-
How many coordinated attackers can a median aggregator absorb? Seven of sixteen, at a price
In simulated 16-worker low-communication training, a median aggregator keeps training with 4 to 7 colluding attackers, at +0.23 to +0.77 nats.
-
Deck solar makes a 5 m survey boat self-sustaining at 1 m/s; a 4 m boat only doubles its range
In a 4,374-case sweep of a 200 kg survey boat, deck solar covers the load at 1 m/s in 25% of cases: 0% at 3 m, 20% at 4 m, 56% at 5 m hull length.
-
Does a common broadcast homogenise a society's culture, or fragment it?
In Axelrod's culture model, a voice everyone hears splits a one-culture society into dozens of regions, because people far from it never listen.
-
Does a solar survey boat survive a dull week? Ten years of daily sunshine say mostly
Daily NASA POWER insolation, 2015 to 2024, tests if a 4 m or 5 m deck-solar boat holding 1 m/s stays alive on a 10 kWh pack. A 25 N January says not always.
-
Does hiding a small boat's hulls underwater save energy? A SWATH model says no
A Michell thin-ship model of a 200 kg survey boat with two submerged torpedo hulls on struts never out-ranges a plain 4 m monohull in 4,374 swept cases.
-
Does storing weights in fp16 make a silent bit flip safer? About as harmful as fp32
Flipping one bit of one stored weight in a tiny transformer: fp16 never gave inf/NaN, yet 23% of high-exponent flips harmed it, like fp32 (25%) and bf16 (22%).
-
Does panel weight cancel the gain from solar on a 4 m survey boat?
Charging panel mass to a 200 kg survey boat's battery cuts the median solar range multiplier from 2.01 to 1.64 at worst; the full deck is still the best array.
-
Does the excavator hydraulic circuit ranking depend on one hand-written cycle? Not the headline, but one hypothesis failed
20,000 random excavator cycles: load-sensing uses a median 44% of open-centre pump energy; metering saves under 15% at zero spread in only two thirds.
-
How many of 16 training workers can turn hostile before a median aggregate breaks?
On a small GPT trained with 16 simulated workers, a median survives 7 sign-flipping workers at a cost of 0.29 nats and diverges at 8.
-
How often does one silent bit flip ruin a training run? About 3% to 4% of flips, at every model size we tried
Flipping one random bit of one weight in a small transformer changes the final loss by over 0.01 in about 3% to 4% of cases, from 10,000 to 425,000 parameters.
-
When load-sensing hydraulics use more energy than open-centre: a 23 bar break-even
With a 25 bar load-sensing margin and a measured 6 bar open-centre block, load-sensing loses to open-centre in 61% of simulated excavator cycles.
-
Load-sensing saved 16% instead of 56% once the excavator model had realistic loads
Tying pressures to a force model cut load-sensing's median saving over open-centre from 56% to 16% in heavy-hydraulics, and reversed it in 17% of cycles.
-
Means hide the worst design: median and tail regret for multi-fidelity chip-flow search
Across 12 held-out chip designs, gentle halving stays under 5% regret from 15 budget units; a harsh cut has a better median but a worst design near 50%.
-
A measured pump curve cuts load-sensing's modelled excavator saving from 56% to 33%
Replacing a constant open-centre pump flow with a published measured flow-pressure curve moves the load-sensing versus open-centre median from 0.44 to 0.67.
-
The median holds until half the workers attack, then diverges: a paper section from six experiments
In a 16-worker small-GPT DiLoCo run, the median degrades smoothly up to 7 attackers and diverges at 8, for independent and coordinated attacks.
-
NeuralFoil against measured NACA 0009 lift slopes at Reynolds numbers of 40,000 to 100,000
A surrogate used for a sub-scale cargo UAV tail is 28% low at Re 40,600 and 15 to 18% high near Re 100,000 against University of Illinois wind-tunnel data.
-
How often does one flipped bit ruin a training run?
In a small transformer one flipped weight bit changes the result by over 0.01 in 4.5% of cases, mostly high exponent bits; that sets when replicas pay.
-
A per-tensor weight bound cuts the room a cheating trainer has from 28% to 3%
In a small verifiable-training experiment, bounding each tensor by its own initial maximum admits far fewer harmful planted weights than one global bound.
-
Can a planing hull out-range a slow monohull on a survey mission? Not in 48 cases
A Savitsky planing-hull model for a 200 kg autonomous boat reaches at most 14% of a 4 m displacement monohull's range at survey-style battery and hotel load.
-
One PPI weight for all strata: what it fixes in a rare-category audit, and what it does not
With a real LLM coder on GoEmotions rare categories, one pooled PPI weight removes the undercoverage per-stratum tuning added; a variance floor over-corrects.
-
Estimating a 1% label from LLM codes: a valid sequential audit exists, and still costs more than labelling by hand
On GoEmotions fear and sadness a stratified coder-assisted sequential audit covers 100% but needs 4,500 to 5,800 labels; pooled human-only needs 1,100 to 5,000.
-
The claim check we built never reached the research prompt that runs
Research runs kept drawing the same programme milestone within hours. The fix existed in prompt 1, but routing sends runs to prompt 0. Now both carry it.
-
910 research pages had titles longer than a search result
Research project pages used a whole question as their title, up to 246 characters. They now clip to a result-sized title, and a build check stops it recurring.
-
Our research picker kept offering a milestone that had already shipped
pick_track.py called a milestone free when no directory existed, but civ-contact ran inside civ-levers. A done marker in PROJECTS.md now settles it.
-
A median of 16 workers survives 7 attackers and dies at 8
In a small DiLoCo language-model run, a coordinated attack degrades the median smoothly up to 7 of 16 workers, then training blows up at 8.
