastrophy % 10 == 0:
+        mus.append((mu, mu_hit, mu_len))
+        print(f"\033[31mA catastrophy erased the entire population. Not all is lost, a new {prodigy=} awakes.\033[0m")
+        mu = prodigy
+        mu_hit = prodigy_hit
+        mu_len = prodigy_len
+        print(f"\033[31mIn a last attempt to preserve their culture the best known mus are engraved in a stone wall:\033[0m {mus=}")
+        sd = round(random.uniform(0.00001, 1.0)*(up-lo))
+        print(f"\033[31mEvolution restarts with {sd=}\033[0m")
+        population += 1
+        generation = 1
+
+    print(f"\033[35mGeneration {generation} (Population {population}):\033[0m {mu=}, {sd=}, {lo=}, {up=}")
+
+
+    for kin in truncnorm.draw(30000):
+        kin = round(kin)
+
+        reg = array('q', [kin, reg_b, reg_c])
+        vm = VM(reg, mem, pc)
+        vm.run()
+
+        if len(cmp_bytes(vm.outbuf, mem))>mu_hit:
+            mu = kin
+            mu_hit = len(cmp_bytes(vm.outbuf, mem))
+            mu_len = len(vm.outbuf)
+            print(f"\033[32mFound new best kin: {mu=}, {mu_hit=}\033[0m", end=" ")
+            vm.flush()
+            continue
+
+        if len(cmp_bytes(vm.outbuf, mem))==mu_hit and kin<mu:
+            mu = kin
+            mu_hit = len(cmp_bytes(vm.outbuf, mem))
+            mu_len = len(vm.outbuf)
+            print(f"\033[32mFound new best kin: {mu=}, {mu_hit=}\033[0m", end=" ")
+            vm.flush()
+            continue
+
+        if kin < mu and kin < prodigy and len(cmp_bytes(vm.outbuf, mem)) > 3:
+            prodigy = kin
+            prodigy_hit = len(cmp_bytes(vm.outbuf, mem))
+            prodigy_len = len(vm.outbuf)
+            print(f"Found new prodigy: {prodigy=}, {prodigy_hit=}", end=" ")
+            vm.flush()
+
+        if len(vm.outbuf) < mu_len and kin>lo:
+            lo = kin
+            lo_hit = len(cmp_bytes(vm.outbuf, mem))
+            lo_len = len(vm.outbuf)
+            print(f"Found new lower bound: {lo}, {lo_hit=}, {lo_len=}", end=" ")
+            vm.flush()
+            continue
+
+        if len(vm.outbuf) > mu_len and kin<up:
+             up = kin
+             up_hit = len(cmp_bytes(vm.outbuf, mem))
+             up_len = len(vm.outbuf)
+             print(f"Found new upper bound: {up}, {up_hit=}, {up_len=}", end=" ")
+             vm.flush()
+             continue
+
+        assert(lo < mu)
+        assert(mu < up)
+
+
+    sd = sd//generation
+    if sd<100:
+        sd = round(random.uniform(0.00001, 1.0)*(up-lo))
+        print(f"\033[36mA fungus infected the generation and changed mutations. New mutation with {sd=}\033[0m")
+
+    print(f"{sd=}")
+    truncnorm = Truncnorm(mu, sd, lo, up)
+    generation += 1
+
+# we evolved to a species capable of genetic engineering the very best
+# prodigy we found is 164278764924544. We even found that multiple
+# times. Now brute force ourselves to the top of the food chain.
+#
+# Our probabilistic algorithm is good a finding a a very good solution
+# (up to 15 elements correct) but it is actually bad at pinpointing
+# the exactly best solution.
+
+mu = 164278764924544
+for breed in range(mu-(2**(5*3)), mu+(2**(5*3))):
+    reg = array('q', [breed, reg_b, reg_c])
+    vm = VM(reg, mem, pc)
+    vm.run()
+
+    if vm.cmp(mem):
+        print(f"Self replicating {breed=}")
+        vm.flush()
+
+print(reg_a)
+vm.flush()