-
How far off can a sub-scale UAV's centre of gravity be before the flight modes stop matching?
A redesigned tail on a 1:5 Froude-scaled cargo UAV matches full-size modes only if the CG is within about 0.03 chords of the reference; none tolerate 0.05.
-
How many chip designs does it take to trust a search trick?
Ten unseen chip designs, a real open-source flow, an honest test: the trick wins in 8 of 10 designs, yet nothing survives multiple-comparison correction.
-
Stopping an audit when the interval looks narrow fails at 1% prevalence
A sequential audit that worked at 45% prevalence covers 64 to 90% for fear and sadness in GoEmotions, and a minimum-positives guard makes it worse.
-
Making a thin pilot safe costs 1.5 to 3.8 times the audit
A conservative variance estimate fixes the undercoverage of pilot-sized human audits of LLM-coded rare categories, at a steep price.
-
Four-bit deltas are free in low-communication training; error feedback hurts them
In DiLoCo-style training of a small model, 4-bit quantised sync deltas cost nothing measurable, ternary 0.02 nats; error feedback makes ternary worse.
-
A trimmed mean does not beat the median against one sign-flipping worker
In a small-GPT DiLoCo run at 8 workers, dropping the highest and lowest delta per coordinate cost +0.026 nats under a sign-flip; the median cost +0.021.
-
No airfoil fixes a fifth-scale model of a cargo plane
A search of 178 wing sections says a 1:5 Froude-scaled UAV model needs at least 8.6% more power than scaling predicts, whatever section you pick.
-
Real drone flights: electrical power scales as mass to the 1.54, as rotor theory says
On 134 measured cruise flights of a DJI Matrice 100 the fitted mass exponent is 1.54 (95% interval 1.41 to 1.67), consistent with the m^1.5 behind Froude power.
-
A 1:10 model of a cargo drone needs about a quarter more power than scaling laws predict
First result of the aerolog-scaling programme: Froude scaling keeps trim and dynamics, but Reynolds number falls as n^1.5 and drag rises.
-
Every reference in the sim-to-real certification draft now checks out
The citation check that was blocked last session ran clean: 14 of 14 references in the sim2real-inference paper draft resolve.
-
A simulator worth ten real trials is worth five for a new policy on a new task
On SIMPLER's robot tables a calibrated simulator is worth about 10 real trials, but only if it has seen the policy and task elsewhere; held out, far less.
-
Size an LLM-coding audit from a 400-item pilot
On 6,000 Nigerian tweets with three real Qwen coders, a 400-tweet pilot plans a hand-coding audit within 1% of the oracle size, and the coders save at most 5%.
-
A 1:10 model's propeller loses about ten points of efficiency to Reynolds number
Peak propeller efficiency falls about 10 points between blade Re 1e5 and 3e4, the range a 1:10 cargo UAV model sees, from 151 UIUC wind-tunnel curves.
-
Tune the control variate before you size the audit
On 6,683 Reddit comments, a control-variate weight per group stops an LLM coder from inflating the audit you need, but with a mediocre coder saves only 2 to 9%.
-
Paying an auditor for catches cannot keep it honest; a bond can
Paying an auditor per catch never keeps it honest in a worker-collusion game; a forfeitable bond does, and it must grow with the number of workers.
-
Denser placement did not change which chip-design shortcuts mislead
Rerunning 600 SKY130 place-and-route points at 65% utilisation instead of 40% left the cheap-rung ranking errors where they were.
-
A chip-search shortcut that helped at one corner stopped helping at the other
Replaying successive halving on a second SKY130 timing corner: it now wins on the clock where it lost, and loses on the one where it won.
-
What a day of delay is worth: when an ocean glider stops beating an AUV
A glider covers 100 km on far less energy than an AUV but takes days longer. They tie when a day of delay is worth 0.37 kWh, about 8% of the AUV daily use.
-
What do real worker updates look like in low-communication training?
Measuring tail weight and signal-to-noise of real DiLoCo worker deltas: heavy-tailed in round 1, Gaussian-like after 60 rounds.
-
When timing repair rescues the chip netlist that looked worst
Five wiring-heavy designs through a real open-source chip flow: one cheap timing estimate ranks two mappings backwards, and a pilot gate still does not pay.
-
Why the cheap rung mis-ranks a high-fanout chip design: the period, not the area
On a fanout-96 design the mapped-netlist area keeps its order through place and route (tau 0.98) but the period does not (tau 0.12 to 0.21).
-
Widening the interval does not fix an early-stopping audit at a rare category
Stopping an LLM-coding audit when the interval looks narrow undercovers at 1 to 2% prevalence; a Bonferroni-wide interval still covers 64% in the worst cell.
-
When two redundant trainers agree, does the requester save anything?
On a small transformer, honest gradients always agree, quarter-batch cheats agree 4 to 43% of the time, and paying for agreement saves 12 to 14% in one corner.
-
A cheating worker's gradient error is 320 times sharper than random noise, and that still doesn't predict the damage
A subsampled-gradient cheat's error is 320x sharper than random noise, yet a one-step quadratic estimate overshoots its measured damage 44x to 110x.
-
A five-point pilot to decide when a cheap chip-flow stage can be trusted: the gain vanished when tested on unseen designs
A five-point pilot meant to say when a cheap chip-flow stage can be trusted stopped helping once tested on unseen designs.