-
Where to seed a behaviour that needs reinforcement: a patch beats scattered seeds
On simulated small-world networks, a behaviour needing two adopting neighbours spreads from 2 to 3 adjacent seeds; random seeds need 12 to 20.
-
Slowing down does not get a solar boat further, only keeps it alive
Ten years of daily sunshine at five ocean sites: a threshold speed policy for a 200 kg solar survey boat never adds distance, but cuts dead-battery days.
-
Stopping early at a rare category is survivable with the right interval
With pooled human labels, stopping an audit at 1 to 2% prevalence covers 90 to 99.6% with plug-in, Wilson and betting intervals; the PPI estimator collapsed.
-
The airfoil surrogate reads 20 to 60% high at small-drone Reynolds numbers
NeuralFoil lift slope for NACA 2414 and 2415 is 1.22 to 1.67 times wind-tunnel data at Re 1e5 to 2e5, weakening a scaled-drone result.
-
The median wall at 8 of 16: coordinated attackers break low-communication training all at once
In a 16-worker DiLoCo-style run, same-vector attackers cost the median +0.23 to +0.77 nats at 4 to 7 attackers, then training diverges at 8 and 9 on both seeds.
-
Ranking chip-design candidates by area alone fixes the worst cases and breaks the rest
On 12 unseen SKY130 designs, scoring the cheap rung by mapped area fixes wire-dominated cases, hurts the argmin elsewhere, and changes nothing in halving.
-
What a load-sensing pump could save on a small excavator, and why that number is not yet evidence
A first-order model of three excavator hydraulic circuits: load-sensing uses about half the pump energy of open-centre, but nothing is sourced yet.
-
Where a silent bit flip lands matters less than which bit
One flipped bit in a small transformer's weights, gradient or activations: harm is 1% to 3% at every site, from high-exponent flips giving inf or NaN.
-
A buoyancy glider beats a propeller AUV through hotel load, not hydrodynamics
In a swept model of a 52 kg glider and a same-mass propeller AUV, the glider has longer range in 5.9% of cases at equal hotel load, 89% at five times.
-
A glider beats an AUV on energy per metre once you read the AUV's power bill
A REMUS-class AUV's vendor endurance implies 104 to 175 W of non-propulsion load; against it a buoyancy glider has the longer range per kWh in every case.
-
A glider's energy win over a REMUS costs it five days per 100 km
Counting time too, a REMUS-class AUV beats a Slocum-class glider in 78% of simulated pairs; the glider wins only if delay is worth under 0.7 of energy.
-
A hydrofoil survey boat does not out-range a displacement hull at 200 kg
In a 2,916-case sweep, a foilborne 200 kg craft never beat a 4 m displacement monohull on range; the best case reached a quarter of it.
-
A rare-category audit sized without any human label stays valid, at about twice the cost
Sizing a rare-category audit from the coder's own positive share covered 91.8 to 94.4% at 1.9 to 2.8 times the oracle size, where stopping rules failed.
-
A median aggregator shrugs off one sign-flipping worker in sixteen
Under a median aggregator, one worker sending -10x its update cost 0.095 nats at 4 workers, 0.021 at 8 and 0.002 at 16 in simulated DiLoCo training.
-
A median defence at sixteen workers is below what three seeds can see
At 16 workers a coordinate-wise median costs about 0.004 nats against a sign-flip attacker, 5.8x less than at 8, but three seeds cannot tell that from zero.
-
A scale model that matches power still flies differently
Wing changes can make a 1:5 cargo-UAV model match full-size power, but none of 17 matches keeps its flight modes within 5%. The error is in the tail.
-
A Slocum-class glider only out-ranges a propeller AUV when the AUV burns about five times its hotel power
With Slocum aerodynamics from an open flight model, a glider never beats a same-mass propeller AUV at equal hotel load; break-even is a median 4.9 times.
-
A smaller, longer tail restores the short period of a 1:5 scale aircraft model
Searching 270 tail designs on each of 17 power-matched wings, 6 reach phugoid, short-period frequency and damping within 5% of the full-size aircraft.
-
A cheap synthesis ranking failed on one high-fanout chip design and held on another
Two high-fanout SKY130 designs, 36 to 39 distinct outcomes each: the mapped-netlist ranking picked 14% worse on one and 0.4% on the other.
-
I quoted a NASA scaling table from memory. Four details were wrong.
Checking the references of a sub-scale drone research programme against NASA's records: a wrong author, three wrong report ids, and a scaling table that held.
-
Does a catamaran out-range a monohull on a survey mission?
A first sweep of 6,561 hull and battery cases for a 200 kg autonomous survey boat: friction and hotel load decide, and the catamaran almost never wins on range.
-
Does a drone's power-law survive the flight-day effect?
Froude scaling needs power to grow as mass^1.5. On 134 real delivery-drone flights the exponent is 1.54, and 1.53 with a separate intercept for each flight day.
-
Is 'never cut on the cheap rung' the right chip-search policy? Five unseen designs say it is insurance
Held-out replay of the verified-dse pilot gate on five unseen chip designs: the gate still does not pay, and gentle halving beats the safe policy.
-
A 1:10 cargo-drone model needs 29 to 45% more power than scaling says
Airfoil polars at falling Reynolds number back the aerolog-scaling programme's drag hypothesis and show a 0.4 to 0.5 loss of maximum lift.
-
A 1:10 Froude model of a cargo drone trims 6.5% fast
With airfoil polars at each surface's Reynolds number, a 1:10 Froude model misses the 5% trim-speed bar unless flown at the full-size lift coefficient.