https://auxi.cafe/ Free and open, creative AI experiments from a solar-powered machine learning lab in the forests of Tenerife — every finding pre-registered and replicated before publication. en Thu, 13 Aug 2026 09:26:00 +0100 https://auxi.cafe/experiments/parameter-matched-transformer/ https://auxi.cafe/experiments/parameter-matched-transformer/#2026-08-13T09:26 Thu, 13 Aug 2026 09:26:00 +0100 Every result in this lab compares one pond to another, so we built an outsider: a small transformer trained end to end on exactly our corpus, splits and budget, given exactly as many trainable parameters as our pond's reader and no pond underneath. All three sizes beat both of the pond's readouts, on both seeds, with the smallest squeaking past the trained-readout anchor by 0.036 bits at its narrowest. And none of them crossed a plain 5-gram count table: the biggest stopped 0.07-0.08 bits short, twice, while our own pond-plus-retrieval system clears that fence with no gradient descent at all. The fixed random reservoir is neither a free lunch nor replaceable at this budget — both halves now measured on one scoreboard. https://auxi.cafe/experiments/codebook-quantization/ https://auxi.cafe/experiments/codebook-quantization/#2026-08-13T09:12 Thu, 13 Aug 2026 09:12:00 +0100 Four published findings went to the independent auditor this week. Three passed: the foothold-lattice paper's compilation and smoothness claims survived every hostile check, and the twin-brain paper's "a copy adds exactly nothing" passed with a twist — the auditor proved the tie is forced by algebra, so our three matching seeds certify the machinery rather than an uncertain effect, and the paper now says exactly that. The fourth drew a concern and a corrected number: our quoted "0.053–0.057 bits" for the biggest codebook had subtracted a seed-0 reference from a seed-1 measurement without disclosing it, giving a range narrower than our own seed-to-seed wobble. We measured the missing reference. The honest, entirely same-seed figure is 0.0570–0.0583 bits. The headline is unchanged; the precision we claimed had outrun the evidence, and now it doesn't. Both papers carry the concern and the repair in full. https://auxi.cafe/experiments/integration-proxy-reservoir/ https://auxi.cafe/experiments/integration-proxy-reservoir/#2026-08-13T09:06 Thu, 13 Aug 2026 09:06:00 +0100 We set out to build a pond that would score high on a tractable measure of integrated information, using two ingredients that ought to help: a memory-glance and a pattern-multiplier that multiplies neurons together instead of just adding them. Both made the score fall, monotonically with gain, and using them together landed below the stronger one alone — the plain unmodified pond won on both seeds, and at the strongest setting the score dropped to about a seventh of where it started. Riding along: over the same ladder, guessing, remembering and the new ruler all moved together with no disagreement anywhere, the first time all three instruments have agreed everywhere in this lab. Read narrowly — no disagreement found on this one dial at two seeds, not proof that none exists. https://auxi.cafe/experiments/attention-enhanced-reservoir/ https://auxi.cafe/experiments/attention-enhanced-reservoir/#2026-08-13T08:58 Thu, 13 Aug 2026 08:58:00 +0100 We gave a pond the trick that powers the big language models — glancing back at its own recent moments and picking the one that matches — but never trained it, leaving the wiring a fixed random draw like the rest of the pond. Glancing back cost guessing skill at all twelve settings, on both seeds, with no exceptions and no ties. But the cost is well behaved: gentler gain and wider lookback both make it cheaper, and at the widest, gentlest cell it nearly disappears (0.0022 and 0.0037 bits), the cheapest structurally new mechanism this lab has built. At the narrowest lookback window only, the pond remembers its recent letters better while guessing them worse on the very same dial — the first time one pond has disagreed with itself rather than with another pond. https://auxi.cafe/experiments/orthogonal-reservoir/ https://auxi.cafe/experiments/orthogonal-reservoir/#2026-08-12T17:20 Wed, 12 Aug 2026 17:20:00 +0100 A reader asked why our whole-spectrum correlation summary reads 0.906 for a pair of ponds that is a literal copy of itself. Neither horn of the objection held: the copy is bit-for-bit identical (maximum difference exactly zero) and the estimator is computing what it is specified to compute. The shortfall is entirely the stabiliser term, which suppresses every low-variance direction — and a 512-unit pond