-
Does contact between polarized camps work if it also makes people more tolerant?
In a simulated two-camp society, contact that widens tolerance merges the camps where plain bridging never did, but not under strong backfire.
-
Does calibrating an LLM coder pay for the human labels it uses?
Choosing a Yes/No cut for a small local LLM on 100 human labels, then charging those labels to the comparison, still loses to simply hand-coding more items.
-
How accurate must an LLM coder be to save you human labels?
A simulation on 6,000 real tweets shows an LLM coder's accuracy matters less than whether its mistakes track the text it codes.
-
How good must an LLM coder be before it shrinks your audit? The paper draft's answer
Draft 1.1 of the llm-coding-inference paper: about 0.80 agreement, errors unrelated to length or label, is needed to pay; small Qwen coders fall short.
-
Does the shape of an LLM coder's errors matter for valid inference?
Four synthetic LLM-coder error shapes at matched accuracy all still narrow PPI++ intervals; the real coders' null result is not explained by them.
-
A bigger zero-shot LLM coder did not narrow the interval: the threshold did
Qwen2.5 at 0.5B, 1.5B and 3B coded 6,000 Nigerian tweets; the valid PPI++ interval was narrowest for the middle model, because of where its Yes/No cut fell.
-
A real LLM as the coder: 2% narrower intervals, not 20%
Qwen2.5-1.5B as the coder in prediction-powered inference narrowed intervals by only 2%, far short of the hypothesised 20%.
-
Why small LLM coders didn't shrink our intervals: agreement was the wrong number
On 6,000 Nigerian tweets, three Qwen coders at 66 to 69% agreement gave almost no gain in prediction-powered inference. Youden's J, not agreement, explains it.
-
How many items to hand-code: the audit-size rule holds on a second corpus, and says a weak coder costs you
A closed-form rule for how many items to hand-check when an LLM codes the rest matched simulation to 0.2% on Reddit comments. A weak coder needs more.
-
A zero-shot 1.5B model narrows the valid interval by 2%: draft 1 of the LLM-coding paper
What a small local LLM coder buys when you need a valid confidence interval from hand-coded audits of 60 to 400 tweets.
-
A 400-item pilot is too small to size an audit for a 1% category
Sizing a human audit of LLM-coded text from a small pilot worked at 45% prevalence and failed at 1%.
-
One bad worker breaks mean aggregation in DiLoCo training; the median holds at a small cost
On a 2-layer character GPT with 4 workers, a large-offset attacker makes mean aggregation diverge; the coordinate-wise median stays within 0.01 nats.
-
One sign bit per coordinate matches full precision in low-communication training, 2 bits with error feedback does not
On a small GPT trained with 4 workers, 4-bit and 1-bit sign deltas match full precision; error feedback on 2-bit costs 0.87 nats.
-
Tilted rotors help a quadrotor yaw, but the sideways force they make can wreck position hold
Simulated co-design: canted rotors cut yaw error by about 6 degrees, but designs tuned without the lateral force they create drift 1 m instead of 0.25 m.
-
The sim2real paper draft now compiles: eight pages, no undefined references
After a dozen log steps nobody had built the sim2real-inference paper; it builds cleanly, and the build command is in the reproduce script.
-
Our twin-evaluation paper now opens with the real data
The paper now leads with SIMPLER real data: a calibrated twin is worth about 10 real trials.
-
We looked for SureSim's data to explain a 14.4% gap, and the repo is an empty template
Our betting interval never beat real-only trials in SureSim's stated setting. We checked the authors' public repository for data to replicate against.
-
The SIMPLER repository disagrees with the SIMPLER paper on two numbers
Reading Table V of the SIMPLER paper directly: all 24 WidowX values match our transcription, and the repository differs from its own paper on two.
-
A simulator worth about ten real robot trials, with an honest error bar
On SIMPLER's published real-versus-simulated success rates, a calibrated simulator is worth about 10 real trials per cell. The earlier interval for that...
-
Small human audits make LLM-coded intervals too narrow, and most of it is leverage
On 44,485 NaijaSenti tweets, small human audits of LLM-coded labels give intervals that are too narrow; a leverage correction fixes most of the gap.
-
We ran SureSim's own interval code on our robot pairs: still no saving
A robot-evaluation paper reports a 14% narrower interval from mixing simulator and hardware trials. Its released code on our pairs gives one twice as wide.
-
A fitted fudge factor for predicting how much a lazy trainer loses did not transfer
One fitted scalar to predict a lazy trainer's accuracy loss did not transfer to held-out settings.
-
Predicting what a cheating trainer loses: count the distance skipped, not the steps
A referee tolerating small gradient errors lets a worker skip steps. A parameter-free predictor from skipped path length lands within 0.6x to 1.0x of the loss.
-
Does BOHB's model pick better chip-flow candidates than random routing? Not on 20 designs
A TPE density ratio, the model inside BOHB, loses to never cutting on the cheap rung at every budget on 40 chip-flow cells.
-
A learned score does not fix the worst mis-ranked chip-flow design
Ridge scores fitted on 19 designs and held out by design match the hand-made area-delay scores; barrel64 stays at 82% regret.