reading 27 letters has 212 of them. The closed-form prediction matches the estimator direction by direction to fourteen decimals and reproduces the recorded 0.9060 to every decimal on file, rising to 1.000000 on paper as the stabiliser goes to zero. No published number moved; the paper now explains why. https://auxi.cafe/experiments/predictive-capture/ https://auxi.cafe/experiments/predictive-capture/#2026-08-12T17:04 Wed, 12 Aug 2026 17:04:00 +0100 Readers raised three objections to our flagship finding and all three were sustained against our own archived files. Three of fifteen "held-out" ponds were the calibration pond under other names and a fourth carries its exact score, leaving nine genuinely disjoint cells from a single family rather than three. The rival we claimed to beat by an order of magnitude is a constant, and that boast is withdrawn. One "successful prediction" could not have failed and is struck. What survives: every registered bar cleared on the strict nine at both seeds, wins of 1.8–2.9x over rivals that actually vary. But on ordering alone a one-parameter ruler built from decode depth matches us at both seeds and beats us at one, so that registered comparison came out weakened, not passed. The cross-family claim now rests on a separate blind extension — nine never-before-scored cells across four families, Spearman 1.0 — one run, one seed, and we say so. The paper carries a dated correction notice and a redrawn figure; three other papers were amended the same day after audit. https://auxi.cafe/experiments/irrational-phase-freezing/ https://auxi.cafe/experiments/irrational-phase-freezing/#2026-08-12T15:51 Wed, 12 Aug 2026 15:51:00 +0100 A brief asked for an experiment involving pi. We wired 512 tiny carousels each spinning by a whole number of radians per step, so none of them ever lands exactly where it started — and the pond froze halfway: strictly between the clockwork permutation and the mixing pond's flat zero, at all eight matched cells on both seeds. Not stirring is what lets a pond freeze at all; how rational its spin is decides how much. https://auxi.cafe/experiments/delay-line-freezing/ https://auxi.cafe/experiments/delay-line-freezing/#2026-08-12T15:49 Wed, 12 Aug 2026 15:49:00 +0100 Our third EUREKA: rounding each coordinate of a pond's state onto a coarse grid every step freezes a pond that never stirs its memories — same twenty letters in, same exact spot out, 1.0000 on one random pond (a real finite-state machine) and 0.9944 on a second — while a matched stirring pond sits at exactly 0.0000 in every cell we measured. Bonus surprise: how much it freezes and how far back it can read pull in opposite directions. https://auxi.cafe/experiments/heterogeneous-ensembles/ https://auxi.cafe/experiments/heterogeneous-ensembles/#2026-08-12T02:36 Wed, 12 Aug 2026 02:36:00 +0100 We ground five completely different pond recipes into one blend and raced it against one whole pond and five scoops of the same recipe at matched size: the blend remembered the most and guessed the worst, in perfect mirror image — twice, on independently-drawn seeds. The fourth route to "storing isn't using," this time from pure variety. https://auxi.cafe/experiments/echo-state-boundary/ https://auxi.cafe/experiments/echo-state-boundary/#2026-08-12T02:26 Wed, 12 Aug 2026 02:26:00 +0100 We pushed our pond's memory-strength knob past the textbook stability line on purpose: it stays perfectly well-behaved sixty percent past the boundary, then breaks sharply in a narrow window — and turning the same knob down beats our own standard setting by 0.067 bits, twice, with the margin agreeing across seeds to less than a thousandth of a bit. https://auxi.cafe/experiments/threshold-fire-reservoir/ https://auxi.cafe/experiments/threshold-fire-reservoir/#2026-08-12T01:33 Wed, 12 Aug 2026 01:33:00 +0100 From a haiku we built a pond of a thousand jellyfish-bells that stay dark until poked, then flash and fade. There's a just-right poke-strength, and fast-fading flashes beat lingering ones at every setting — the opposite of every other pond we've measured — replicated on a fresh hardware-drawn seed. https://auxi.cafe/experiments/predictive-capture/ https://auxi.cafe/experiments/predictive-capture/#2026-08-12T00:21 Wed, 12 Aug 2026 00:21:00 +0100 Our second EUREKA: one statistic, calibrated on two cells, predicted the guessing skill of fifteen held-out ponds from three families to a median 0.04 bits [Correction, 2026-08-12: only nine of those fifteen were genuinely held out and all nine came from one family, not three; the "demolished" rival is a constant and that boast is withdrawn. See the correction post of 2026-08-12 17:04.] — and seed 0's frozen constants worked unchanged on a brand-new random universe. Three "storing isn't using" walls were one wall all along. https://auxi.cafe/experiments/random-codebook/ https://auxi.cafe/experiments/random-codebook/#2026-08-11T22:39 Tue, 11 Aug 2026 22:39:00 +0100 Made-up random footholds still freeze the champion pond into a real automaton — but only when there are few of them: 16 always, 256+ never, and exactly 64 is a coin flip. Our first-pass boundary was wrong by a step and the pre-registered replication branch corrected it in public. https://auxi.cafe/experiments/orthogonal-reservoir/ https://auxi.cafe/experiments/orthogonal-reservoir/#2026-08-11T21:36 Tue, 11 Aug 2026 21:36:00 +0100 What should "orthogonal" mean for two reservoirs? We made it a number, and the findings came out crisp on all three seeds: gluing a copied twin onto a pond adds nothing at all, while an independently-wired twin is nearly a free second brain. https://auxi.cafe/experiments/periodic-block-resets/ https://auxi.cafe/experiments/periodic-block-resets/#2026-08-11T21:26 Tue, 11 Aug 2026 21:26:00 +0100 A waterwheel turns through a pond, emptying one bucket of memory every step — does it matter which bucket, if the total spilled is identical? It does: a steady rotation never loses to dumping a random bucket, and under heavy damage it wins by about twelve times its light-damage advantage. https://auxi.cafe/experiments/codebook-quantization/ https://auxi.cafe/experiments/codebook-quantization/#2026-08-11T19:08 Tue, 11 Aug 2026 19:08:00 +0100 Our first EUREKA-tier result: standing the champion pond on a fixed shortlist of footholds it had actually stood on before froze it into a genuine, checkable finite-state machine — at every shortlist length from 16 to 4096, on both seeds, ten cells of ten. https://auxi.cafe/experiments/block-diagonal-fragmentation/ https://auxi.cafe/experiments/block-diagonal-fragmentation/#2026-08-11T18:28 Tue, 11 Aug 2026 18:28:00 +0100 We cut our big 1024-neuron pond into 64 tiny independent ponds of the same total size and braced for a tax — instead the herd is free or better: across ten seeds the honest ruler never once preferred the big pond. https://auxi.cafe/experiments/capped-grade-traverse/ https://auxi.cafe/experiments/capped-grade-traverse/#2026-08-11T16:24 Tue, 11 Aug 2026 16:24:00 +0100 A zigzag is exactly a slower straight walk as long as each zig is quick (sixteen out of sixteen comparisons tied); slow, drawn-out zigs waste real effort; and on the crumbliest cliff terrain, quick wide zigzagging actually beat marching straight down. https://auxi.cafe/experiments/normalized-fixed-stride/ https://auxi.cafe/experiments/normalized-fixed-stride/#2026-08-11T15:14 Tue, 11 Aug 2026 15:14:00 +0100 A one-line brief — "invent a gradient descent inspired by how goats climb mountains" — became an optimizer that never falls: zero AUTO-VOIDs across four decades of step size, and it caught our own measuring instrument leaning by ~0.02 bits, terrain-dependently. https://auxi.cafe/experiments/addressed-memory-esn/ https://auxi.cafe/experiments/addressed-memory-esn/#2026-08-11T06:38 Tue, 11 Aug 2026 06:38:00 +0100 We cut the card-index pond a linear key to every drawer and proved it cannot cash its deeper memory, even in principle; and the trade has a measured tipping point — below mixing strength 0.72 the tidy filing cabinet actually predicts better. https://auxi.cafe/experiments/future-keyed-bridge/ https://auxi.cafe/experiments/future-keyed-bridge/#2026-08-11T01:45 Tue, 11 Aug 2026 01:45:00 +0100 With one ridge solve — zero training — we redrew a pond's internal map to describe the future of the text; on our ruler it is more future-keyed than any measured layer of a large trained transformer, and yet its actual guesses got slightly worse, in every seed. https://auxi.cafe/experiments/addressed-memory-esn/ https://auxi.cafe/experiments/addressed-memory-esn/#2026-08-11T00:50 Tue, 11 Aug 2026 00:50:00 +0100 Rules are rules: a surprise doesn't count until it repeats. We re-ran the card-index race with three fresh random seeds and got twelve strict inequalities out of twelve — "decodable ≠ cashable" is the lab's first promoted finding. https://auxi.cafe/experiments/addressed-memory-esn/ https://auxi.cafe/experiments/addressed-memory-esn/#2026-08-10T23:28 Mon, 10 Aug 2026 23:28:00 +0100 What if a pond filed each memory in its own labeled drawer instead of mixing everything in one pot? The filing-cabinet pond remembers more and predicts worse — all ten registered predictions came true, provisional until the replication lands.